Aller au contenu
Corporate & Investment Banking Mon parcours English

Accueil / Capital Markets & Securities Services / Securities Services / Module 24

24Reconciliation & Controls

The control that catches errors before they become losses · How a bank proves its records agree, finds the breaks, and increasingly automates the discipline

Prouver que deux enregistrements concordent, et comprendre pourquoi ils divergent.

Lecture
1 h 50
Sections
69
Questions de rappel
42
Questions de QCM
15
Simulateurs
1

Le texte de ce module est encore en anglais : la version française arrive. Les outils interactifs, eux, sont déjà en français.

1Reconciliation is the bank's independent check on itself

Before any of the machinery, one idea. A bank runs on records, and records can be wrong. Reconciliation is how the bank proves — to itself, to its clients, and to its regulator — that the records it acts on actually match reality. This part builds that idea from zero and places it in the flow of a trade.

A bank cannot take its own word for what it holds

Start with a question a bank must be able to answer at any moment: do we actually hold the cash and securities our books say we do? It sounds trivial. It is not. A large bank books millions of movements a day across hundreds of accounts, dozens of systems and a global network of other banks and custodians who hold its money and its securities. Every one of those movements is an instruction that may or may not have happened the way the books assume. Cash that was supposed to arrive may not have. A trade that was supposed to settle may have failed. A fee may have been charged that nobody recorded. Somewhere in that torrent, the bank's own picture of the world drifts away from the truth.

Reconciliation is the discipline that closes that gap. It takes two records that should describe the same thing — the bank's own ledger and an outside statement, say — and compares them field by field. Where they agree, the record is confirmed. Where they disagree, the difference is isolated, named, and handed to someone to explain. That difference has a name that will recur on every page of this module: a break. A break is not noise to be swept away. It is the output the whole exercise exists to produce, because a break is where an error, a fraud or a loss first shows itself.

The reason a bank cannot simply trust its own systems is that a copy of a mistake is still a mistake. If the desk books a trade wrong and that wrong booking flows into the risk system and the ledger, all three systems will agree with each other perfectly — and all three will be wrong. Internal consistency proves nothing on its own. What proves something is comparison against an independent source: the agent bank that actually holds the cash, the custodian that actually holds the shares, the counterparty that saw the other side of the trade. Reconciliation is valuable precisely because at least one side of the comparison comes from outside the place that made the record.

Reconciliation is a detective control, and it sits downstream on purpose

Controls come in two families, and reconciliation belongs firmly to one of them. A preventive control stops an error before it happens: input validation that rejects a malformed trade, a four-eyes check that requires a second person to authorise a payment, a limit that blocks an oversized position. A detective control assumes something will slip through the preventive net anyway, and catches it after the fact. Reconciliation is the classic detective control — the safety net strung beneath every preventive check, on the working assumption that prevention will sometimes fail. You need both. Prevention alone is overconfidence; detection alone is negligence.

Figure 1 Reconciliation is a detective control inside a wider framework — the backstop that assumes prevention will sometimes fail

It is worth naming the preventive controls that sit around reconciliation, because reconciliation assumes they exist and will still sometimes fail. On the way in, input validation rejects a malformed trade before it is booked; a four-eyes check requires a second person to authorise a payment or a static-data change; limits block an oversized or out-of-mandate position; and access controls stop people touching systems they have no business in. Each of these tries to stop an error at source. Reconciliation is what remains after all of them — the detective control that assumes, correctly, that some error will slip past every preventive gate, and goes looking for it. That is why the two families are complementary rather than alternatives: prevention narrows the stream of errors, and detection catches what the narrowing misses.

Reconciliation sits downstream in the trade lifecycle, and that position is deliberate. A trade is captured at the front, settles through agents and custodians, and its records flow onward into the books, risk and regulatory reporting. Reconciliation is not a single stop on that journey. It reaches across every stage and asks, at each one, "does what we think happened actually match what the outside world recorded?" It runs after the action, comparing the front-office view against a source that the front office does not control. That downstream, independent vantage point is what lets it see a false booking that everything upstream has already accepted as true.

Because it is independent, reconciliation cannot be quietly overruled. If a trader could reconcile their own book, they could make an inconvenient difference disappear by simply declaring it matched. The whole point is that the person and the data confirming the records are separate from the person and the desk that created them. This is why reconciliation lives in Operations, Product Control and Finance — the middle and back office — and never on the trading desk. Independence is not bureaucracy. It is the source of the control's authority. We return to it in depth in Part 3.

The same discipline, applied to different pairs of records

Every reconciliation has the same shape, but they are named by what they compare. Underneath, the engine, the keys and the break workflow are identical; what changes is which two records are being lined up and what "agreement" means for them. Two big families dominate: cash and securities — nostro reconciliation proves the cash, position reconciliation proves the stock. Around those two sit trade, intersystem, client-asset, collateral and intercompany reconciliations that keep the moving parts consistent. The table below is the map of the field — every type you will meet is one row of it.

Reconciliation type Side A (our record) Side B (the other side) The question it answers
Nostro / cash Our internal cash ledger The agent bank's account statement Did the payments we expected actually happen?
Position / stock Our securities sub-ledger The custodian's or CSD's holding statement Do we really hold the stock our books claim?
Trade Our trade record The counterparty's, broker's or clearer's record Do both sides agree on the terms of the deal?
Intersystem One internal system (e.g. front office) Another internal system (e.g. the ledger) Do our own systems tell the same story?
Front-to-back The trade as captured at the front The position, cash and P&L through to the GL Does the trade stay consistent all the way down?
Depot / client-money Our record of client assets and money The banks and custodians actually holding them Are the assets we safeguard for clients really there?
Collateral / margin Our record of collateral posted or held The counterparty's, clearing house's or triparty agent's Do both sides agree on the margin that should move?
Intercompany One group entity's record Another group entity's record Do internal trades and balances agree and net to zero?

The common thread is worth stating plainly: every reconciliation compares an internal record against an external source, or one internal record against another. Once you see that, the apparent zoo of reconciliation types collapses into a single idea applied to different objects. A EUR nostro recon and a client-asset depot recon feel like different worlds to the teams that run them, but the engine underneath is the same: two sides, a set of keys, a match rule, and a break at the end where they disagree. It is still worth walking each type once, because what "agreement" means — and what a break reveals — differs enough to be worth knowing before the machinery of Part 2.

Nostro / cash reconciliation

The cash family. Our internal cash ledger against the agent bank's statement, asking whether the payments we expected actually happened. A break here can be a failed receipt, an unbooked fee, an FX difference, or, at worst, cash moving where it should not. It is the most fundamental reconciliation because a bank that cannot reconcile its cash does not really know its own liquidity.

Position / stock reconciliation

The securities family. Our sub-ledger of securities held against the custodian's or CSD's holding statement, asking whether we really hold the stock our books claim. For each security and account the quantity we record must equal the quantity they hold; a difference is a position break, and it often reveals a settlement fail or a mis-processed corporate action.

Trade reconciliation

Our record of a trade against the counterparty's, the executing broker's, or the clearer's, asking whether both sides agree on the terms of the deal. A trade break caught early prevents a settlement fail later, which is why trade matching sits close to the point of execution.

Intersystem reconciliation

One internal system against another — front office against back office, risk against the ledger — asking whether our own systems tell the same story. It has no external side; both records are ours, which makes it a pure test of internal consistency as data flows between systems.

Front-to-back reconciliation

A specific, demanding form of intersystem check: that a single trade stays consistent from front-office capture through the position, the cash and the P&L, all the way to the general ledger. It is where the cascade principle of Part 4 becomes operational — a check that nothing was dropped or transformed wrongly along the chain.

Depot / client-money reconciliation

The heavily regulated case. Client assets and client money held, against the records of the banks and custodians holding them. Here the reconciliation is not just prudent but a legal protection — the CASS regime in the UK and EU equivalents require it, and an unreconciled client position is a serious control and regulatory failure in its own right.

Nostro reconciliation is the reference case for the whole module

We anchor the module on nostro reconciliation because it is the most fundamental and the easiest to picture. A nostro is the bank's own cash account held at another bank — "our money at their place." To pay and receive in every currency the bank operates in, it holds nostro accounts at agent banks around the world: euros at one correspondent, dollars at another, yen at a third. The nostro recon compares two views of that account. Side A is the bank's internal cash ledger — what it expected to happen. Side B is the agent bank's statement — what actually happened. Every expected payment in or out should match a movement on the statement. Anything unmatched is a break.

Nostro breaks read like a list of things that can go wrong with money, and that is exactly why the recon matters. An expected receipt that never arrives is a failed or late payment. A movement on the statement the bank never booked is an unexpected item — a fee, an error, or worse, a payment sent somewhere it should not have gone. A payment for the wrong amount is usually an unbooked charge or an FX difference. And in the worst case, cash moving where it should not be is fraud or misdirection, which nostro reconciliation is often the first control to catch. A bank that cannot reconcile its cash quickly does not really know its own liquidity — and in Europe, intraday liquidity is a growing supervisory focus in its own right.

Collateral and margin reconciliation is the fastest-growing family

The families so far — cash and securities — predate the 2008 crisis; the crisis created a third that now rivals them in importance: collateral and margin. When two parties trade a derivative, each is exposed to the other defaulting, so they exchange collateral to cover that exposure — and every exchange of collateral is a set of records that must agree. Two flavours matter. Variation margin (VM) is collateral posted, usually daily, to cover the change in a trade's market value; initial margin (IM) is a larger buffer posted up front to cover potential future exposure. Both are calculated, called, posted and held — and every step is reconciled.

Margin reconciliation compares the two sides' views of what should be posted, and it has teeth because a difference is money. If our systems say the counterparty owes us EUR 5 million of variation margin and theirs say EUR 4 million, that EUR 1 million gap is a margin dispute — a break with a direct funding and credit consequence, not a cosmetic difference. Regulation has made this mandatory and daily: cleared derivatives are margined by the clearing house, and bilateral derivatives fall under the Uncleared Margin Rules (UMR), which require large participants to exchange and reconcile IM and VM. On the securities side, triparty collateral — where an agent bank manages the collateral between the two parties — adds a further reconciliation, between each party's records and the triparty agent's.

For the business, this is where reconciliation meets credit risk head-on. A cash or position break is usually an operational error; a margin break can mean the bank is under-collateralised against a counterparty that could default tomorrow. That is why margin reconciliation runs daily, disputes are escalated fast, and the discipline sits close to the collateral-management and credit-risk functions — and why it is one of the busiest areas of reconciliation transformation. All figures illustrative.

Intercompany reconciliation keeps the group consistent with itself

A large bank is not one entity but many legal entities under one group, and they trade with each other constantly. When the London entity books a trade with the Paris entity, each records its own side, and at group level the two must agree and net to zero — an internal trade cannot create or destroy value for the group as a whole. Intercompany reconciliation checks exactly this: that matched pairs of internal positions, balances and P&L agree across entities, so that the consolidated group accounts are clean.

It matters most at the consolidation boundary, where a surviving break becomes a misstatement. An intercompany difference left open to period-end distorts the group's consolidated balance sheet and can trigger audit findings and restatements. Because both sides are internal, it is a pure intersystem reconciliation — there is no external truth to appeal to — which makes consistent booking conventions and shared reference data across entities essential to keeping it clean. Illustrative.

Cash reconciliation underpins intraday liquidity management

Nostro reconciliation is not only a control against error; it is the raw material of the bank's view of its own cash, and Treasury depends on it. A bank must know, through the day, how much cash it actually holds in each currency at each agent — to fund its obligations, to avoid costly overdrafts, and to meet supervisory expectations on intraday liquidity. That picture is only as accurate as the reconciliation behind it: an unreconciled or delayed nostro means Treasury is managing liquidity on a number it cannot fully trust.

This is why the move toward intraday and real-time cash reconciliation is a Treasury story as much as a control story. Reconciling cash once, overnight, tells Treasury yesterday's position; reconciling continuously tells it where the money is now — which is what intraday liquidity management actually needs. As supervisors sharpen their focus on intraday liquidity, expecting banks to monitor and control their positions through the day, the cadence of cash reconciliation becomes a direct input to a regulatory and funding discipline, not merely operational hygiene.

Why this became a formal discipline

Reconciliation has always existed in some form — merchants have tied their books to their bankers' statements for centuries — but it became a *regulated, independent discipline* for a specific reason: things went catastrophically wrong when it did not exist. A run of back-office failures and trading scandals through the 1990s and 2000s made the same point again and again — that a control operated by the people it was meant to check is no control at all. Each episode sharpened the expectation, now codified, that reconciliation and the back office must be genuinely independent of the front office, organisationally and in system access.

Two broader forces turned that expectation into hard rules. The first was the operational-risk framework that emerged under the Basel accords, which required banks to hold capital against exactly the kind of losses that weak reconciliation lets through — and to measure and manage those losses, which means running the control well and recording its failures. The second was the tightening of client-asset protection, brought into sharp focus when large firms failed and clients discovered that money supposedly held for them had not been properly segregated or reconciled. Out of these forces came the modern picture: reconciliation as a mandatory, independent, evidenced key control, supervised directly by regulators. The discipline in this module is not a matter of good housekeeping; it is the institutional memory of what happens without it.

Position reconciliation is the securities twin of the nostro

If nostro reconciliation proves the cash, position reconciliation proves the stock — and it carries risks that are easy to underestimate. It compares our internal record of securities held — the sub-ledger, what we think we hold — against the statement of whoever actually holds them for us: the custodian, sub-custodian or CSD, what is actually held. For each security, identified by its ISIN, and each account, the quantity we record must equal the quantity they hold. A difference is a position break.

Position breaks matter more than their frequency suggests, because a wrong position corrupts more than one number. If our record of a holding is wrong, our risk figures and our P&L are wrong too — we may be hedging or valuing a quantity we do not hold. A break often reflects a settlement fail: stock we thought had arrived but has not, or vice versa. Corporate actions are a classic source, because dividends, splits and mergers change holdings and a mis-processed event breaks the position quietly. And for client assets, an unreconciled position is not merely an operational nuisance but a safeguarding failure — the entire point of the reconciliation is that the assets we hold for clients are demonstrably there. In Europe, securities are held through CSDs such as Euroclear and Clearstream and through custodian networks under CSDR, which also introduced settlement discipline with penalties for fails, making the accuracy of position reconciliation a supervised concern.

Why reconciliation is not optional

It prevents real losses. An unnoticed break can be a missing payment, a wrong position, or a fraud growing quietly in the dark. Caught early, it is cheap to fix — a phone call, a corrected booking. Left to age, the same break can become money that is genuinely gone and no longer recoverable. Reconciliation's economic value is that it moves the moment of discovery as early as possible, when remediation still costs almost nothing.

It is a legal obligation, not a courtesy. Client-money and client-asset rules (the CASS regime in the UK, with equivalents across the EU), books-and-records requirements, and SOX-style internal-control regimes all require independent reconciliation and evidence that it was performed. A firm that cannot show complete, timely, well-documented reconciliations is not merely untidy; it is in breach. Regulators inspect the reconciliation evidence directly, and failures bring fines and remediation orders. Part 4 treats these regimes as system specifications.

It is one of the largest data problems in the bank. Millions of records, dozens of source systems, incompatible formats, matched every single day against external statements that arrive late, in different layouts, with different reference conventions. Getting that data clean and aligned is a serious data-engineering challenge, and the health of every reconciliation depends on it. This is the thread that leads, by the end of the module, to why reconciliation is where transformation money increasingly goes.

And it is ripe for automation. A great deal of reconciliation is still done by hand — analysts eyeballing statements, chasing references, keeping the count in spreadsheets. That makes it slow, expensive and fragile, but it also makes it one of the richest targets in the bank for better data, no-code tooling and machine learning. The direction of travel, which Part 5 develops in full, is from manual and batch toward automated, near-real-time and business-owned.

The scale is what makes it a technology problem

It is worth pausing on the sheer size of the thing, because the scale is what turns a simple idea into an engineering discipline. A large bank may hold hundreds of nostro accounts across dozens of currencies and agent banks, custody relationships spanning many markets and CSDs, and a trade population running to millions of transactions. Every one of those accounts and populations is reconciled — many daily, some intraday — against a statement that arrives in its own format, on its own schedule, with its own reference conventions. The number of individual record comparisons a large institution performs in a single day runs into the tens of millions. These magnitudes are illustrative, but the order of magnitude is real.

At that scale, three consequences follow that shape everything in this module. Manual reconciliation does not survive contact with the volume — a person can eyeball a hundred items, not a million, so matching must be automated and only the exceptions can be human. Small percentages become large absolute numbers — a 2% false-break rate on a million records is twenty thousand items for someone to clear, which is why shaving that rate is worth real money. And consistency of data becomes a first-order problem — when millions of records from dozens of sources must line up, a single upstream format change can manufacture thousands of breaks overnight. The scale is precisely why reconciliation lives where technology and control meet, and why an analyst who understands its data plumbing is the one who can improve it.

The pressure on reconciliation is rising, not falling

Two forces are making reconciliation harder and more scrutinised at the same time. The first is settlement acceleration. As markets move to shorter settlement cycles — the move to T+1 in several major markets, with Europe preparing to follow — the window to catch and fix a break shrinks. A reconciliation that ran comfortably overnight on a T+2 timetable has far less slack when settlement is a day faster, which pushes firms toward intraday and real-time reconciliation. The second force is regulatory. Client-asset protection, derivative-reporting quality under EMIR, settlement discipline under CSDR, and operational-resilience rules under DORA have all tightened, and every one of them treats reconciliation as core evidence. The result is a discipline under simultaneous pressure to run faster, cover more, and prove itself more rigorously than ever.

  1. Why is comparing an internal system against a copy of itself not a control?

  2. What is the difference between a preventive and a detective control, and which is reconciliation?

  3. Name the two big families of reconciliation and what each proves.

  4. What is a nostro account, and what two records does a nostro recon compare?

  5. Give two reasons reconciliation is not optional.

  6. Why does the scale of reconciliation make it a technology problem?

  7. Why is the pressure on reconciliation rising?

  8. What is a margin reconciliation, and why does a margin break matter more than a typical operational break?

  9. What everyday activity is reconciliation identical to?

1

  1. Why is comparing a system against a copy of itself not a control?

  2. Is reconciliation preventive or detective?

  3. What does a nostro reconciliation prove?

2The lifecycle turns raw feeds into a cleared, evidenced control

This is the engine room. Every reconciliation, whatever it compares, runs the same pipeline: load the two sides, normalise them, match them, isolate the exceptions, investigate and resolve them, then report the outcome. This part walks that pipeline stage by stage, opens the matching engine that sits at its heart, and follows a single break from detection to a fix that stops it ever coming back.

The whole run has six stages, and the effort is not spread evenly

A reconciliation run is a pipeline with a fixed shape. Data comes in, gets cleaned, gets matched, and whatever fails to match becomes work for a human. The table below is the spine of this entire part — six stages, each with a job, an owner and an output. Read it once now; we open each stage in turn.

Stage What happens Mostly done by Output
1. Load / ingest Pull both data sets from files, feeds, messages or APIs into the platform Automation Two raw data sets, side A and side B
2. Normalise / transform Map both sides to a common format: field names, dates, signs, currencies Automation + config Two aligned, comparable data sets
3. Match Apply the keys and rules in passes; auto-match as much as possible The matching engine Matched pairs + a pile of unmatched items
4. Identify exceptions Classify what did not match cleanly as breaks; queue them Engine + workflow A prioritised break queue
5. Investigate / resolve Find each break's cause; correct, chase, claim or write off Ops analysts (human) Cleared breaks + cause codes
6. Report / evidence Record outcomes, produce control evidence, feed the metrics Automation + sign-off Attestation, KPIs, audit trail

Two facts about where the effort goes decide whether a reconciliation is cheap or expensive. First, most of the data work lives in stages 1 and 2 — loading and normalising. This is the unglamorous plumbing, and it is where most reconciliations quietly succeed or fail: get the data clean and aligned, and matching is easy; feed the engine messy data, and it will manufacture breaks that are not real. Second, most of the human work concentrates in stages 4 and 5 — the exceptions. Matching itself is highly automated and nearly free. Every unmatched item, by contrast, is a person's time. A healthy reconciliation therefore maximises auto-matching and minimises manual investigation, and it does so mainly by fixing the data upstream rather than by loosening the rules.

Figure 2 The reconciliation lifecycle: data effort concentrates early in load and normalise; human effort concentrates later on the exceptions

Stage 1 — Loading pulls two sides that were never designed to meet

Loading sounds mechanical, and it is where the first breaks are born. The platform must pull side A — usually the bank's own records, from a ledger or sub-ledger — and side B — the external statement or the other internal system — and get both into one place on a schedule. The two sides were built by different organisations for different purposes and never intended to be compared. Side A might arrive as a clean internal feed; side B might be a SWIFT statement message, a CSV file emailed overnight, or a file dropped on a server by a custodian. A late file, a missing file, or a file whose layout changed without warning breaks the whole run before a single record is matched. Feed fragility is the single most common operational cause of a failed reconciliation.

Stage 2 — Normalisation is the unglamorous step that decides everything

Normalisation maps both sides onto a common language so that "the same thing" really looks the same. The two records may describe an identical payment yet disagree on the surface in a dozen mechanical ways: one side writes the date as 16/07/2026 and the other as 2026-07-16; one uses a minus sign for a debit and the other a separate direction field; one carries the amount as 1,250,000.00 and the other as 1250000; one names the currency EUR and the other 978. None of these differences is a real error, but every one of them will stop two records matching if it is not resolved first. Normalisation aligns field names, date formats, signs, currency codes and reference conventions so that genuine agreement is visible and cosmetic differences vanish.

This is the highest-leverage step in the whole pipeline, and it is almost entirely a data problem. A format mismatch that survives normalisation does not just fail to match — it creates a false break, an item flagged as a difference that reflects a data issue rather than a genuine error. False breaks are the dominant cost in most reconciliations, and they are manufactured here. Getting normalisation right, and keeping it right as upstream formats drift, is the core of what a reconciliation business analyst delivers. We return to false breaks in depth once the matching engine is on the table.

A concrete before-and-after makes the point. Suppose our ledger records a debit as "-1250000" with the date "16/07/2026" and the currency "EUR", while the agent's statement records the same movement as "1,250,000.00" in a separate amount field, with a direction flag "DR", the date "2026-07-16", and the currency code "978". Every field describes the identical payment, and every field looks different. Normalisation resolves the sign convention (a leading minus versus a direction flag), the thousands separators and decimals, the two date formats, and the numeric-versus-alphabetic currency code — so that after the step both records read "account NOSTRO-EUR-01, value date 2026-07-16, EUR, debit, 1250000.00." Only now can the engine see that they agree. Skip any one of these conversions and the payment breaks for a reason that has nothing to do with the money.

Stage 3 — The matching engine pairs records in passes, tightest logic first

The matching engine takes the two normalised data sets and pairs records using configured keys and rules, running in passes. A matching key is the set of fields used to decide that a record on side A and a record on side B describe the same thing — typically account, date, amount and a reference. The engine does not try to be clever all at once. It runs the strictest logic first and falls back to looser logic only on what is left, so that clean matches are made with confidence and fuzzy matches are a last resort rather than a first guess.

Figure 3 The engine runs rules in passes — exact, then tolerance, then fuzzy or grouped — and whatever survives every pass becomes an exception

Four matching approaches, from strict to forgiving

Exact matching demands that the keys and amounts agree precisely. It is simple, strict and ideal for clean data with reliable references. Most items in a healthy cash reconciliation match here, in the first pass, and never trouble a human.

Tolerance matching treats two records as agreeing if their difference falls within a small threshold. A rounding of a few cents or a tiny FX difference should not generate a break; a sensible tolerance absorbs it. But a tolerance is a risk decision disguised as a convenience — set it too wide and it hides real errors under the guise of noise. The threshold is a judgement about how large a difference the firm is willing to ignore, and it should be set deliberately, not casually.

Fuzzy matching pairs records on approximate keys — a reference that is close but not identical, a name that nearly agrees — to cope with messy data where exact keys are unreliable. It is powerful and dangerous in equal measure, because a fuzzy match that is wrong pairs two records that do not belong together and hides two real breaks at once.

One-to-many and many-to-many matching handle the common case where the records do not line up one for one. A single payment may settle several invoices, or a batch on side A may net against a group on side B. The engine groups records and matches on the sum — for example, treating a group as matched when the total on side A equals the total on side B. Real cash and securities flows are full of these netted movements, so grouped matching is not an edge case; it is core.

The auto-match rate is the single number that summarises the engine

The auto-match rate is the percentage of records paired automatically, with no human touch — and it is the headline efficiency metric of any reconciliation. A clean cash reconciliation often runs at 95% or higher; harder reconciliations with messy references or complex netting run lower. The number matters because every unmatched item is manual effort: a higher auto-match rate means lower cost, a faster close and fewer opportunities for human error. You raise it with better input data, smarter keys, sensible tolerances, and rules that learn from how past breaks were resolved. But there is an honest limit. Chasing 100% by loosening the rules hides real breaks under a false calm. The goal is a high auto-match rate achieved with tight tolerances — efficiency and rigour together, not one bought with the other.

The matching rule is configuration a business analyst owns

Here is the crucial point about the whole engine: the keys, passes and tolerances are configuration, not code. A reconciliation is defined by its rule set, and that rule set is a document a business analyst writes, versions and gets audited — not a program an engineer compiles. This is why the reconciliation BA is a technology-adjacent role that needs no software development: the deliverable is a precise specification of how two data sets should be compared. The table below shows the same rule set that governs our running EUR nostro reconciliation, laid out as a specification.

Config element Setting for the EUR nostro recon What it does
Sources A = internal ledger; B = agent statement Names the two sides being compared
Matching keys account, value date, payment reference Identifies the same payment on both sides
Pass 1 — exact compare amount, currency, direction; tolerance 0.00 Pairs clean, precise matches first
Pass 2 — tolerance compare amount, currency; allow a flat ±EUR 5 Absorbs trivial rounding and tiny FX
Pass 3 — grouped group by account + value date; match when sums equal Handles one-to-many netted movements
Unmatched route to exception queue "CASH_OPS" Sends every break to the right team
Aging SLA warn at 2 days, escalate at 5 days Sets the clock that drives escalation

Read the tolerance line carefully, because it decides the fate of our running example. The EUR 5 threshold in pass 2 would have quietly absorbed a few cents of rounding. It would not absorb the EUR 750 difference on our example payment — that is 150 times the tolerance. So the payment matches on all three keys but fails every amount pass, and the engine does exactly what it should: it raises a break. No code changed to make any of this happen. The entire behaviour of the reconciliation is set by the configuration above, which is precisely why it can be owned, versioned and audited as a control artifact.

The artifact — a match placed beside a break

Nothing makes the mechanic clearer than seeing two records that should describe the same payment laid side by side. The table below is our running nostro case at the level of the actual fields. Side A is our ledger; side B is the agent's statement. The three keys agree, so the engine is confident these two records are the same payment — but one compared value does not agree, which is what turns a confident pairing into a break.

Field A — our ledger B — agent statement Role
account NOSTRO-EUR-01 NOSTRO-EUR-01 matching key — agrees
value date 2026-07-16 2026-07-16 matching key — agrees
reference PMT-55231 PMT-55231 matching key — agrees
currency EUR EUR compared — agrees
direction CREDIT CREDIT compared — agrees
amount 1,250,000.00 1,250,750.00 compared — BREAK
counterparty CLIENT-ACME CLIENT-ACME classifying — agrees

Read the outcome the way the engine records it. The result is not "unmatched" — it is matched with a difference. The pairing is correct; a single value is wrong. The break field is the amount; the delta is +EUR 750 with side B higher; the break type is a value difference; and the status opens as OPEN, assigned to cash-recon Ops. That is the complete, structured description of a break that flows into the exception workflow — and it is exactly the object an investigator picks up in stage 5. Every field has a job: the keys prove identity, the compared values decide agreement, the classifying fields tell the investigator where to look, and the lifecycle fields carry the break to resolution. All figures illustrative.

Stage 4 — Exceptions are the point, and they come in a few shapes

Whatever survives every matching pass is an exception — a break — and breaks fall into a small number of recognisable types. Naming the type is the first step of every investigation, because the type points at the likely cause.

A record missing on one side exists on A but has no counterpart on B, or the reverse. Something is unrecorded somewhere — a payment that never happened, or one side that has not yet loaded. A value difference is the most common and most investigable break: the records pair up on their keys but a field disagrees — an amount, a quantity, a date. Our EUR 750 case is exactly this. A duplicate or one-to-many mismatch arises when one record should match several, or when duplicates exist, and the matching rule was not smart enough to group them. A timing difference is the benign case: both sides are correct but recorded on different dates because the two feeds were captured at different moments. A timing break is temporary and should self-clear on the next run — which is why aging matters, as we will see.

Most breaks are not errors — they are dirty data

The uncomfortable truth of reconciliation is that the large majority of breaks are false breaks: data differences, not genuine errors. They come from format mismatches that slipped through normalisation, from the same payment carrying different references on each side so the keys never lined up, from cut-off timing where the two feeds were snapped at different moments, and from stale reference or static data — an account or instrument mapping that is out of date on one side. None of these is a real problem with the money or the stock. All of them cost real effort to clear.

False breaks are the biggest lever in the whole discipline for two reasons. They drown the signal — a flood of meaningless breaks buries the handful of genuine ones that actually carry risk, so the control gets worse precisely when it is busiest. And they cost real money — Ops spends most of its time clearing noise instead of investigating the items that matter. The durable fix is never a wider tolerance, which only hides risk. It is cleaner source data and correct reference mapping — fixing the problem upstream so the false break is never generated. Standardising formats, references and static data is a classic data project, and it is where reconciliation transformation creates most of its value.

Stage 5 — Working a break, from open to resolved

A break has its own lifecycle, and the health of a reconciliation is measured by how that lifecycle runs — not by whether breaks exist at all. Breaks are expected and normal; a reconciliation with zero breaks is either trivial or, more likely, badly configured. What matters is how quickly and cleanly each break is investigated, resolved and prevented from recurring.

Figure 4 A break runs from detection to resolution while an aging clock forces escalation — the point is not to avoid breaks but to clear them fast and fix their cause

The break lifecycle has five steps. It opens when the engine raises it, classifies it by type, side and value, and assigns it to an owner or queue. Ops then investigates, gathering evidence — trade details, statements, the history of similar breaks — to work out what actually happened. It categorises the cause against a standard set of codes: timing, our error, their error, a fee, an FX difference, a data or reference issue, a corporate action, or something more serious. It acts on that cause — correcting a booking, chasing the counterparty, raising a claim, or escalating. And finally it resolves or writes off — closing the break when both sides agree, or, for a small residual that cannot be reconciled, writing it off with the required approval.

What makes a break serious is not one thing but a combination. Size — a large value at risk. Age — an old, unexplained break is a red flag in itself. Direction — could we be short of cash or stock, rather than merely holding a harmless surplus? Pattern — a break that recurs points at a systemic problem upstream. And impact — anything touching client money or a reportable position is serious regardless of size. Aging and materiality together drive escalation, which we treat as a governance question in Part 3.

Standard cause codes turn investigation into a language

Categorising a break against a standard set of cause codes is what makes root-cause analysis possible, so it is worth seeing the codes an Ops team actually uses. A shared vocabulary means every resolved break contributes to the same trend analysis, and the code points directly at the fix.

Cause code What it means Typical durable fix
Timing Both sides correct, captured at different cut-offs Align cut-offs; let it self-clear next run
Our error A booking or processing mistake on our side Correct the booking; fix the process that made it
Their error A mistake on the counterparty or agent side Chase and correct; raise a claim if needed
Fee / charge An unbooked agent or correspondent fee Capture the fee schedule; auto-book known charges
FX A currency or rate difference Standardise the FX source and rounding rules
Data / reference A stale or wrong mapping or format Correct the reference data at its governed source
Corporate action A mis-processed dividend, split or entitlement Fix the corporate-action logic and reference data
Duplicate The same movement loaded or booked twice Deduplicate; fix the feed that double-loaded

The power of the codes is cumulative, not individual. One break tagged "fee" is a chore; a thousand breaks tagged "fee" over a quarter is a business case for auto-booking agent charges. The code on a single break tells the analyst what to do now; the distribution of codes across the estate tells the manager what to fix permanently. This is the concrete mechanism behind the claim that reconciliation is a continuous improvement engine — it runs on cause-code data, and a function that does not categorise its breaks consistently is flying blind.

Root-cause analysis turns a daily check into an improvement engine

Clearing a break fixes today; fixing its cause stops it recurring — and that is where reconciliation earns its keep beyond mere compliance. The mechanism is a feedback loop. First, categorise every break with a standard cause code as it is resolved. Second, look for concentrations across those codes over time — which causes, systems, feeds or counterparties generate the most breaks? Third, find the systemic driver behind a recurring cluster, which usually traces to one upstream problem: a bad feed, a mapping gap, a broken process step. Fourth, fix it upstream — auto-book the known fee, correct the mapping, repair the feed — so the whole class of break disappears. Fifth, measure the improvement by watching the match rate and break volume afterward to confirm the cause is gone and the saved effort is real.

A reconciliation that only ever clears breaks, and never fixes their causes, stays expensive forever. The teams that treat cause data as fuel for continuous improvement steadily raise their auto-match rate and shrink their manual pile; the teams that treat each break as an isolated chore run to stand still. The EUR 750 fee break is the miniature of this: cleared once, it is a chore; its cause fixed, it removes a recurring class of false break across hundreds of payments. Multiply that discipline across every recurring cause and it becomes a genuine transformation programme.

Trade and confirmation matching keeps the two sides of a deal honest

Not every reconciliation waits for a statement; some happen at the point of the trade itself, and trade confirmation matching is the most important of them. When two parties agree a trade, each books its own version — its price, size, direction, settlement date and counterparty. If those versions disagree, the trade will fail to settle later, often expensively. Confirmation matching compares the two sides' records of the same trade soon after execution, so a discrepancy is caught in hours rather than surfacing as a settlement fail days later.

The mechanics are a reconciliation in miniature, run forward rather than after the fact. Each side sends its trade details; a matching process — increasingly an electronic platform that both parties connect to — pairs them on trade identifiers and compares the economic terms. A clean match is affirmed and flows to settlement; a mismatch is an exception worked exactly like any other break, by chasing the counterparty and correcting whichever side is wrong. The earlier this happens, the cheaper it is: a price disagreement caught at affirmation is a quick correction, while the same disagreement discovered at settlement can mean a failed trade, a penalty under settlement-discipline rules, and a scramble to fund or borrow.

Confirmation matching also illustrates why coverage matters so much. A trade that is never confirmed — because it was structured to evade the process, as in the Société Générale case in Part 5, or simply because a product type was not covered — is a blind spot, and blind spots are where the largest losses hide. This is why a mature reconciliation function inventories every product and every flow and asks, for each, "where and when is this confirmed and reconciled?" A control is only as strong as its least-covered corner.

The cadence of a reconciliation is a risk decision

How often a reconciliation runs is not an operational detail; it is a statement about how much risk the firm will tolerate between checks. A high-value cash nostro may reconcile intraday, several times a day, because a break left uncaught for hours could mean a funding shortfall or a fraud in flight. A lower-risk intersystem check may run daily. Some reconciliations — certain reference or period-end tie-outs — run monthly. The rule is that the cadence should match the speed at which a break can hurt: the faster a difference can turn into a loss, the more often the reconciliation must run.

Settlement acceleration is forcing that cadence up across the board. When settlement took two days, an overnight batch reconciliation left a comfortable margin to catch and fix a break before it settled. As markets move to T+1, that margin shrinks, and a reconciliation that only runs end-of-day may report a fail after the money has already moved. This is the operational reason behind the industry's push toward intraday and real-time reconciliation, and it is why cadence — not just accuracy — is now a live design question for every high-value control.

Stage 6 — Reporting is the evidence that the control actually ran

The final stage records what happened and turns it into evidence. Every resolved break carries commentary, supporting evidence and a resolution code. The run's outcome — completed on time, breaks cleared, residuals written off with approval — becomes the control evidence that management attests to and that auditors and regulators inspect. The metrics fall out of this same step: the auto-match rate, the break volume and value, the aging profile, the mix of causes. A reconciliation that is performed but not evidenced is, from a regulator's point of view, a reconciliation that did not happen. Evidence is not paperwork bolted on at the end; it is the proof that the control exists.

  1. Name the six stages of the reconciliation lifecycle in order. — Load, normalise, match, identify exceptions, investigate/resolve, report.

  2. Why does the quality of a reconciliation depend on load and normalise more than on the engine?

  3. In what order does a matching engine run its passes, and why?

  4. What is the auto-match rate, and why is chasing 100% dangerous?

  5. Why is a matching rule described as configuration a BA owns rather than code?

  6. What is a false break, and what is the durable fix?

  7. Why is a reconciliation's health measured by how breaks are worked rather than by having none?

  8. What does root-cause analysis add beyond clearing a break?

2

  1. Why does quality depend more on load and normalise than on the engine?

  2. In what order do matching passes run?

  3. Which is the last stage of the lifecycle?

3Independence is an organisation chart, not a slogan

Reconciliation only works because of who does it and who is kept away from it. This part maps the people: the internal functions that run and oversee the control, the external parties that supply the independent other side, and the trading desk that is deliberately left outside the loop. The organising idea is independence — and independence is enforced by reporting lines, data sources and sign-offs, not by good intentions.

Four camps, and the bank's control functions sit between them

The reconciliation ecosystem divides into four camps, and understanding it starts with seeing who is allowed near the control and who is not. The first camp is the internal control functions — Operations, Product Control, Finance, Risk and Compliance, Internal Audit, and the IT and data teams that build the platform. These run and oversee the reconciliation. The second is the trading desk — the source of the trades and positions being checked, and deliberately excluded from checking them. The third is the external holders — the agent banks, custodians, sub-custodians and central securities depositories that actually hold the bank's cash and securities, and the trade repositories that hold derivative reports. These supply side B, the independent record. The fourth is the oversight camp — internal and external audit and the regulators who inspect that the control genuinely works.

Figure 5 Reconciliation Ops sits in the middle: internal control functions run and oversee it, external holders supply the independent 'B' side, and the desk is kept outside the loop

The mental model is simple once the camps are clear. The internal control functions run and oversee the control; the external actors provide the independent other side it checks against; and the front office is intentionally outside the loop. That last point is the whole design. The desk is the source of the records under scrutiny, so letting it near the reconciliation would defeat the purpose. Every arrow of authority and every data feed is arranged to keep the checker separate from the checked.

Every reconciliation is a set of information flows between these camps

The camps are connected by a defined set of flows, and naming them makes the ecosystem concrete. The table reads From / To / What flows / When — the major instruction and information movements a reconciliation depends on.

From To What flows When
Trading desk Books & records Trade and position data (source only) At capture, continuously
Agent bank Reconciliation Ops Cash account statement (side B) Daily / intraday
Custodian / CSD Reconciliation Ops Securities holding statement (side B) Daily
Reconciliation Ops Team lead / manager Escalated aged and material breaks On threshold breach
Product Control General ledger Validated P&L and positions Daily / period close
Bank Trade repository Derivative trade reports (EMIR) On trade, per deadline
Trade repository Bank Pairing and matching results After both sides report
Reconciliation manager Risk, audit, regulator Control evidence, aging, attestation Periodically / on inspection

The teams inside the bank, and where the change work lands

A reconciliation function is not one job but a blend of four: people clearing breaks, controls governance, independent finance, and heavy technology. Knowing who does what tells you who to talk to and where transformation actually happens. The table below is the standard cast around any reconciliation function.

Role Function Core responsibility Primary technology
Reconciliation / Ops analyst Operations Investigates and clears breaks, records cause codes, chases counterparties The exception-management workflow
Recon / controls manager Operations / Controls Owns the recon inventory, aging, escalation and attestation; runs the controls forum Metrics dashboards, attestation tooling
Product controller Finance Independent P&L validation and general-ledger tie-out; P&L attribution P&L and ledger systems
Business analyst (recon) Change / Technology Specifies keys, tolerances, normalisation and workflow; onboards new recons The recon platform configuration
Risk & compliance officer 2nd line Oversees control effectiveness, client-asset rules and regulatory recon GRC and surveillance systems
IT / data engineer Technology Builds feeds, transformations, the matching platform and the metrics pipeline Data pipelines, the platform, APIs

Each role deserves a closer look, because a transformation lands differently on each one.

Reconciliation / Ops analyst

The front line of the control. The analyst picks up the day's breaks, gathers the evidence, works out what happened, records a cause code, and clears or escalates. It is investigative work — part detective, part accountant — and the analyst's accumulated knowledge of why things break is one of the most valuable and least documented assets in the function.

Recon / controls manager

Accountable for the health of the whole estate. The manager owns the inventory of reconciliations, the aging profile, the escalation ladder and the attestation, and runs the controls forum where material and aged breaks are reviewed. When a regulator asks "how do you know this control works?", it is the manager who must answer.

Product controller

The independent finance voice. Beyond validating the desks' P&L, the product controller ties positions and profit back to the general ledger and investigates any unexplained difference through attribution. The role is where reconciliation meets the firm's reported earnings, and its independence from the desk is as non-negotiable as Ops's.

Business analyst (recon)

The core change deliverable. The BA specifies the keys, tolerances, normalisation rules and workflow for each reconciliation and onboards new ones. This is the person who translates "we need to reconcile these two things" into a precise, versioned, auditable configuration — and who, in a no-code world, increasingly configures it directly.

Risk & compliance officer

The second-line overseer. This role does not run reconciliations; it checks that the control environment is effective, owns the regulatory obligations (client-asset rules, EMIR), and forms a view on provisioning and systemic causes. It is the challenge function that keeps the first line honest.

IT / data engineer

Where automation is actually delivered. The engineer builds the feeds, the transformations, the matching platform and the metrics pipeline. Because the module's central claim is that data quality decides everything, this is the role that most directly moves the auto-match rate — by hardening the feeds and mappings the whole control rests on.

The business-analyst and data-engineering roles are where most transformation work sits, and that is worth dwelling on. The Ops analysts clear the breaks the current process produces; the manager keeps the control defensible; the product controller validates the numbers. But the durable improvements — onboarding a new reconciliation, killing a class of false break, raising the auto-match rate, building self-service tooling — are delivered by the people who own the configuration and the data. When a firm decides to transform its reconciliation estate, it is these two roles it staffs up, because change here is a data and configuration exercise, not a trading one.

The three lines of defence place reconciliation precisely

Banks organise their controls into three lines, and knowing which line owns what removes most of the confusion about accountability. The first line is the business that runs the controls day to day — here, Operations performing the reconciliation and clearing breaks. The second line is Risk and Compliance, which oversees whether the controls are designed and operating effectively, owns provisioning judgements, and holds the regulatory obligations. The third line is Internal Audit, which independently tests that the first and second lines are doing their jobs, and reports to the board rather than to the business. External audit and the regulator sit beyond even the third line as the ultimate check.

Reconciliation is a first-and-second-line control, and usually a *key* control — the subset the firm formally relies on for the accuracy of its financial reporting. A key control is documented, owned by a named person, performed on a defined schedule, tested periodically and attested by management. There are no orphan controls: every reconciliation has an owner, a frequency and evidence of operation. Placing reconciliation in this framework matters because it is never designed in isolation — it is one detective control inside a documented control environment, and it must interlock with the preventive controls around it and the testing above it.

Segregation of duties is what gives the control its authority

A control you mark yourself is no control at all, and this is the single most important governance idea in the module. Segregation of duties means the people who reconcile are organisationally separate from the people who trade and who book. The reasoning is stark. If a trader could reconcile their own book, they could hide a loss or a rogue position simply by declaring the inconvenient difference matched. An independent function comparing against an external source cannot be quietly overruled that way. This is exactly why reconciliation lives in Operations, Product Control and Finance — the middle and back office — and never on the desk.

Independence is enforced through four concrete mechanisms, not through a policy statement. Separate reporting lines mean reconciliation and controls report up through Operations and Finance, not through the trading business they check, so the desk cannot pressure the checkers. An independent data source means side B comes from outside the desk — an agent, a custodian, a separate system — so the desk cannot supply the very evidence used to check it. Four-eyes and defined sign-off mean break resolution, and especially write-offs, require a second reviewer with the authority to approve. And access controls with a full audit trail mean system permissions stop unauthorised changes and every action is logged — a SOX-style requirement that makes the control tamper-evident.

Breaks climb a defined escalation ladder, automatically

Independence would mean little if a difficult break could simply be sat on, so breaks climb a defined ladder — and they climb it by rule, not by anyone's decision. The principle is that escalation is automatic and evidence-based: value and age thresholds push a break up the hierarchy without anyone choosing to raise or hide it. That removes the discretion that concealment depends on.

Level Who Role in the break's life
1 Ops analyst Owns and investigates day-to-day breaks; resolves routine ones; records cause codes and evidence
2 Team lead / supervisor Reviews the queue, reallocates, approves write-offs within limits, clears aging warnings before they escalate
3 Management / controls forum A regular forum reviews aged and material breaks, trends and remediation across the book
4 Risk & finance (2nd line) Oversee provisioning, systemic causes, and whether the control environment is effective
5 Audit & the regulator Internal audit independently tests; regulators can inspect the evidence, aging and client-asset compliance

The ladder is what makes the whole control defensible. Clear ownership at every level, plus an audit trail of who did what and when, means the firm can show management, audit and the regulator exactly how any break was handled. A break that crosses an age or value threshold escalates on its own — the SLA in the configuration, from Part 2, is the mechanism — so the escalation cannot be quietly suppressed by the person who would most like it suppressed. This is segregation of duties extended through time: not just separate people at the start, but an enforced path upward that no single person controls.

The controls forum and the indicators that feed it

Independence and an escalation ladder need a place where judgement is actually exercised, and that place is the controls forum. At a regular cadence — often monthly — the reconciliation and controls forum brings together Operations, the controls manager, and second-line risk to review the estate: the aged and material breaks, the trends in root causes, the reconciliations that missed their run, and the remediation actions in flight. It is where a pattern that no single analyst would see — the same cause code climbing across several reconciliations — gets recognised and owned. The forum is the human counterpart to the automated escalation ladder: the ladder pushes individual breaks up, and the forum looks across all of them at once.

The forum runs on key risk indicators, which are the reconciliation KPIs read as early warnings rather than performance scores. A rising 30-day aging bucket, a falling auto-match rate, a reconciliation that has missed its completion window, a cause code that is suddenly dominant — each is a KRI that signals the control environment is drifting before a loss crystallises. The distinction matters: a KPI asks "how well did we do?", while a KRI asks "what is about to go wrong?". A well-run function watches the same numbers through both lenses, and the forum is where the KRI lens is applied with enough seniority to act on what it shows.

The operating model decides where and how the work is done

Who runs reconciliation, and from where, is a strategic choice as much as an organisational one. The routine work — clearing predictable breaks, chasing counterparties, producing evidence — is high-volume and process-driven, which is why banks have long moved it to lower-cost locations: offshore and nearshore operations centres that handle day-to-day break management while the judgement-heavy roles (the controls manager, product control, the business analyst) stay closer to the business. The target operating model, or TOM, for reconciliation is essentially a decision about which activities sit where, and how the hand-offs between them are controlled.

Three sourcing models sit on a spectrum, and each trades control against cost and scale. Keeping everything in-house gives the most control and the least leverage of scale. A managed service, where a vendor runs the reconciliation platform and sometimes the break investigation on the bank's behalf, trades some control for scale and speed — attractive, but subject to the outsourcing and operational-resilience rules that govern any critical third party. A mutualised utility, where several banks share one reconciliation infrastructure for a common problem, promises the greatest efficiency on standardised, non-differentiating recons, at the cost of bespoke control. Most large banks run a blend, keeping differentiating and client-asset reconciliations close and pushing commoditised ones toward shared or managed models.

Whatever the model, accountability cannot be outsourced. A bank can move the operation of a reconciliation offshore or to a vendor, but it remains accountable to its regulator for the control working — which is why sourcing decisions always come wrapped in service-level agreements, oversight, and the same evidence and attestation the control would need in-house. The lesson mirrors the technology one from Part 4: the tool or the location can change, but ownership of the control stays with the bank.

Product control extends reconciliation from cash to profit

Reconciliation is not only about cash and stock; the same discipline is applied to the trading desks' profit, and that is what product control does. Product control is the independent finance function that owns and validates the desks' P&L. Its central task is a reconciliation of profit: the desk's own estimated, or flash, P&L against the independently calculated actual P&L from the books. A difference between the two is a break, exactly like a cash break, and it is worked the same way.

The tool that explains a P&L break is called P&L attribution, and it is root-cause analysis by another name. Attribution decomposes the difference into its causes — market moves, new trades, fees, and adjustments — until the profit is fully explained. An unexplained P&L difference that survives attribution is a genuine break, and it can reveal a mismarked position, a booking error, or a rogue trade. The parallels to everything in Part 2 are exact: product control is independent of the desk, so the trader cannot mark their own profit unchallenged; it explains rather than merely matches; it feeds validated numbers to the general ledger; and it catches error and misconduct that no preventive control would stop. Product control is where reconciliation and the firm's reported earnings meet.

  1. What are the four camps in the reconciliation ecosystem?

  2. Why is the trading desk deliberately kept outside the reconciliation loop?

  3. Which two roles carry most of the transformation work, and why?

  4. Where does reconciliation sit in the three lines of defence?

  5. Name the four mechanisms that enforce segregation of duties.

  6. Why is automatic, threshold-based escalation important?

  7. How is product control a reconciliation, and what technique explains its breaks?

  8. Name the three sourcing models for the reconciliation operating model, and what stays fixed across them. — In-house, managed service, and mutualised utility; accountability to the regulator for the control working stays with the bank.

3

  1. Why is the trading desk kept outside the reconciliation loop?

  2. Which roles carry most transformation work?

  3. Who supplies side B of a reconciliation?

4Under the control sits one of the bank's biggest data problems

Reconciliation is a data-engineering discipline wearing a control's clothes. This part opens the machinery: the functional stack a reconciliation flows through, the vendors who supply it, the lineage of a trade front to back, the anatomy of the record and the break, the identifiers and messages that tie records together, the integration patterns that carry the data, and the regulations that specify how the whole chain must behave.

Ask what functions a reconciliation needs before naming any vendor

Before any product name, the clean question is: what must any reconciliation platform do, from raw feed to evidenced control? The answer is five functional blocks in a chain, resting on a set of cross-cutting foundations. Getting the functional picture first means you can judge any vendor against what the work actually requires, rather than being led by a feature list.

Figure 6 The recon platform stack: five functional blocks from ingestion to reporting, resting on reference data, workflow, access control and resilience
Functional block What it does Where it breaks
Data ingestion Pulls both sides from files, feeds, APIs and messages on a schedule A late, missing or changed file breaks the whole run
Normalisation Maps, cleanses and standardises fields, formats, signs, currencies A silent upstream format change manufactures false breaks
Matching engine Applies keys and rules in passes; auto-matches; raises exceptions Poor keys or loose tolerances degrade match quality
Exception management Assigns, investigates, comments, resolves, escalates; ages breaks Weak workflow lets breaks age unowned
Reporting & metrics Dashboards, KPIs, attestation and audit evidence for risk and finance No evidence means the control cannot be proven

Underneath all five blocks sit the foundations that make the difference between a healthy estate and a fragile one. Reference and static data — the account and instrument mappings — feed every layer, and a gap there blocks matching everywhere. A workflow and audit trail records who did what, so a break is traceable back through matching and normalisation to its source. Access control and segregation of duties keep the checker separate from the checked. And operational resilience keeps the platform running on the days it cannot be allowed to fail. These are not add-ons; they are what turns a matching tool into a control.

The vendor landscape is a map of buy-versus-build decisions

A handful of platforms run most of the market's reconciliation, and each is known for something specific. The buy-versus-build pattern here is itself informative: banks buy the matching engine and the workflow, because they are commodities that no bank gains advantage from writing itself, but they keep the configuration — the keys, tolerances and reference-data mappings — in-house, because that is where the institution-specific knowledge lives. The table maps the names an analyst will meet.

Platform Known for Position
SmartStream TLM Enterprise cash and securities recon at high volume The long-standing incumbent; deep, powerful, a large estate to modernise
Duco Cloud, self-service, no-code recon The poster child for business-owned, fast-to-change reconciliation
Gresham Clareti (CTC) Real-time transaction recon and a control framework Strong in cash, positions and integrity control
Broadridge Recon inside a wide post-trade platform Often bundled with the broader operational stack
FIS (and peers) Recon modules in large banking and markets suites IntelliMatch heritage and estate tooling
In-house & spreadsheets Bank-built tools and, dangerously, spreadsheets Legacy and tactical — the prime target for consolidation

The last row is the important one, because it is where most of the risk and most of the opportunity sit. A large share of reconciliations still run on bank-built tools or on spreadsheets maintained by a single analyst — slow, undocumented, error-prone, and a key-person and audit risk all at once. The transformation opportunity is consistent across banks: consolidate scattered reconciliations onto a modern, self-service platform; harden the feeds and normalisation to kill false breaks; automate the exception workflow, aging and attestation; and then, on top of clean data, add analytics and machine learning. In Europe this shift from on-premise incumbents toward cloud, self-service tools is a live trend, shaped by data-residency, outsourcing and operational-resilience rules that constrain which cloud services can be used and how.

Choosing and running the platform in practice

The vendor table names the players; choosing between them is a judgement about volume, change appetite and control maturity. A firm running enormous, stable, high-volume cash and securities reconciliations may value the depth and scale of an enterprise incumbent even at the cost of slower change. A firm whose pain is the speed of change — a long backlog of reconciliations waiting to be onboarded or tweaked — will lean toward a no-code, self-service platform where the business owns the rules. Most large banks end up with a mix, and the transformation question is less "which single tool?" than "how do we consolidate a sprawling estate onto as few, as governed, as self-service platforms as possible?"

Whatever the tool, the same operating disciplines decide whether it succeeds. Every reconciliation needs a named owner, a documented purpose, and a defined cadence; the reference data it depends on needs a single governed source; the feeds it consumes need monitoring so a late or malformed file is caught before it breaks the run; and the whole estate needs an inventory so that coverage can be proven and no reconciliation runs unowned. A powerful platform run without these disciplines is a fast way to produce ungoverned reconciliations; a modest platform run with them is a genuine control. The tool is necessary but never sufficient — the operating model around it is what makes it work.

Matching a million records fast is an engineering problem

The idea of matching is simple; doing it on tens of millions of records within a batch window is not, and the engineering behind it is worth seeing. A naive engine that compared every record on side A against every record on side B would perform a million times a million comparisons — impossibly slow. Real engines avoid that by keying and indexing: they compute a key from the matching fields and group records by that key, so only records that share a key are ever compared. This turns an impossible all-against-all problem into a fast lookup, and it is why the choice of matching key is a performance decision as much as a logical one.

Fuzzy and grouped matching are where the cost returns, and they are managed with a technique called blocking. Because fuzzy matching cannot rely on an exact key, a careless implementation would again drift toward comparing everything with everything. Blocking narrows the field first — grouping records into plausible candidate sets (say, same account and same day) and running the expensive approximate comparison only within each block. The engine trades a little completeness for a large gain in speed, which is acceptable precisely because fuzzy matching is the last resort after the exact and tolerance passes have already cleared the bulk.

These constraints shape the whole platform, not just the engine. Batch windows are finite — a reconciliation that must finish before the morning has only so many hours — so throughput, parallelism and the efficiency of the keys decide whether it completes on time. As volumes grow and cadences move intraday, the engineering shifts from overnight batch toward streaming and incremental matching, where each new record is matched as it arrives rather than reprocessing the whole population. The performance engineering is invisible when it works and business-critical when it does not: a matching run that overruns its window is a control that did not complete.

Follow one trade front to back, because errors cascade downstream

A single trade is represented in many systems, and reconciliation is what keeps them telling the same story. The trade is captured at the front office, valued in the risk system, recorded in the product sub-ledger, posted to the general ledger, and reported to regulators. Each system may hold the trade slightly differently — different fields, different timing, different transformations — and every feed between them is a place where data can be lost, duplicated or mangled. Intersystem, or front-to-back, reconciliation checks that trade counts and positions agree system to system, that risk and the ledger reflect the same book, and that nothing is dropped or transformed wrongly as it flows down.

Figure 7 A trade front to back: recon checkpoints at each junction catch an error before it reaches the ledger and the regulatory report

The principle that makes lineage matter is the cascade: an error upstream reaches every downstream system. A wrong booking at the front does not stay at the front — it flows into risk, into the sub-ledger, into the general ledger, and into the regulatory report, corrupting each in turn. Because the systems copy from one another, they will happily agree with each other while all being wrong. Only a reconciliation against an independent source at each junction breaks the cascade. This is also why regulatory and financial reports are only ever as trustworthy as the front-to-back consistency behind them: the report at the end of the chain inherits every unreconciled error along it.

The anatomy of the record — and why deduplication is hard

Everything the engine does turns on the fields of a single record, so it is worth opening one up. A reconciliation record — a cash movement, a position line, a trade — carries a handful of fields that each drive something downstream. The matching keys (account, value date, reference) decide which records pair. The compared values (amount, quantity, currency, direction) decide whether a pair agrees. The classifying fields (counterparty, instrument, product) route a break to the right queue and tell an investigator where to look. And the lifecycle fields (status, owner, age, cause code, resolution) carry the break through its workflow. A missing or wrong value in any one of these does not stay local; it changes which records match, which break, and where the break goes.

Laid out as a table, the record's fields divide into four jobs, and knowing which job a field does tells you what a wrong value will break. This is the anatomy an analyst configuring a reconciliation works with directly.

Field group Example fields What it drives What a wrong value causes
Matching keys account, value date, reference Which records pair A missed or false pairing — the record breaks against the wrong item or none
Compared values amount, quantity, currency, direction Whether a pair agrees A value-difference break, or a real error absorbed by too-wide tolerance
Classifying fields counterparty, instrument, product Routing and investigation A break sent to the wrong queue or investigated with the wrong context
Lifecycle fields status, owner, age, cause code, resolution The break's workflow A break that ages unowned, or clears without evidence

Deduplication is the deceptively hard problem hiding inside all of this. When one payment on side A legitimately corresponds to several items on side B, or when the same movement appears twice because a feed was loaded twice, the engine must decide what is one economic event and what is an accidental copy. Get it wrong in one direction and you match records that do not belong together, hiding real breaks; get it wrong in the other and you raise breaks that are pure artefacts of duplication. Reliable deduplication depends on good keys and clean reference data — which is why, once again, the data foundation decides the health of the control.

Immutable audit and the configuration as the real asset

Two properties separate a reconciliation platform from an ordinary data tool: its audit trail is append-only, and its configuration is the institution's genuine intellectual property. A key control must be able to prove, after the fact, exactly what it did and who touched it — which means every action on a break, every rule change, every sign-off is recorded in a log that cannot be edited or deleted, only added to. Immutability is not a technical nicety; it is what makes the control's history admissible to an auditor or a regulator. A platform whose records could be quietly altered would be a control that certifies nothing.

The configuration — the keys, tolerances, normalisation rules and reference mappings — is where the value concentrates, and it is why banks guard it even when they buy the engine. The matching engine itself is a commodity; any vendor's engine will pair records against rules. What is not a commodity is the accumulated knowledge of how this institution's data actually behaves — which references drift, which agents net, which instruments break every dividend season, what tolerance is safe on which account. That knowledge, encoded as configuration and reference data, is the real asset, and it is why the buy-versus-build line falls where it does: buy the engine, own the rules. A firm that outsourced its configuration would be handing away the one part of the control that competitors cannot replicate.

Changing a reconciliation safely is a control in its own right

The module has argued that a reconciliation's rules are configuration a business analyst owns — which raises a sharp question the earlier parts left open: how do you change the rules of a key control without breaking the control? A tolerance widened by mistake, a key mistyped, a mapping changed carelessly — any of these can silently stop the reconciliation catching what it should, and because the control still runs, the failure is invisible until something goes wrong. So changing a reconciliation is itself a governed process, with the same rigour the bank applies to code changes elsewhere.

The safe-change discipline has a recognisable shape. A proposed change is specified and reviewed; it is tested in a non-production environment against known data, including deliberately seeded breaks, to prove it still catches what it must and does not manufacture new false breaks; it is signed off under four-eyes; it is versioned so the exact configuration in force on any date can be reconstructed; and it is released in a controlled way, often with the old and new logic run in parallel for a period to confirm the change behaves as intended. Each step has a purpose: the test proves the control still works, the version history makes it auditable, and the parallel run catches surprises before they matter.

No-code platforms make change faster, which raises the stakes on this discipline rather than lowering them. When a business user can alter a live reconciliation in minutes, the guardrails — testing, sign-off, versioning — become the only thing standing between agility and an ungoverned control. The maturity of a reconciliation function shows less in how quickly it can change a rule than in how safely it can, and a regulator inspecting a key control will ask to see exactly this change history. Speed without the safe-change discipline is not transformation; it is risk.

Reference and static data is the quiet foundation of every match

Underneath the transactional flow sits a slower-moving layer that decides whether any of it can match: reference and static data. Where transactional data is the stream of movements — payments, trades, positions — reference data is the relatively stable information that gives those movements meaning: which internal account maps to which external nostro, which instrument code on our side corresponds to which ISIN on the custodian's, which counterparty short-name resolves to which LEI. A matching engine cannot pair two records if it does not know that "NOSTRO-EUR-01" on our side and the agent's own account label refer to the same account. That knowledge lives in reference data, and when it is wrong or missing, matching fails silently.

Reference data is dangerous precisely because it drifts slowly and invisibly. A counterparty is renamed after a merger; a custodian re-tiers an account; an instrument is reorganised in a corporate action. Each change is small, and each one, if not reflected in the mapping, turns a stream of clean matches into a stream of false breaks overnight — with no obvious cause, because nothing in the transactional data changed. Fragmented reference data, held in many places and maintained by many hands, is one of the most common hidden causes of matching failure across a whole estate. This is why a serious reconciliation transformation almost always includes a reference-data programme: consolidating the mappings into a governed, single source is often worth more than any change to the engine itself.

Normalisation maps to a canonical model, fed by master data

Behind the plain word "normalise" sits a specific architectural idea: both sides are mapped onto a single *canonical model* — one agreed internal representation of a payment, a position or a trade — rather than each reconciliation inventing its own field names and formats. When every source maps to the same canonical shape, a new reconciliation reuses mappings that already exist, and a fix to one mapping benefits every reconciliation that uses it. Without a canonical model, each reconciliation re-solves the same translation problem in its own way, and the estate accumulates dozens of incompatible interpretations of the same data — a hidden driver of both cost and false breaks.

The canonical model is only as good as the reference data feeding it, which is why mature functions treat that reference data as a managed master. Master data management (MDM) is the discipline of maintaining a single, governed "golden source" for the account, instrument and counterparty mappings, rather than letting each system keep its own drifting copy. The connection to reconciliation is direct: the golden source is what lets the canonical model resolve "this account on our side is that account on theirs," and when it is fragmented, matching fails in ways no amount of engine tuning can fix. The unglamorous work of MDM is, once again, where the leverage sits — a theme this module keeps returning to because it keeps being true.

Data quality has dimensions, and each one breaks matching differently

"Clean data" is vague; the useful version breaks data quality into dimensions, because each dimension fails in its own way and needs its own fix. An analyst diagnosing a troubled reconciliation runs down this list, because the remedy depends entirely on which dimension is at fault.

Dimension What it means How it breaks a reconciliation
Completeness Every expected record is present A missing record breaks as "unmatched on one side"
Accuracy Each value is correct A wrong amount or quantity breaks as a value difference
Consistency The same thing is represented the same way on both sides Format or reference mismatches create false breaks
Timeliness Data arrives within the reconciliation window A late feed breaks the run or creates timing differences
Uniqueness No unintended duplicates Duplicates create one-to-many and phantom breaks
Validity Values conform to expected formats and ranges Malformed data fails normalisation and cannot match

Reading a break through this lens is faster than guessing. A cluster of "missing on one side" breaks points at completeness or timeliness — a feed problem. A cluster of value differences points at accuracy — a genuine error, or a tolerance set wrong. A flood of breaks that appear identical points at consistency — a normalisation or reference-data gap. The dimension names the fix, and the fix is almost always upstream. This is the disciplined version of the claim made throughout the module that data quality is the biggest lever: it becomes actionable once you can say which quality is failing.

Identifiers are the fields that let records find each other

Records from different systems can only be matched if they name the same thing the same way, and that is what standard identifiers provide. Three matter most. The ISIN (International Securities Identification Number) identifies the security — the specific bond or share — so that a position on our books and a holding at the custodian can be recognised as the same instrument. The LEI (Legal Entity Identifier) identifies the legal entity — the counterparty, the client, the fund — so that a trade can be tied to the right party across systems and reports. And the UTI (Unique Transaction Identifier) identifies the derivative trade itself, and is the key on which a trade repository pairs the two sides' reports under EMIR. Where these identifiers are clean and consistent, matching is easy; where they are missing or inconsistent, the keys fail to line up and false breaks multiply.

Identifier Names Where it bites in reconciliation
ISIN The security (bond, share) Pairs our position against the custodian's holding for the same instrument
LEI The legal entity (counterparty, client) Ties a trade or account to the right party across systems and reports
UTI The individual derivative trade The key a trade repository uses to pair both sides' EMIR reports

Connectivity and observability decide whether the data even arrives

Before any matching happens, the data has to get in — and how it arrives, and whether anyone notices when it does not, is a first-order operational concern. Reconciliation feeds travel over a handful of channels: managed file transfer (SFTP and its enterprise equivalents) for scheduled files, message queues (MQ) for guaranteed delivery between systems, streaming buses (Kafka-style) for real-time events, and REST APIs for on-demand pulls. Each has its own failure modes — a stalled file transfer, a backed-up queue, a broken API — and each is a place a reconciliation can fail before it even starts.

This is why observability matters as much as the pipeline itself. A mature platform does not wait to discover a missing feed when the reconciliation comes up short; it monitors the feeds actively — expecting a file by a certain time, alerting when one is late, malformed or the wrong size — so operations can chase the source before the run is affected. The same discipline watches the runs themselves: whether they completed, how long they took, whether volumes moved unexpectedly. Feed fragility was named in Part 2 as the most common operational cause of a failed reconciliation; observability is the direct answer to it, turning a silent failure into an early alert. A control you cannot see failing is a control you cannot rely on.

How the data moves — batch, API, streaming, and the push to intraday

The reconciliation you can run depends on how the data reaches you, and there are three integration patterns in ascending order of modernity. Batch is the legacy default: files delivered on a schedule, typically overnight. It is simple and universal, but its weakness is latency — you only know about a break the morning after, and a late file stalls everything. API is the modernisation: data pulled on demand, when you need it, rather than waiting for a file. Event streaming is the frontier: systems push each movement as it happens onto a message bus, so the reconciliation sees data in real time. Each step up the ladder shortens the gap between an event and its check.

Pattern How data arrives Strength Weakness
Batch Files on a schedule, typically overnight Simple, universal, works with any counterparty High latency; a late file stalls the whole run
API Pulled on demand when needed Timelier; no waiting for a scheduled file Both systems must expose and consume the interface
Event streaming Pushed movement-by-movement onto a bus Real-time; systems stay consistent by construction Requires modern infrastructure on both sides

Cloud, tenancy and data residency reshape where reconciliation runs

The shift from on-premise incumbents to cloud and SaaS reconciliation is not only a delivery-model change; it raises specific questions a bank must answer before it can move a control to the cloud. A SaaS reconciliation platform is typically multi-tenant — many clients share one system, with their data logically separated — which is efficient but forces hard questions about isolation and security. Reconciliation data is among the bank's most sensitive: it contains account numbers, counterparty identities and, in client-asset recons, client personal data, so data residency (where the data physically sits) and data protection (GDPR obligations over that personal data) become gating constraints on which cloud services can be used and in which region.

Two controls sit on top of the platform regardless of where it runs: encryption and entitlements. Data must be encrypted both in transit and at rest, so that a compromised channel or store does not expose account and client information. And entitlements — fine-grained access rules — must ensure each user sees only the reconciliations and data they are authorised for, which matters especially where one platform serves many desks, entities or client segments. These are the concrete mechanisms behind the "access control and segregation of duties" foundation named in the architecture: in a cloud, multi-tenant world they are not optional hardening but the precondition for putting a sensitive control there at all — and, as the resilience discussion below notes, DORA scrutinises exactly these third-party cloud dependencies.

The messages that carry the two sides

Reconciliation data arrives as standardised financial messages, and two worlds dominate the post-trade layer that feeds cash and position recon. The incumbent standard is SWIFT MT — structured messages exchanged between banks and agents, including the account statements that form side B of a nostro reconciliation. These are migrating to ISO 20022 (MX), a richer, XML-based standard that carries far more structured data per message, which improves matching because more fields arrive clean and machine-readable rather than buried in free text. The table shows the layers an analyst meets.

Layer Standard Carries Timescale
Bank-to-agent statements SWIFT MT940 / MT950 End-of-day and interim cash account statements (side B for nostro) Daily / intraday, migrating to ISO 20022
Cash reporting (ISO 20022) camt.052 / .053 / .054 Intraday, end-of-day and debit/credit notifications Real-time to daily
Securities / settlement ISO 15022 / 20022 (sese, semt) Settlement instructions and custody holding statements (side B for position) Around settlement, daily
Derivative reporting ISO 20022 (EMIR REFIT) Trade reports to the repository, paired on the UTI Per reporting deadline

The richer the message standard, the healthier the reconciliation, which is why the ISO 20022 migration matters to this discipline specifically. A EUR nostro reconciliation fed by structured camt statements has more clean fields to match on than one fed by an MT940 whose narrative line has to be parsed for the reference. More structured data means better keys, fewer false breaks, and a higher auto-match rate — the migration is a data-quality upgrade that lands directly on the control.

The messaging sequence, end to end

It helps to trace how a nostro statement actually reaches the reconciliation, because the message path is where much of the latency and fragility lives. During the day, the agent bank processes payments in and out of the firm's nostro account. It sends intraday notifications — increasingly ISO 20022 camt.054 debit/credit messages — as movements occur, and at end of day it sends a full statement, historically a SWIFT MT940, now migrating to a camt.053. The firm's platform ingests that statement as side B, aligns it against the internal ledger as side A, and runs the reconciliation. On the securities side the pattern mirrors this: settlement instructions and confirmations flow as ISO 15022 or 20022 messages, and the custodian's holding statement arrives as the position side B.

The sequence exposes two structural weaknesses that reconciliation must absorb. The first is timing: intraday notifications and the end-of-day statement snapshot the account at different moments, so a movement captured in one but not yet the other produces a timing difference that should self-clear — but only if the reconciliation understands the cut-offs. The second is format drift: as agents migrate from MT to ISO 20022 on their own timetables, a firm may receive the old and new formats from different agents simultaneously, and every format the platform must parse is another place a silent change can manufacture false breaks. The messaging migration is therefore both an opportunity — richer, cleaner data — and a live operational risk to be managed through the transition.

Regulatory reporting reconciliation is a family of its own

EMIR is the example this module has used, but derivative reporting is only one member of a whole family of reconciliations whose "other side" is a regulator or a reporting mechanism. Modern regulation requires banks to report vast volumes of transactions, and a report is only useful if it is complete and accurate — so the bank must reconcile what it reported against what it should have reported, and against what the receiving mechanism accepted. Under EMIR, both counterparties report each derivative and the trade repository pairs and matches them on the UTI. Under MiFIR transaction reporting, the bank reports executed transactions to an approved reporting mechanism (ARM) and must reconcile its submissions against its own books and against the ARM's acknowledgements, chasing every rejection. Under SFTR, the same discipline applies to securities-financing transactions.

The mechanics are ordinary reconciliation; the stakes are regulatory data quality. Each of these is two records that should agree — what we reported versus what we should have, or what the repository holds versus our own record — worked as breaks against reporting-specific cause codes: a missing report, a rejected report, a field that fails validation, a pairing that does not match the counterparty's. Supervisors increasingly judge a firm by its reporting quality metrics — completeness, accuracy, pairing and matching rates — and poor numbers attract exactly the fines and remediation orders described in Part 5. Reconciliation is therefore not only a control over the bank's own books; it is the control that keeps the bank's regulatory reporting honest, and it is a large and fast-growing part of the estate.

Operational resilience is where "mission-critical" becomes measurable

Some reconciliation moments have no second chance, and that is what makes resilience a first-order concern rather than an IT footnote. A missed client-money reconciliation, a settlement-day position check that does not run, a period-close tie-out that fails — these happen on days when the control simply must operate, and cannot be deferred to tomorrow. Resilience is the set of guarantees that keep the platform running through failure, and the terms have precise meanings.

Concept What it guarantees Why it matters to reconciliation
Redundancy No single point of failure; components are duplicated The platform survives the loss of one server, feed or site
Automatic failover Traffic switches to a standby without manual intervention The reconciliation still runs when primary infrastructure fails
Disaster recovery A tested ability to restore service at another site A regional outage does not stop the close or the client-asset check
Business continuity People and processes keep operating through disruption Breaks still get worked when the office or a key vendor is down

In Europe, resilience is now codified rather than merely expected. The Digital Operational Resilience Act (DORA) treats reconciliation platforms and their feeds as critical operations, raising the bar on resilience testing, incident response and third-party risk — which matters directly because so many reconciliations depend on external cloud vendors and data feeds. Redundancy alone is not enough: a duplicated system that has never been failover-tested is a hope, not a guarantee, and DORA is explicit that the capability must be rehearsed, not assumed.

Five regimes shape every system in the chain

Reconciliation is not merely good practice; a cluster of regulations *mandates* it, and each reads as a specification for the systems, not just a legal summary. Understanding them as system requirements is what lets an analyst design a platform that is compliant by construction.

Client-asset protection (CASS and EU equivalents)

Requires both an internal reconciliation, proving the firm's own records of client money and assets are consistent, and an external one, proving those records agree with the banks and custodians actually holding the assets. Crucially, if a shortfall is found, the firm must fund it from its own money until it is resolved — so the system must not only detect a client-asset break but flag it for immediate funding and prompt remediation.

Books-and-records accuracy

The foundational obligation that firms keep accurate records; reconciliation is the evidence that the books reflect reality. The system requirement is coverage — every account and product inventoried and reconciled, with no blind spots, and the inventory itself attested.

SOX-style internal control

Treats reconciliation as a documented key control: owned, performed on schedule, tested and attested for the reliability of financial reporting. The system requirement is an immutable audit trail and enforced sign-off, so that the control's operation can be evidenced after the fact.

EMIR pairing and matching

Requires both counterparties to report each derivative to a trade repository, which pairs the two reports on the UTI and matches their fields; firms must keep pairing and matching rates high and fix breaks. The system requirement is a self-reconciliation of the firm's own reporting against the repository and the counterparty — reconciliation applied to regulatory data itself.

Operational resilience (DORA)

Requires that the reconciliation platform and its feeds, as critical operations, meet raised expectations on resilience, testing and third-party risk. The system requirement is redundancy, tested failover and managed vendor dependencies.

Across all five, national regulators such as the AMF and ACPR in France, and ESMA at EU level, inspect the reconciliation evidence, the aging and client-asset compliance directly. Poor pairing rates and aged client-asset breaks attract fines and remediation orders. The through-line is that none of these regimes asks for a document at the end; each asks for a property of the system — coverage, evidence, an audit trail, tested resilience — which is why a platform designed around them is compliant by construction rather than by retrofit.

  1. Name the five functional blocks of a reconciliation platform. — Data ingestion, normalisation, matching engine, exception management, reporting and metrics.

  2. In the buy-versus-build pattern, what do banks keep in-house and why?

  3. What is the cascade principle, and how does reconciliation break it?

  4. What do ISIN, LEI and UTI each identify?

  5. Why does reference data cause breaks even when transactional data is correct?

  6. Why is moving from batch to event streaming more than a speed upgrade?

  7. Why is redundancy alone insufficient for operational resilience?

  8. Under CASS, what must a firm do if a client-asset reconciliation reveals a shortfall?

  9. Why is changing a reconciliation's configuration itself a governed process?

  10. How does a matching engine avoid comparing every record against every other?

  11. Beyond EMIR, name two other regulatory-reporting reconciliations.

4

  1. What do banks keep in-house?

  2. How does reconciliation break the cascade?

  3. Which is NOT one of the five blocks?

5When reconciliation is weak, the losses are real

This part turns from how reconciliation works to what happens when it does not. It maps the risk landscape as a single taxonomy, sets out the metrics that reveal whether the control is healthy, examines three real cases where weak reconciliation turned into catastrophic loss, counts the true cost of a break, and then looks forward to the transformation — clean data, no-code tooling and machine learning — that is reshaping the discipline.

Reconciliation risk falls into six types, and the taxonomy is a map of controllability

The risks around reconciliation are not a random list; they form a taxonomy, and the order of that taxonomy tells you how much control you have over each one. Some risks are directly within reconciliation's power to catch and fix; others are systemic and reconciliation can only ever be one defence among several. Seeing the six types as a spectrum from controllable to systemic is more useful than memorising them, because it tells you where to spend effort.

Figure 8 The six reconciliation risk types, ordered from those the control can directly manage to those that are more systemic
Risk type What it is Who owns it Main defence
Operational / settlement A break from a failed settlement or a processing error Operations Timely reconciliation and prompt break resolution
Financial / funding Funding cost of a cash break, provisions, write-offs Finance / Treasury Fast clearance and aging discipline
Data quality False breaks that drown the signal and waste effort Change / IT Clean upstream data and reference mapping
Regulatory / compliance A CASS breach or an EMIR pairing failure 2nd-line Risk & Compliance Complete coverage, evidence, prompt remediation
Fraud & conflict A rogue position or misappropriation hidden by false matching 1st & 2nd line together Segregation of duties, independent data
Reputational Fines and public failure that outlast the money lost The whole firm Every control above, operating well

The KPIs are the dashboard that tells you whether the control works

Reconciliation is measured by a small set of metrics, and each answers a different question about the control's health. Management, audit and the regulator all read this dashboard, so an analyst who wants to improve a reconciliation must know what each number means and how they interact.

The auto-match rate — the percentage matched with no human touch — measures efficiency; higher is cheaper and faster. Break volume and value measure the raw workload and the money at risk. The aging profile — how long breaks stay open, bucketed as 0–2 days, 3–5 days, 6–30 days, 30 days and over — is the single best indicator of control health, because an aging pile is where hidden losses accumulate. Timeliness and completion ask whether every reconciliation ran on time for the daily or period close; a missed reconciliation is a control gap regardless of the match rate. The root-cause mix shows where the systemic problems are and where improvement effort should go. And operational-risk losses and write-offs show whether breaks are actually turning into money lost, feeding the firm's operational-risk capital.

Laid out with indicative targets, the dashboard becomes something an analyst can actually manage against. The ranges below are illustrative — every reconciliation has its own realistic benchmarks — but they show the shape of a healthy control.

Metric What good looks like (illustrative) What it warns of when it slips
Auto-match rate 95%+ on clean cash; 80%+ on harder recons Rising manual effort, or tolerances set too loose
Break volume Stable or falling week on week A new upstream data or feed problem
Aging — 30d+ bucket Near zero, tightly justified Hidden losses accumulating unexplained
Completion / timeliness 100% of recons run on time A control gap or coverage hole
Root-cause concentration No single cause dominating A systemic upstream driver to fix
Write-offs Small, rare, fully approved Breaks turning into real losses

The metrics are only meaningful read together, because each can be gamed alone. A high auto-match rate with a growing 30-day bucket is a control being flattered while it quietly rots. Perfect completion with a rising write-off trend means the recons run but do not protect. The dashboard works as a system: efficiency (auto-match), risk (volume, aging), reliability (completion), and whether you are actually fixing things (root-cause mix, write-offs). A transformation programme is judged by whether it moves this whole system, not one flattering number on it.

Three real cases, three ways reconciliation fails

When reconciliation is weak, ignored or bypassed, the losses are not hypothetical — and three real cases map the three distinct ways the control fails. They are not random disasters; each isolates a different failure mode, and together they cover the space.

Barings Bank, 1995 — the control had no independence

The failure mode is a collapse of segregation of duties. At Barings' Singapore operation, a single trader, Nick Leeson, controlled both the trading desk and the back office that was supposed to check it. He hid mounting losses in an error account — the now-infamous account 88888 — and, because he also ran the settlement and reconciliation function, he could reconcile his own book and make the breaks disappear. The losses grew unchecked to around £827 million and destroyed a 233-year-old bank. There was a reconciliation; it caught nothing, because the person concealing the problem was also the person certifying there was none. This is the textbook demonstration that a control you mark yourself is no control at all.

Société Générale, 2008 — the control was evaded by concealment

The failure mode is a detective control defeated by someone who understood exactly how it worked. A trader on the Paris arbitrage desk, Jérôme Kerviel, built enormous unauthorised directional positions and masked them with fictitious offsetting trades. Because he had previously worked in the middle office, he knew precisely which controls and reconciliations would look at which fields and when, and he timed and structured the fake trades to evade them — cancelling and re-entering them before they could be confirmed or reconciled with counterparties. When the positions were finally unwound, the loss was about €4.9 billion. The lesson is uncomfortable: reconciliation that checks only what it always checked can be gamed by an insider who knows its blind spots, which is why coverage, confirmation matching and behavioural surveillance must reinforce it.

JPMorgan client-money failure, 2010 — the control failed a regulatory obligation

The failure mode is a client-asset reconciliation that did not protect what it was meant to protect. The UK regulator fined J.P. Morgan Securities £33.32 million for failing to properly segregate client money over a period of nearly seven years — money that should have been ring-fenced and reconciled as client assets was instead left mixed with the firm's own funds. No client ultimately lost money, but the control obligation itself was breached, and the fine was, at the time, the largest the regulator had levied. The lesson is that client-asset reconciliation is a legal duty in its own right: a firm can be penalised heavily for the failure of the control even when no loss crystallises, because the whole point is to protect clients before anything goes wrong.

Case Failure mode What reconciliation should have caught Lesson
Barings (1995) No independence A trader reconciling his own hidden loss account Segregation of duties is non-negotiable
Société Générale (2008) Control evaded Fictitious offsetting trades that never confirmed Coverage and confirmation matching must reinforce recon
JPMorgan (2010) Regulatory breach Client money not segregated and reconciled Client-asset recon is a legal duty; failing it is itself the loss

Four recurring patterns turn weak reconciliation into loss

Step back from the individual cases and the same failure patterns recur across the industry, which is what makes them worth naming. Hidden rogue positions — in the famous trading-loss cases, weak or bypassed reconciliation let unauthorised positions grow undetected until they were catastrophic. Client-money shortfalls — firms have been fined heavily where client-money reconciliation failed to spot that segregated money was short. Aged breaks becoming losses — an unexplained break left open too long turns out to be real money gone, no longer recoverable. And fraud via false matching — where the same person books and reconciles, misappropriation can be hidden behind fake matches.

From those patterns fall four lessons that generalise to any reconciliation. Independence is everything: whenever booking and reconciling were not genuinely separated, the control failed, which is why segregation of duties is non-negotiable. Aging cannot be ignored: old breaks are not administrative backlog, they are unexplained risk, and escalation and provisioning must actually bite. Completeness matters: a reconciliation that misses accounts or products leaves blind spots, so coverage must be inventoried and attested. And automate the boring but watch the odd: automation raises quality, but a healthy control keeps human judgement on the unusual and the material rather than letting a machine wave it through.

The economics of a break explain why the control pays for itself

Every break carries a cost even when no money is ultimately lost, and adding those costs up is what justifies investment in the control. There is the investigation effort — analyst time to research and clear it, small on each break but enormous in aggregate across millions. There is funding and interest — a cash break can mean an overdrawn or idle balance with a real funding cost, which matters more as intraday liquidity comes under supervisory scrutiny. There are provisions and write-offs — aged, unresolvable breaks that become actual losses. And there is capital — operational-risk losses feed the op-risk capital the bank is required to hold, so weak reconciliation is expensive even on the balance sheet.

A rough illustration makes the aggregate cost tangible. Suppose a reconciliation estate raises 30,000 breaks a month, of which 80% are false breaks from dirty data. If each false break takes an analyst ten minutes to clear, that is 24,000 breaks costing 4,000 analyst-hours a month — the equivalent of dozens of full-time staff spending their days clearing noise. A reference-data project that removed even half of those false breaks would free thousands of hours and sharpen the control, because the real breaks would stop being buried. The numbers are illustrative, but the shape is exactly why the business case for data hardening is so strong: the cost of false breaks is enormous precisely because each one is individually trivial and there are so many of them.

Against that sits the far larger cost of the control actually failing. The loss you did not catch — a fraud, a rogue position, a client-money shortfall — is the loss reconciliation exists to prevent, and the three cases above show its scale. Regulatory penalties, remediation orders and mandated skilled-person reviews follow client-asset and reporting failures. Fixing a failed control after the fact, plus the reputational damage, costs far more than running it well. And on the other side of the ledger sits the efficiency dividend: cutting false breaks and automating the routine work frees Ops to focus on genuine risk while lowering cost — which is the business case for every transformation in the next section.

The efficiency dividend is the business case that funds everything else

Everything in this part points to one commercial conclusion: running reconciliation better does not merely cut cost, it improves the control — and that double win is what funds transformation. Cutting false breaks frees analysts from clearing noise; automating the routine moves people to the exceptions that carry risk; hardening the data stops whole classes of break at source. Each of these lowers cost and sharpens the control, because in every case the genuine breaks stop being buried under the noise. That is unusual: in most of the bank, control and efficiency trade off against each other, and spending less means catching less. In reconciliation they move together, because the dominant cost is waste — false breaks and manual effort that protect nothing.

This is why the case for investment is strong even in a cost-conscious institution. A programme that removes false breaks pays for itself in freed analyst time, reduces operational-risk losses and the capital held against them, cuts the funding cost of aged cash breaks, and lowers the chance of a regulatory penalty — while making the control more likely to catch the loss it exists to prevent. The unglamorous data work that this module keeps returning to is not a cost centre to be minimised; it is the highest-return investment available in the operational control stack. An analyst who can articulate that dividend is the one who gets the transformation funded.

Generative AI and machine learning — real uses, honestly governed

Machine learning is making matching and break-handling smarter, and there are five uses worth taking seriously — assessed honestly rather than hyped. ML matching trains models to pair records that rigid rules miss — messy references, fuzzy names — raising the auto-match rate on exactly the hard items that clog the queue. Break classification auto-tags a break's likely cause from its features and history, speeding investigation. Break prediction flags items likely to break before they do, so they can be pre-empted. Suggested resolution proposes how to clear a break by analogy to how similar ones were cleared before. And LLM assistants summarise evidence, draft break commentary, and help onboard a new reconciliation by reading its documentation.

ML matching

Rigid rules match what they are told to match; a model can learn to pair records that the rules miss — a reference that is close but not identical, a name recorded two different ways, a netted group with no clean key. Trained on how past items were matched and resolved, it lifts the auto-match rate on precisely the hard, messy items that otherwise clog the exception queue. It does not replace the rule engine; it sits on top of it, catching what the deterministic passes leave behind.

Break classification

When a break is raised, a model can predict its likely cause from its features and the history of similar breaks — flagging it as a probable fee, a timing difference, or a data-reference issue before an analyst even opens it. That head start speeds investigation and routes the break to the right place, turning the first, slowest step of the break lifecycle into something closer to instant.

Break prediction

Rather than waiting for a break to occur, a model can flag items likely to break before the run — a payment whose reference looks malformed, a position whose corporate action is due — so the underlying problem can be pre-empted. Prediction is the most forward-looking use and the one that most changes the economics, because a break prevented costs nothing to clear.

Suggested resolution and LLM assistants

For a break that does occur, a model can propose a resolution by analogy to how similar ones were cleared, and a language model can draft the break commentary, summarise the supporting evidence, and help a business analyst onboard a new reconciliation by reading its documentation. These are genuine time-savers on the paperwork around the control — but they are assistants to a human decision, not the decision itself.

Use of AI / ML The gain The governing risk
ML matching Higher auto-match on messy, hard-to-pair records A wrong match hides two real breaks — needs justification
Break classification Faster investigation via auto-tagged causes Mis-classification misroutes and delays a real break
Break prediction Pre-empt breaks before they occur False positives create work; false negatives give false comfort
Suggested / auto-resolution Cuts manual effort on routine breaks A wrongly auto-closed break hides real risk
LLM assistants Faster commentary, evidence, onboarding MNPI and confidentiality; hallucinated facts in a control record

The honest caveats matter more here than in most parts of the bank, because reconciliation is a *key* control. Data comes first and AI second — none of these techniques beats simply cleaning the upstream data, which is usually the real bottleneck. Explainability is not optional: a key control must be auditable, so a black-box match or auto-resolution needs a justification a regulator can follow. Auto-resolution is genuinely risky, because a wrongly auto-closed break hides real risk, so humans must stay on the material items. And the whole thing is an assistant, not an autopilot — AI augments matching and investigation, but the control and its accountability stay human. European banks pursue all of this under model-governance, audit and the emerging EU AI Act, where explainability for a key control is essential, and the near-term value is higher auto-match and faster investigation on top of clean data, not autonomous resolution of material breaks.

  1. Why is the risk taxonomy described as a map of controllability?

  2. Why can a high auto-match rate be misleading?

  3. Name the three real failure cases and the failure mode each represents.

  4. What does the JPMorgan case show about client-asset reconciliation?

  5. What is the filter for judging a transformation?

  6. Why must AI in a key control remain explainable, with a human in the loop?

5

  1. Why can a high auto-match rate mislead?

  2. Which case shows a control evaded by concealment?

  3. Which risks can reconciliation fix itself?

The six ideas that matter most

If the whole module compresses to six ideas, these are they. Each is the through-line of a part, and together they are reconciliation in one page of argument.

Idea 1 — Independence is what makes it a control

A comparison is only a control if at least one side is independent. Checking a system against a copy of itself catches nothing, because a copy of a mistake is still a mistake. The authority of reconciliation comes entirely from segregation of duties and an external source of truth — separate people, separate data — whether the recon proves cash, stock, margin, the group's own consistency, or a regulatory report. Every rogue-trading collapse is, at root, this idea failing.

Idea 2 — Data cascades, so errors are never contained

An error in one field of one record does not stay there. Because every downstream system copies from the one before it, a wrong amount or a stale mapping propagates through risk, the ledger and the regulatory report, and the systems will agree with each other while all being wrong. Reconciliation against an independent source at each junction is what breaks the cascade — which is why data quality at the source is the highest-leverage work in the discipline.

Idea 3 — The break is the point, and aging is the signal

A break is not a failure; it is the output the control exists to produce. Health is measured not by having zero breaks but by how fast and how cleanly they are investigated, resolved and prevented from recurring. An aging pile of unexplained breaks is the real warning sign, because that is where hidden losses accumulate. Root-cause analysis turns the daily check into a continuous improvement engine.

Idea 4 — Compliance is core architecture, not an afterthought

Reconciliation is mandated by regulation and specified by it. CASS client-asset protection, EMIR, MiFIR and SFTR reporting reconciliation, UMR margin exchange, SOX-style key controls, CSDR settlement discipline and DORA operational resilience each read as a system requirement — coverage, evidence, an immutable audit trail, tested failover. A firm that treats these as paperwork bolted on at the end has built the wrong system; a firm that treats them as architecture builds a control that is compliant by construction.

Idea 5 — Most breaks are noise, and the noise is the opportunity

The large majority of breaks are false breaks — data differences, not real errors — and they are where the cost and the opportunity both sit. They drown the genuine breaks and consume Ops effort, and the durable fix is never a wider tolerance but cleaner upstream data. Fixing the data does not just cut cost; it sharpens the control, because the real breaks stop being buried. This is the quiet economics behind every reconciliation transformation.

Idea 6 — Real transformation moves cost or risk, including GenAI

The genuine improvements change the economics or the risk of the control, not its appearance. No-code self-service, consolidation, automation, data hardening and real-time reconciliation all pass that test; a prettier dashboard does not. Machine learning belongs in the same frame — valuable where it raises the auto-match rate on hard items or speeds investigation, governed with explainability and a human on anything material. And however fast the tooling lets you change a recon, the safe-change discipline — test, version, sign off — is what keeps agility from becoming an ungoverned control. Clean data first, AI second, accountability always human.

Process Cheat Sheet — The Whole Discipline on One Page

A single reference for the entire reconciliation flow — the lifecycle, the numbers to know, and the discipline in one sentence.

The lifecycle at a glance

Stage Core activity Key output Lead and systems
1. Load / ingest Pull both sides from files, feeds, APIs, messages Two raw data sets (A and B) Automation · ingestion connectors
2. Normalise Align fields, formats, signs, currencies, references Two comparable data sets Config + BA · normalisation layer
3. Match Apply keys and rules in passes; auto-match Matched pairs + unmatched pile Matching engine
4. Exceptions Classify unmatched items as breaks; queue them Prioritised break queue Engine + workflow
5. Investigate / resolve Find cause; correct, chase, claim or write off Cleared breaks + cause codes Ops analysts (human)
6. Report / evidence Record outcome, produce evidence, feed metrics Attestation, KPIs, audit trail Automation + sign-off

The break lifecycle at a glance

Step Activity What clears it
Open Engine raises, classifies and assigns the break —
Investigate Gather evidence: trade, statement, break history —
Categorise Tag the root cause with a standard code —
Act Correct, chase, claim or escalate —
Resolve / write off Close on agreement, or write off with sign-off Four-eyes approval

Key numbers to carry in your head

Metric Rule of thumb (illustrative)
Auto-match rate — clean cash recon 95%+
Auto-match rate — harder / messy recon 70–90%
Pass order Exact → tolerance → fuzzy → grouped
Tolerance on a EUR nostro (example) ±EUR 5 (flat cap)
Aging buckets 0–2d · 3–5d · 6–30d · 30d+
Typical aging SLA (example) Warn at 2d, escalate at 5d
Share of breaks that are false (data noise) The large majority
Client-asset shortfall rule Fund from the firm's own money until resolved
EMIR reporting Both sides report; TR pairs on the UTI
Margin recon (VM/IM) Daily; a mismatch is a margin dispute (funding + credit)
Regulatory reporting recon EMIR · MiFIR (ARM) · SFTR — chase every rejection
The three identifiers ISIN (security) · LEI (entity) · UTI (derivative trade)
Settlement direction of travel T+2 → T+1, pushing toward intraday recon

The whole discipline in one sentence

Reconciliation compares two records that should agree — using keys to line them up and a rule to decide agreement — and flags every difference as a break, so that an independent function can investigate it, resolve it, and fix its cause before it becomes a loss.