Leaving this course?

Lesson 1 of 14

One payment, many owners

A payment leaving your institution passes through more controls than almost any other object in financial services. Authentication proves somebody asked for it. Authorisation decides whether they could. A fraud score asks whether they meant it. Sanctions screening asks whether either party is designated. The message format decides what information travels. Settlement decides when it becomes irreversible. Monitoring looks at it again afterwards. Reconciliation checks that what you think happened is what happened.

Each of those is competently built in most institutions. Each is also owned by a different team, designed against a different standard, measured on a different number, and changed on a different release cycle. The failures are rarely inside a control. They are at the seams.

A seam is where one control's assumption is another control's output. Screening assumes the name field contains a name. Monitoring assumes the stated purpose of the relationship is on file. Recall assumes somebody is reading the inbox. Each assumption is reasonable, none is tested end to end, and the gap between two correct controls is invisible from inside either of them.

This course follows a single payment through the chain and stops at each seam. It assumes you already know what the individual controls do — if you want one of them on its own terms, the catalogue has a course for it. What it adds is the thing no single-control course can: the behaviour of the system when all of them are switched on at once.

A note on why this is worth a course of its own. Most compliance subjects are about a judgement — is this customer acceptable, is this activity suspicious — and judgement can be improved by better information and more time. Payments are about a decision taken under a deadline you do not control, repeated millions of times, where the information arrives in a fixed-format message somebody else composed. That changes what "good" looks like: not better judgement, but a chain whose behaviour you can predict and state.

One idea to carry through. Payments are different from almost every other compliance surface because of irrevocability. A wrong decision about a customer can be revisited next week. A wrong decision about a payment may be final in seconds. Every design choice in this course is downstream of that one fact, and the controls that matter most are the ones that act before the point of no return rather than the ones that report accurately afterwards.

Lesson 2 of 14

Authentication is not intention

The first control is also the most widely misunderstood. Strong authentication proves that the customer initiated the payment. It says nothing about whether the customer wanted to.

That distinction is the whole of the authorised push payment problem. A customer who has been deceived into sending money authenticates perfectly: correct credentials, correct device, correct second factor, biometric match, no anomaly of any kind. Every signal a security control is designed to read says this is the legitimate account holder, because it is. The fraud is in the reason, and authentication cannot see reasons.

This also explains why the industry's response to scams has moved towards warnings, delays and confirmation of the payee rather than stronger authentication. Those measures attack the reason rather than the identity: they interrupt the deception, give the customer a moment outside the conversation they are being steered through, or tell them the name on the account does not match the one they were given. None of them is an authentication improvement, and an institution that keeps strengthening authentication in response to scam losses is treating the wrong variable.

So separate the three questions and know which control answers which:

  • Is this the account holder? Authentication. Credentials, devices, biometrics, possession.
  • Is this person permitted to make this payment? Authorisation. Mandates, limits, roles, joint-signature rules.
  • Does this payment make sense for this person? Behavioural and fraud controls, which compare it to what this customer normally does and to patterns seen elsewhere.

Institutions that treat the first as covering the third are the ones surprised by scam losses, and the surprise is structural rather than a failure of effort — they built an excellent answer to a question nobody was asking.

What the third control actually needs, and rarely gets, is context that is not in the payment. The payment message says an amount, a beneficiary and a reference. It does not say that the customer has been on the phone for forty minutes, has never paid this beneficiary before, has just changed their contact details, or is making a property payment for the first time. Each of those is knowable and most of it lives in a different system from the one making the decision. The practical work in this area is less about better models than about getting the signals into the same place as the decision.

Lesson 3 of 14

What must travel with it

The FATF's Recommendation 16 is now titled Payment transparency — the October 2025 text and its interpretive note both use that name, in place of the older "wire transfers" framing. The change of title is a change of emphasis worth noticing: the subject is not a payment type, it is whether a payment can be understood by everyone who handles it.

The requirement has three limbs, and the first is the one most systems were built for. Countries should ensure that financial institutions "include required and accurate originator information, and required beneficiary information, on payments or value transfers and related messages. This information should be structured to the extent possible and should remain with such payment or value transfer or related message throughout the payment chain."

Read the three qualifiers separately, because each is a distinct engineering obligation.

  • Required and accurate. Not merely present. A field populated with a reference number, a branch name or the string "customer" satisfies a schema and defeats the purpose.
  • Structured to the extent possible. A single free-text block containing a name, a street and a city is parseable by a human and hostile to every automated control downstream. This is the clause that makes data structure a compliance matter rather than a technical preference.
  • Remain with the message throughout the chain. The information has to survive every hop, every format conversion and every intermediary. The commonest way to fail this limb is not to strip the data deliberately; it is to convert between message formats and silently lose the fields the new format has no room for.

That last failure mode is worth dwelling on, because it is invisible locally. Each institution in the chain can be fully compliant with what it received, and the payment can still arrive at the far end stripped of the information that would have identified it — because somewhere in the middle a well-behaved system mapped a rich message into a narrower one. Only an end-to-end test finds that. No participant's own monitoring will.

Lesson 4 of 14

Detecting what is missing

Recommendation 16's second limb is an obligation most institutions discover late: countries should ensure that financial institutions "monitor payments or value transfers for the purpose of detecting those which lack required originator and/or beneficiary information and take appropriate measures".

So you are required to detect absence, which is harder than it sounds and quite different from the work of monitoring for suspicious patterns. A control looking for unusual activity is looking at what is there. This one is looking for what is not — a beneficiary with no address where one was required, an originator name that is plainly a placeholder, a field that is present but meaningless.

"Appropriate measures" is deliberately unspecified, and the operational choice matters:

  • Reject the payment and send it back, which is clean, defensible and visible to the sender.
  • Enquire, holding the payment while you ask the previous institution for the missing information — slower, and the delay is itself a decision with consequences for the customer.
  • Process and record, where the risk is low and the gap is explicable, which is sometimes right and must be a documented policy rather than a default produced by a queue nobody works.

What is not defensible is a fourth option that occurs by accident: detect, log, and never act. An institution that generates a daily report of incomplete payments which nobody reads has created evidence that it knew, which is a materially worse position than not having looked.

There is also a reciprocal duty worth noticing, because it is the one institutions forget: if you are required to detect payments that lack information, then somebody downstream is making the same check on what you send. A rejected payment is not only an inbound problem. Measuring your own outbound completeness — before a correspondent measures it for you, and tells you about it in a review — is cheap, and almost nobody does it until after the first complaint.

One practical point about placement. Detecting missing information is cheapest at the boundary, where the message arrives, and most expensive once the payment is in your books. Pushing the check to the edge is both better control and lower cost, and it is one of the few places in this course where those two things point the same way.

Lesson 5 of 14

Cards, and the thresholds that change the rules

Not every payment carries the same obligations, and two distinctions in the interpretive note to Recommendation 16 decide a great deal of system behaviour.

Card payments have their own rule. For transfers flowing from a transaction carried out using a credit, debit or prepaid card for the purchase of goods or services, the interpretive note says the card number should accompany all transfers flowing from the transaction, and the name and location of the card issuing and merchant acquiring financial institutions should be made available upon request. Note the asymmetry: the card number travels, the institution details need only be available. That is a different design from a credit transfer, where the party information itself travels, and a system built on one model will under-populate the other.

The de minimis threshold changes what must be present. The interpretive note separates cross-border payments above the de minimis threshold, which must contain required and accurate originator information and required beneficiary information, from those below it, which must contain the name of the originator and the beneficiary and an account or reference number. Below the threshold the obligation is lighter — it is not absent.

Two consequences for anyone designing the controls:

  • Thresholds are an attack surface. Any rule that changes behaviour at an amount invites activity just underneath it. The structuring pattern is the oldest in this field, and the control is not a lower threshold but detecting the pattern of amounts clustered below one.
  • Batches are a separate case. The interpretive note addresses batch files, requiring that the batch contain required and accurate originator information and full beneficiary information, fully traceable within the beneficiary country. A batch is one object to your system and many payments to the standard, and controls written for single messages routinely treat a batch as a single item — which is how a thousand payments pass through a check designed for one.

And where technical limitations prevent required information accompanying a cross-border transfer from remaining with a related domestic transfer, the interpretive note requires a record to be kept for at least five years by the receiving intermediary of all the information received. That is the standard conceding that the chain sometimes breaks, and placing a record-keeping duty exactly where it breaks.

Lesson 6 of 14

Screening in the payment path

Recommendation 16's third limb is the sanctions one. Countries should ensure that, in the context of processing payments or value transfers, financial institutions take freezing action and prohibit conducting transactions with designated persons and entities, under the relevant United Nations Security Council resolutions — the text names resolution 1267 (1999) and its successors, and resolution 1373 (2001).

That is an obligation to act, in the payment path, before the money goes. It is not an obligation to detect eventually.

The Wolfsberg Group's guidance draws the distinction that organises this work: institutions deploy transaction screening and customer screening, and they do different jobs. "Transaction screening is used to identify transactions involving targeted individuals or entities. Customer or Name screening is designed to identify targeted individuals or entities during on-boarding or the lifecycle of the customer relationship." Neither substitutes for the other. A designated party who is not your customer appears only in transaction screening; a customer who becomes designated after onboarding appears only if you re-screen.

The guidance is also unusually candid about configuration, and it is worth quoting because it cuts against how screening is often discussed: "There is no clearly defined approach to technology or configuration that is better or worse, and each will have its own strengths and limitations. Understanding those strengths and limitations is critical." Institutions are expected to document how their systems are configured, so as to demonstrate the configuration is reasonably expected to detect the risks they face and "to ensure transparency of any system limitations or risk based decisions which the screening controls are not designed to detect".

That final clause is the professional standard in this area. You are not expected to catch everything. You are expected to know and state what you have chosen not to catch — which data attributes you screen, when you screen, which lists you use, and how exact or fuzzy the filter is set. An institution that cannot answer those four questions does not have a screening programme; it has a screening product.

Lesson 7 of 14

The latency budget

Every control in the payment path spends time, and in instant payments the total available is a second or two. That turns control design into a budgeting problem, and the budget is the most under-discussed constraint in payments compliance.

Work out what the chain actually costs. Authentication, authorisation, limit checks, a fraud score, a sanctions screen against several lists with fuzzy matching, a format validation, a liquidity check. Any one of them is fast. All of them in series, with a network hop between some, is frequently not fast enough — and the pressure to meet the scheme's timing is immediate and commercial, while the pressure to screen thoroughly is occasional and regulatory. Guess which one wins by default.

Be careful about where the budget is measured, too. A control's own latency is usually monitored by the team that owns it, in isolation, and looks fine. What matters is the end-to-end time for a payment that exercises every control — which is a different number, owned by nobody, and typically only discovered when a scheme reports that you have breached its timing.

Three legitimate responses, and one that is not:

  • Move work earlier. Anything that does not depend on this specific payment can be done before it: list changes screened against the customer base overnight, beneficiary checks done when the beneficiary is first added rather than at each payment, risk scores precomputed.
  • Move work later, deliberately. Some checks genuinely belong after the event, and saying so explicitly is better than pretending a one-second window accommodates them.
  • Fail closed on the things that must not pass. A sanctions screen that times out must not default to release. That is a configuration choice, it is frequently set the wrong way for availability reasons, and it is the single highest-consequence default in the chain.

The illegitimate response is to quietly loosen the matching to make the budget. Widening a filter's tolerance to save thirty milliseconds is a risk decision disguised as a performance fix, and it will not appear in any risk record. If the budget cannot accommodate the control, that is a finding to escalate, not an engineering problem to absorb.

One more thing worth measuring, because almost nobody does: the distribution of your own control latency, not its average. The payments that time out are not random; they are the ones with the most name matches to resolve, which is to say the ones most likely to matter.

Lesson 8 of 14

Irrevocability, and the point of no return

The fact that organises everything else: at some moment a payment stops being something you can undo and becomes something you can only ask about. Knowing exactly when that moment is, for each rail you use, is the most load-bearing operational knowledge a payments control function has — and it is frequently held informally, by two people, in their heads.

Before that moment, you have control. You can hold, release, amend or cancel unilaterally. After it, you have requests. A recall, a return, an indemnity claim — all of them depend on another institution choosing to act, and most of them depend on the funds still being there.

So map it explicitly, per rail: when is the payment final, who can stop it before that, how long the window actually is in clock time rather than in scheme documentation, and what the mechanism is afterwards. Then design the escalation path to fit inside the window rather than discovering the window during an incident.

Three things that regularly surprise people:

  • The window is often shorter than the batch cycle. A control that reviews a file after it is submitted is reviewing payments that have already gone.
  • A return is not a reversal. The money comes back as a new payment, which has its own screening, its own value date and its own possibility of failure. Treating a return as an undo misstates the risk to the customer.
  • Finality is a legal property, not a technical one. Whether a payment can be unwound may depend on insolvency law, scheme rules and the jurisdiction of each participant, not on whether a system has a cancel button.

The practical test of whether an institution understands its own irrevocability: ask what happens if a payment is identified as fraudulent four minutes after release, at 02:00, on a rail that settles in seconds. If the answer involves an email and a working day, the control framework is built for a world that no longer exists.

Lesson 9 of 14

The payment message as an attack surface

Everything so far assumes the payment instruction is genuine and the question is whether to let it through. The other threat is an instruction that was never given by anybody authorised, created inside the institution's own systems by somebody who has taken control of them.

This is a different problem from fraud against a customer, and it inverts the usual logic. The controls that protect a customer's payment assume the institution is trustworthy. When the institution's own payment creation is compromised, those controls are the attacker's tools: the message is well formed, correctly authenticated at the system level, within limits, and screening will clear it because the beneficiary is not designated.

What actually defends against it is structural rather than analytical:

  • Separation between creating, approving and releasing a payment, enforced by the systems rather than by procedure, so that one compromised credential cannot complete the chain.
  • An approval step that uses a different path from the one used to create the instruction. Two approvals from the same compromised terminal are one approval.
  • Limits on the payment system itself, not only on customers — a maximum value and volume for the institution's own outbound traffic, so that an anomalous burst hits a ceiling even if every individual message passes.
  • Alerting on the absence of the expected, such as a reconciliation report that does not arrive, or logging that stops. Attackers interfere with the evidence, and a missing report is a signal rather than an inconvenience.
  • Independent confirmation of what was sent, from a source the payment system does not control.

The last one is the most valuable and the most commonly weakened for convenience. If your record of what you sent comes only from the system that sent it, you have no way to detect that the system is lying to you. A statement from the correspondent, a feed from the scheme, or a report from an independent copy of the traffic is the thing that makes the compromise visible — which is the subject of the next lesson, because that independent comparison is a security control that almost nobody in security owns.

Lesson 10 of 14

Reconciliation as a security control

Reconciliation is usually filed under finance, treated as a hygiene task, and given to the most junior available person. It is also the control most likely to detect a sophisticated payments compromise, and the mismatch between those two sentences is one of the most consequential organisational errors in this field.

The reason is simple. Every other control asks whether a payment should happen. Reconciliation asks whether what happened matches what you believe happened — and it is the only control that compares two independently produced records. A compromise that defeats authentication, authorisation, limits and screening still has to make two sets of books agree, and making them agree requires compromising both.

What follows from treating it as a control rather than a chore:

  • A break is an incident until proven otherwise. The default assumption for an unexplained difference should be that something is wrong, not that somebody will find the keying error. Most breaks are keying errors. The control has value precisely because it treats them all as potentially not.
  • Frequency is a security parameter. Daily reconciliation on a rail that settles in seconds gives an attacker a day. The right frequency is a function of how fast money can leave, not of when the accounting team works.
  • Independence is the whole value. If both sides of the comparison derive from the same system, the exercise proves internal consistency and nothing about truth.
  • Ageing matters more than volume. A hundred small breaks cleared within the hour is a healthy process; three breaks open for a fortnight is a hole, and the second is usually reported as better because the number is lower.

One organisational recommendation that costs nothing: give the reconciliation break report a route into the security incident process, not only into the finance queue. The signal exists in most institutions already. What is usually missing is anybody in a position to recognise what it means.

Lesson 11 of 14

After the event, and the suspicion

Some things can only be known afterwards. A single payment is unremarkable; the twelfth to the same beneficiary in a week is not. Post-event monitoring exists because pattern needs time, and the payment path cannot wait.

This is the seam where payment controls meet the anti-money-laundering (AML) regime, and the hand-off has two failure modes that look nothing alike.

The first is the payment stopped and nothing reported. A control held a payment, a human reviewed it, the customer gave an explanation, the payment was released — and the suspicion that triggered the hold evaporated with the release. The FATF's Recommendation 20 is short and does not care whether the payment went through: where an institution suspects, or has reasonable grounds to suspect, that funds are the proceeds of criminal activity, it should be required by law to report promptly to the financial intelligence unit. Releasing a payment and reporting a suspicion are not alternatives. They are independent decisions, and conflating them is the commonest reporting failure in payments.

The second is the report made and the payment released without anyone noticing the tension — or, worse, the customer being told why. Recommendation 21 prohibits disclosing that a report is being filed, and a payment operations team asking a customer direct questions about a transaction under suspicion is exactly where that prohibition gets breached by people acting in good faith.

Note also that Recommendation 20 sets no number of days. Promptness is required, the deadline is a national choice, and it is often very short. Payments teams that work to a monthly reporting cycle are frequently working to a cycle nobody chose.

The practical arrangement that prevents both failures is unglamorous: one named owner for the reporting decision on each payment case type, recorded on the case rather than agreed informally, and a rule that a hold released is still a case with a reporting question attached to it.

Lesson 12 of 14

Data quality is the control

Everything in this course rests on fields. Screening matches on names, monitoring aggregates on counterparties, transparency obligations attach to specific attributes. When the data is poor, every control above it degrades at once — and it degrades silently, because each control reports that it ran.

Recommendation 16's "structured to the extent possible" is the compliance hook, and it deserves to be read as an instruction to engineering. A structured field can be compared; a free-text blob can only be searched. The difference shows up as false positives, missed matches and aggregation that silently splits one counterparty into four.

Where the damage concentrates:

  • Names. Transliteration, name order, titles, matronymics, abbreviations, company suffixes. Every one of these is a reason two records of the same party do not match, and the fix is partly in the matching and mostly in the capture.
  • Addresses and countries. A country in a free-text address is not a country code, and risk rules keyed on geography will miss it.
  • Identifiers. An account number, a legal entity identifier or a tax number is worth more than any amount of name matching, and it is the field most often omitted because it is "not required".
  • Purpose and reference fields. Frequently the only statement of what a payment is for, and frequently filled with an internal code meaningful to one party.

The Wolfsberg guidance makes the governance point for screening specifically, and it generalises: the risk decisions include "what data attributes to screen, when to screen, what lists to use and how exact or 'fuzzy' to set the screening filter", and those decisions need to be "clearly articulated, documented and supported by analysis and testing". The first of those four is a data decision, not a configuration one.

A format migration is the moment to fix this and the moment most likely to make it worse. Richer structured formats create room for better data; they do not populate it. A migration that maps old free-text into a new structured field, unparsed, has moved the problem and spent the budget.

Lesson 13 of 14

Testing the chain, not the controls

Each control in this course is tested. Authentication has test cases, the fraud model has a validation set, the screening filter has a tuning exercise, the message validator has a schema suite. The chain, as a chain, is usually not tested at all — and the seams are where the failures are.

What end-to-end testing means here is specific: inject a payment with a known property at one end and observe what every control did with it. Not a review of configurations, not a walkthrough. An actual payment, through the actual path.

Properties worth injecting:

  • A name that should match a list at the fuzziness you believe you have configured, to see whether it does.
  • A message missing a required field, to see whether the absence is detected, by which control, and what then happens to the payment.
  • A format conversion in the middle of the chain, to see which fields survive it.
  • A payment just below each threshold, to confirm the behaviour change is where you think it is.
  • A control that times out, to see whether the chain fails open or closed. This is the test institutions most avoid running and most need to.
  • A batch, to confirm the controls look inside it.

Two disciplines make this work. Know the expected result before you run it, because a test whose expected outcome is decided afterwards always passes. And test after every change to any control in the chain, since the seam that breaks is rarely in the component that changed.

The output worth keeping is not a pass list. It is a short document saying: this is what our chain does to a payment with each of these properties, as measured on this date. That document is the only honest answer to a supervisor asking how you know your controls work together, and almost no institution has one.

Lesson 14 of 14

Owning the seams

The last lesson is organisational, because every technical point in this course has an owner problem behind it.

Ask who owns the payment end to end and you will usually get a list rather than a name: security owns the rails, fraud owns the scoring, compliance owns screening and reporting, operations owns the queues, product owns the customer experience, technology owns the message formats. Everything is owned. The spaces between are not, and that is where this course has spent fourteen lessons.

What helps, in rough order of effect:

  • A named owner for the chain, with the authority to require a change in a control they do not own. Without the authority the role is a coordinator, and coordinators lose to release schedules.
  • A single map of the chain that shows every control, what it assumes about its input, and what it guarantees about its output. Most of the seam failures in this course are visible on such a map and invisible without one.
  • A change process that asks the seam question: what else assumes something about what this control produces? One question, added to an existing form, catches a surprising proportion.
  • Shared measurement. If fraud is measured on losses, compliance on filing timeliness and operations on throughput, the seams will be optimised away by people doing their jobs correctly. At least one measure should belong to the chain — payments released that should have been stopped, for instance, regardless of which control should have stopped them.
  • A standing forum with the authority to decide, because seam questions are cross-team by definition and the alternative to a decision is the status quo.

And when it does go wrong, the incident review should ask the question this course is built on. Not "which control failed", which usually has the answer "none of them, individually". But: which control assumed something another control did not guarantee? That question finds causes. The other one finds owners to blame, which is a different and much less useful activity.

Knowledge check

Fifty questions

Answer all fifty questions, then check your answers. You need 45 out of 50 to pass and receive a certificate. If you score less, you will see which answers were right and wrong, and then go through the course again before you retake the check. Your answers, progress and times are kept only in this browser.

Standards and frameworks covered

The published frameworks this course teaches toward. Saolix is not affiliated with, endorsed by or accredited by any of these bodies; naming a standard is a citation, not a claim about its author.

Sources

The official documents this course relies on. Laws and guidance change, so check the current version.

  1. The FATF Recommendations (Recommendation 16 and its interpretive note, Recommendations 11, 20 and 21, updated October 2025) · Financial Action Task Force
  2. Revised Recommendation 16 on payment transparency (June 2025) · Financial Action Task Force
  3. Wolfsberg Guidance on Sanctions Screening (2019) · The Wolfsberg Group
  4. Payment Transparency Standards (2023) · The Wolfsberg Group