Leaving this course?

Lesson 1 of 14

What transaction monitoring is for

customer due diligence (CDD, which much of the industry calls KYC) tells you who a customer is and what they said they would do. Transaction monitoring checks what they actually do, and flags activity that does not fit.

It is a different control from screening. Screening asks whether a name matches a sanctions or politically exposed persons list. Monitoring asks whether a pattern of activity looks like money laundering, terrorist financing or fraud. A customer can pass every screening check and still use their account to launder money.

Monitoring happens in two ways:

  • Before or as a payment happens, so that a transaction can be held for review before the money leaves.
  • After the event, looking back over days or weeks of activity to find patterns no single transaction reveals.

Ongoing monitoring is the fourth customer due diligence measure in Recommendation 10, and the Interpretive Note to it requires institutions to examine the background and purpose of all complex, unusual large transactions and unusual patterns that have no apparent economic or lawful purpose. National law turns that into a duty you can be examined against. In Kenya, for example, section 44(1) of the Proceeds of Crime and Anti-Money Laundering Act, 2009 requires reporting institutions to monitor, on an ongoing basis, all complex, unusual, suspicious and large transactions, whether completed or not. They must also watch for unusual patterns, and for small but repeated patterns of transactions, that have no apparent economic or lawful purpose. The law allows an institution that forms a suspicion to complete the transaction while it reports (section 44(10)); separately, the Financial Reporting Centre can direct an institution not to proceed with a transaction for up to five working days (section 44A).

Most monitoring compares activity with a baseline: the customer's own history, what they declared at onboarding, and what is normal for similar customers. That is why good due diligence and good monitoring depend on each other. An alert is not proof of anything; it is a question that someone must answer.

The purpose is not to generate alerts. It is to find activity that should be reported to the authorities, and to leave a clear record of how each decision was made.

Terminology: the FATF standards say suspicious transaction report (STR), while the United States and the United Kingdom say suspicious activity report (SAR). They are the same filing, and this course uses whichever term its own sources use.

Lesson 2 of 14

Red flags and typologies

A red flag is a sign that something may be wrong. A typology is a known method criminals use. Recognising both is the core skill of monitoring. Common examples:

  • Structuring: splitting cash deposits or transfers into amounts just below a reporting threshold, such as the FATF designated threshold of USD/EUR 15,000, which many countries adopt as their cash reporting figure and Kenya sets at US$15,000, often across days, branches, agents or accounts.
  • Pass-through activity: money arrives and leaves again quickly, with little left in the account and no business reason.
  • Mule accounts: accounts, often in other people's names, that receive money from many unrelated senders and pass it on. Students and job seekers are sometimes recruited for this, knowingly or not.
  • Activity that does not fit the profile: a salaried customer receiving large business payments, or a small shop moving far more than its size suggests.
  • Sudden change: a dormant account that becomes very active, or a new pattern of cross-border transfers.
  • Unusual links: many customers sharing one phone number, address or device, or transfers to parties with no apparent connection.
  • Trade-based laundering: invoices priced far above or below market value, or goods described vaguely, used to move value across borders.

Where mobile money is widely used, the same patterns show up there as well as in banks, and that is most of the world now: the GSMA counted around 108 billion mobile money transactions worth over $1.68 trillion in 2024, across 2.1 billion registered accounts. Kenya is one heavily mobile market among many, and the signals look alike in all of them: many small transfers from unrelated senders, rapid movement between wallets and bank accounts, and agent transactions that do not match the customer's profile.

Red flags are prompts, not conclusions. Many have innocent explanations. The job is to find out which.

Lesson 3 of 14

How monitoring systems work

Most monitoring systems combine several approaches:

  • Rules and scenarios: defined patterns with thresholds, such as "cash deposits totalling more than a set amount within seven days" or "more than a set number of incoming transfers from different senders in a day".
  • Segmentation: grouping similar customers, such as individuals, small businesses and large companies, so each group gets thresholds that suit it. A pattern that is normal for a supermarket is unusual for a student.
  • Behavioural and statistical models, including machine learning, that score how far activity departs from the customer's normal pattern or look for hidden links between accounts.

Every approach involves a trade-off. Set thresholds too low and analysts spend their time clearing harmless alerts, known as false positives; set them too high and real cases slip through. Most alerts, particularly those generated automatically, turn out not to be suspicious, so tuning matters.

Good practice for managing the system:

  • Map each rule to a risk from your risk assessment, so you can show why it exists.
  • Tune with evidence: test what would happen above and below the current threshold before changing it, and record why you changed it.
  • Check data quality: a rule cannot see transactions that never reach it, or fields that are empty or wrong.
  • Review coverage whenever a new product, channel or typology appears.

A supervisor will ask not only "what does your system do?" but "how do you know it works?"

Lesson 4 of 14

Investigating an alert

An alert is the start of an investigation, not the end. A sound review follows the same steps each time:

  • Understand the alert: which rule or model fired, and on what activity.
  • Look at the customer: their due diligence file, risk rating, occupation or business, declared purpose and past alerts.
  • Look at the wider activity: transactions before and after, the counterparties and whether they are linked to other alerts or accounts.
  • Use available information: internal records, public sources and adverse media (negative news reports). If you need to ask the customer for an explanation or documents, do it as part of normal business, and never suggest that a report is being considered.
  • Decide and explain: either the activity has a credible, documented explanation, or it remains suspicious. Record what you found, what you concluded and why, in words a supervisor or court could follow.

Two principles keep decisions sound:

  • Suspicion is enough. You do not need to prove a crime or identify it. If the activity is suspicious and you cannot explain it away, it should be reported.
  • Four eyes for the hard cases. A second reviewer, or a quality-assurance check on a sample of closed alerts, catches both missed cases and inconsistent decisions.

Kenyan law requires a reporting institution, as far as possible, to examine the background and purpose of unusual or suspicious transactions, set out its findings in writing and keep them for at least seven years (section 44(4) and (5) of the Act). Closing an alert as "no concern" with no explanation does not meet that standard, and supervisors look for it. The reasoning is part of the work.

Lesson 5 of 14

Reporting a suspicion

When suspicion remains, it is reported to the country's financial intelligence unit. The FATF standard (Recommendation 20) is that institutions must report promptly when they suspect, or have reasonable grounds to suspect, that funds are the proceeds of a criminal activity or are related to terrorist financing.

Inside the institution, staff do not report to the authorities themselves. They report the suspicion to the Money Laundering Reporting Officer (MLRO), and the MLRO decides whether to file a report. Section 47 of the Act requires these internal reporting procedures. Regulation 12 of the Proceeds of Crime and Anti-Money Laundering Regulations, 2023 requires every reporting institution to appoint an MLRO at management level, with the authority and independence to do the job.

Recommendation 20 requires a report to reach the financial intelligence unit promptly, and sets no number of days: the deadline is a national choice and is often very short. Find yours, and find out whether it is counted in working days or calendar days, because that detail decides whether a report is late. In Kenya, for example, reports go to the Financial Reporting Centre (FRC), and section 44(2) of the Act requires a report within two days after the suspicion arose (see also regulation 38). Attempted transactions must be reported too (section 44(3)). The duty covers money laundering, terrorist financing, proliferation financing and the proceeds of crime.

Reporting institutions must also report every cash transaction of US$15,000 or more, or the equivalent in another currency, whether or not it looks suspicious, normally by the end of the week in which it happened (section 44(6) and the Fourth Schedule of the Act; regulation 40). Reports are filed electronically through goAML, the FRC's online reporting system. Check the FRC's current guidance for the exact forms and channels.

A useful report answers five questions clearly and factually:

  • Who: the customer and other parties, with identifiers.
  • What: the transactions, amounts, dates and accounts.
  • When and where: the period and the channels used.
  • Why it is suspicious: the red flags, what does not fit the profile, and what explanation was sought. Attach the documents that support the suspicion (section 44(7)).
  • What you did: any action taken on the account.

After reporting:

  • Do not tip off. Disclosing to the customer, or to anyone not authorised to know, that a report is being prepared, is about to be sent or has been sent to the FRC is an offence under section 8 of the Act.
  • Decide on the relationship under your procedures: continue with closer monitoring, restrict or exit. Record the decision.
  • Keep the records of the alert, the investigation and the report for at least seven years (sections 44(5) and 46(4) of the Act; regulation 42), or longer if the FRC requires.

Reporting in good faith is protected. Under section 19 of the Act, no suit, prosecution or other legal proceedings lie against a reporting institution or its staff for anything done with due diligence and in good faith under the Act, and section 20 keeps the identity of the person who gave information confidential. This reflects FATF Recommendation 21.

Lesson 6 of 14

Designing a scenario

A typology is a story about how criminals behave. A scenario is a query that runs against data. Getting from one to the other is the work, and it is where most monitoring programmes are weakest, because the typology is usually copied and the query is usually inherited.

"Cash deposits over a threshold" is not a scenario. It is a filter. A scenario specifies five things, and leaving any of them unstated means nobody can tune it, test it or defend it:

  • The population. Which customers it applies to. Retail only, or business? Which segments? A scenario that runs on everybody is tuned for nobody.
  • The trigger. The condition in data terms: which transaction types, which direction, which channels, which fields.
  • The window. Over what period the condition is measured, and whether the window is rolling or calendar. A seven-day rolling window and a calendar week catch different things, and the difference is exploitable.
  • The threshold. The value or count at which it fires, and in what currency, converted how.
  • The exclusions. What is deliberately not caught, and why. Salary credits, internal transfers, known trade settlement. Every exclusion is a hole somebody chose, so each needs a stated reason and a review date.

Write down what the scenario is for. One sentence naming the typology it is meant to detect. Scenarios without this accumulate, and nobody can later say whether a scenario is redundant, because nobody knows what it was for.

Then ask what a guilty case would look like in your data. This is the step that separates a working scenario from a plausible one. If the typology is third-party funding, what fields would show it? If the institution does not capture the payer's name on a deposit, the scenario cannot detect it, and the honest conclusion is a data gap rather than a detection control.

Beware the scenario that cannot fire. Thresholds above anything the customer base does, populations that are empty after exclusions, conditions requiring fields that are never populated. Every programme has some, and they are worse than no scenario at all because they appear in the coverage count.

And beware the scenario that always fires. A condition met by ordinary behaviour produces alerts nobody can clear, which trains the team to close them, which means the one real case in it will be closed too.

Version the definition. A scenario is a piece of logic that will change. Each version should record what changed, when, who approved it, and what the expected effect was — so that a change in alert volume six months later can be explained rather than investigated.

Lesson 7 of 14

Thresholds, and how to set one honestly

A threshold decides what the institution sees. Most are inherited, round numbers that nobody has tested, and changing one is treated as a risk rather than as maintenance.

A threshold is a decision about what to miss. That is the honest framing. Set it low and the institution drowns; set it high and activity passes unexamined. Either way somebody has decided, and the decision should be recorded with its reasoning rather than left implicit in a configuration file.

Set it from the data, not from the round number. Run the condition with no threshold across a period and look at the distribution of what it would catch. Where does the volume sit, what does the population at each level look like, and at what point does the alert population stop resembling ordinary behaviour? That curve is the evidence for a threshold, and it takes an afternoon.

Below-the-line testing is the control nobody runs. Sample the activity that fell just below the threshold and review it as if it had alerted. If a meaningful proportion would have been escalated, the threshold is too high and now there is evidence. Running this periodically is the only way to know what the threshold costs, and almost nobody does it, because it creates work and findings.

Above-the-line testing tells you the other half. Sample what alerted and was closed. If almost everything closes with no action, the scenario is catching ordinary behaviour, and the question is whether the threshold is too low or the scenario is wrong.

Change one thing at a time, and predict the effect first. Before a change, state the expected volume and why. After it, compare. A change that produces three times the predicted volume means the model of customer behaviour was wrong, which is worth knowing independently of the change.

Test changes against history before deployment. Replay the new definition over past data: what would have alerted, what would have stopped alerting, and would any case the institution actually reported have been missed? That last question is the important one, and a change failing it should not deploy.

Expect a ratchet, and resist it. Thresholds drift upward over time because high volumes are painful and nobody is rewarded for a lower one. The countervailing discipline is below-the-line testing and a record of why each level was chosen, which makes an upward change a decision somebody has to defend.

And do not tune for the alert count. The target is not a number of alerts; it is detection the institution can act on. An institution that sets thresholds to fit its team's capacity has decided its risk appetite by headcount, which is worth saying out loud, because sometimes the right answer is more people rather than a higher threshold.

Lesson 8 of 14

Segmentation, and why one threshold fails

A single threshold applied to every customer is the commonest design fault in monitoring. It is simultaneously too low for some and far too high for others, and it produces both noise and blindness from the same setting.

The arithmetic is obvious once stated. A threshold that is meaningful for a salaried individual is invisible to a wholesale business, and one meaningful for the business buries the individual's account in alerts. No single number serves both, so the population has to be divided.

Segment on what the customer does, not only on what they are. Declared segmentation — retail, small business, corporate — is a start and is often wrong, because the categories reflect how products are sold rather than how accounts behave. Behavioural segmentation groups accounts by observed activity: turnover, transaction count, counterparty variety, cash proportion, international proportion. The groups it produces frequently cut across the declared ones, and that mismatch is itself informative.

Keep the segments few enough to manage. Each segment multiplies the tuning work, and a programme with forty segments has forty thresholds nobody will ever test. A small number of well-chosen groups, each with a stated definition and a tuned threshold, beats a fine-grained scheme that is maintained once.

Peer comparison is the useful form. Rather than asking whether an amount is large, ask whether it is large for this account's peers. A business whose turnover is five times its segment's median is interesting at any absolute value, and that question is answerable from data the institution already holds.

Watch for the segment of one. Large customers often sit alone in their segment, which makes peer comparison meaningless and the threshold arbitrary. These accounts need to be monitored against their own history rather than against peers, and they are usually the ones carrying the most value.

Re-segment, and expect accounts to move. A business grows, a retail customer starts trading, a dormant account reactivates. An account in the wrong segment is monitored against the wrong expectation, which is a silent failure. Re-segmentation on a cycle, with a record of which accounts moved and why, is maintenance rather than a project.

Record the segmentation as part of the scenario. A scenario's definition is incomplete without saying which segment's threshold applies, and a change to segmentation is a change to every scenario that uses it — which is a good reason to treat the two as one piece of governance rather than two.

And check the empty and the enormous. A segment with almost no accounts in it, or one holding most of the book, both suggest the definitions are wrong. Counting the population of each segment is a five-minute check that few programmes have ever run.

Lesson 9 of 14

The data underneath

Every scenario is a query, and a query is only as good as what it runs against. Monitoring failures are far more often data failures than logic failures, and they are harder to see because the system keeps producing alerts.

Know which feeds arrive, and whether they all do. A monitoring system is fed from payment systems, card systems, the core ledger, cash handling, and often a different source for each product. A feed that stops is the worst failure in this field: the scenarios run, alerts appear from the other feeds, dashboards look normal, and an entire product is unmonitored. Reconciling transaction counts between source and monitoring system, daily, is the control that catches it, and many institutions do not have it.

Completeness by field, not just by record. The record arrived; did the fields? Counterparty name, counterparty account, originating country, narrative, channel. A scenario looking for third-party funding needs the payer; a scenario looking at corridors needs a country. Where a field is populated for a third of records, every scenario using it is running on a third of the book, and nothing says so.

Currency, conversion and the date used. Thresholds in one currency against transactions in many require a rate, and which rate, as at which date, matters at the margin. Timezones matter too: a transaction booked late on one day in one system may land on the next day in another, which breaks seven-day windows in ways that are maddening to debug.

Late-arriving and amended transactions. Reversals, corrections, back-dated entries. If the monitoring system takes a daily snapshot and never revisits, a transaction amended after the window closed is never seen in its final form. Whether the system reprocesses, and over what period, is a question worth being able to answer.

Static data decays. Occupation, expected activity, address, segment: monitoring compares behaviour against a profile, and the profile is often years old. A scenario comparing activity to expected turnover recorded at onboarding five years ago is comparing against fiction.

Internal transfers and known flows. These are usually excluded, and the exclusion is usually implemented by a rule somebody wrote once. If the account used for internal settlement changes, the exclusion silently starts excluding something else.

Test with known cases. Take transactions from a case the institution actually reported, put them through the current configuration, and see whether they alert. This is the single most informative test available, and it tests data and logic together, which is how they fail.

And own the lineage. For each scenario, which fields it uses, which feed each comes from, and who owns that feed. Without it, a data problem becomes an argument between teams instead of a fix, and the scenario keeps running on nothing.

Lesson 10 of 14

Coverage: proving the system watches for the right things

An institution's risk assessment names the risks it faces. Its monitoring system detects some set of behaviours. Nobody checks that the second covers the first, and that gap is the most common substantive finding in a monitoring review.

Build the matrix. Rows are the typologies and risks the institution has identified, in its own risk assessment and its own reports. Columns are the scenarios in production. The cells say which scenario would detect which typology, and for which population. The exercise takes days and it is the most valuable document a monitoring programme can produce.

The empty rows are the finding. A typology the institution has identified, with no scenario pointing at it, is an uncovered risk. Sometimes that is a reasonable decision — the risk is immaterial, or another control covers it — and sometimes nobody had noticed. Either way it should be a recorded decision rather than a discovery.

The crowded rows are also a finding. Six scenarios all detecting variations of structuring, and none detecting trade-based laundering, is a programme that grew by accretion. Alerts concentrate where scenarios cluster, which makes the team's experience of risk a reflection of the system's history rather than of the institution's exposure.

Check coverage against what you have actually reported. The reports the institution filed last year are a list of things it found. Which scenario found each one? If most came from one scenario, from referrals by branch staff, or from authorities' requests rather than from monitoring at all, that is the real picture of what detects risk here — and it is often uncomfortable.

And check against what you have missed. Where an authority or a correspondent has told the institution about a customer it had not flagged, that is a coverage test with a known answer. Each such case should be traced back: which scenario should have caught it, and why did it not?

Cover the channels, not just the typologies. A scenario running on payments but not on cards, or on domestic but not on international, has a gap defined by product rather than by risk. The matrix should have a population column precisely so these are visible.

Keep it current, and tie it to change. A new product, a new market or a new typology in the risk assessment should each prompt the question: which scenario covers this? Answering it at the point of change is a conversation; answering it two years later is a remediation programme.

And publish the gaps upward. The board approving a risk assessment should be told which of the risks it names are not monitored for, in those words. An institution that cannot produce that list does not know what its monitoring does, whatever its alert volumes suggest.

Lesson 11 of 14

Models, and what they need to be useful

Rules are explicit and brittle. Models learn patterns and are opaque. Most institutions now run some of both, and the governance question is not which is better but what each requires to be trustworthy.

What models are actually used for here. Three different jobs, often confused. Scoring alerts so the queue is worked in a sensible order. Detecting anomalies that no rule describes. And segmenting customers by behaviour. The first is the most common and the least risky; the second is the most ambitious and the hardest to validate.

Supervised learning needs labels, and labels are scarce. To learn what suspicious looks like, a model needs examples. The institution's own labels are its alert outcomes, which are mostly "closed, no action" — and those are decisions made by analysts under time pressure, not ground truth. Reports filed are a better label and are rare. A model trained on alert outcomes learns to predict what analysts closed, which is not the same as learning what is suspicious.

Unsupervised approaches avoid that and buy a different problem. Anomaly detection finds the unusual, and most unusual behaviour is innocent. Without a way to explain why something was flagged, the output is a list nobody can work, and the honest measure is whether analysts can act on it rather than whether the mathematics is sound.

Explainability is an operational requirement, not a philosophical one. An analyst has to write why this activity is suspicious, and a report has to say it in terms an authority can act on. "The model scored it 0.94" is not a reason. Any model in this pipeline needs to surface the features that drove the output, in terms somebody can put in a narrative.

Drift is the failure that creeps. Customer behaviour changes, products change, a new payment rail appears, and a model trained on last year's book slowly stops describing this year's. The symptoms are subtle: a gradual change in alert volume or in the mix of what alerts. Monitoring the model's own inputs and outputs for drift is part of running it.

Treat it as a model, with model governance. Documented purpose, data lineage, training population, performance measures, limitations, an owner, independent validation before deployment, and periodic revalidation. This is ordinary model risk management, and a monitoring model that has never been through it is an unvalidated control in a regulated process.

And keep the rules. Models are good at ordering and at finding the unexpected; rules are good at the known typologies the institution must demonstrably cover. A programme that replaces its coverage matrix with a score has lost the ability to say which risks it watches for, which is the question in the previous lesson and the one a supervisor asks first.

Lesson 12 of 14

Validating the system, and testing every change

A monitoring system is a control, and like any control it has to be tested by somebody other than the people who built it. The distinctive thing here is how much of the testing can be done with data the institution already has.

Independent validation means independent of the builder. Somebody who did not configure the scenarios, reporting to somebody who does not own the outcome, with access to the configuration, the data and the alert history. A review performed by the team that runs the system is a self-assessment, whatever it is called.

What a validation should actually establish. Not whether the system is well regarded, but five concrete things: that the scenarios in production are the ones that were approved; that each one can fire and does; that the data they rely on is complete enough to support them; that the thresholds have evidence behind them; and that the coverage matrix accounts for the institution's stated risks.

Test end to end, not component by component. Put a transaction set through the live path — source system, feed, monitoring engine, alert queue, case file — and confirm it arrives. Component testing passes while feeds are broken, because each component works on the data it was given.

Known-case testing is the strongest evidence available. Replay the transactions from cases the institution reported. Do they alert under the current configuration? Every miss is either a tuning finding, a data finding or a coverage finding, and each is specific enough to act on. Few institutions do this; the ones that start usually find something in the first week.

Synthetic cases cover what history does not. For a typology the institution has never seen, construct the transactions it would produce and inject them into a test environment. This is the only way to test coverage of a risk that has not yet materialised, which is most of the risk assessment.

Every change gets tested before and measured after. Pre-implementation: replay over history, state the expected effect, check that no reported case would have been missed. Post-implementation: compare actual volume against the prediction, and investigate the gap rather than accepting it.

Keep a configuration baseline. What is in production, as a document, reconciled periodically against what is actually running. Drift between the two is common, arises from emergency changes nobody recorded, and is invisible until somebody compares.

And record the validation's findings as findings. With owners, dates and tracking, like any audit result. A validation whose output is a report nobody actions has cost money and changed nothing, which is worse than not having done it, because the institution now believes it has been validated.

Lesson 13 of 14

Running the queue, and closing the loop

A monitoring system produces alerts, and an institution's ability to act on them is a capacity question with a quality consequence. This lesson is about the operating model, and about the feedback that stops the system drifting.

Capacity is a design parameter, not a constraint to absorb. Expected alert volume, average handling time and available analyst hours give the arithmetic. Where volume exceeds capacity the options are more analysts, better scenarios, or fewer alerts by raising thresholds — and that third option is a risk decision masquerading as an operational one. It should be taken explicitly, at the right level, and recorded.

A backlog is a control failure with a date attached. An alert sitting for six weeks concerns activity that has already completed, possibly in an account that has already been emptied. Backlogs should be reported upward with their age distribution, not as a single number, because the tail is what matters.

Prioritise, but keep the floor. Risk-based ordering is sensible: higher-risk customers, larger amounts, higher-scoring alerts first. What is not sensible is a queue where the bottom is never reached, which is a threshold decision taken by accident. If low-priority alerts are not going to be worked, say so and change the scenario.

Quality assurance on closures, not just on reports. Sample closed alerts and re-review them. The question is whether the stated reason supports the decision and whether anything material was missed. This is where under-detection lives, and it is invisible in every volume metric.

Then feed what you find back into the system. This is the loop most programmes never close. If quality assurance finds that a scenario's alerts are almost all closed for the same reason, that reason is a tuning instruction: an exclusion to add, a threshold to move, a segment to split. Each closure reason, counted and reviewed monthly, is the cheapest tuning data an institution has, and it is usually discarded.

Measure the things that mean something. Alerts per analyst and closure rates measure throughput, not detection. More useful: the proportion of alerts leading to a case, the proportion of cases leading to a report, the age distribution of open alerts, the share of reports that originated from monitoring rather than from elsewhere, and the result of below-the-line testing.

Beware the metric that rewards closing. Where analysts are measured on volume closed, the rational behaviour is to close quickly, and the one real case in a hundred is closed with the rest. Any productivity measure needs a quality measure beside it or it will do harm.

And give the analysts a route back. The people reading alerts know which scenarios are useless, which exclusions are missing and which data fields are empty. A programme without a channel from the queue to the configuration is throwing away its best source of tuning information.

Lesson 14 of 14

One worked example, end to end

A mid-sized bank, retail and small business. Eighteen scenarios in production, about 4,000 alerts a month, nine analysts, and a monitoring programme nobody had examined for four years. A new head of financial crime asks three questions.

First: which scenario found the reports we filed? Thirty-one reports in the previous year. Nineteen came from branch staff referrals. Seven came from authorities' requests. Four came from one scenario on cash deposits. One came from a card fraud rule. None came from the other seventeen scenarios. That single table changed the conversation from alert volumes to coverage.

Second: can every scenario fire? The configuration was reconciled against production for the first time. Three scenarios had thresholds above anything the customer base had ever done. Two relied on a counterparty-country field populated for 11% of records. One had an exclusion list containing an internal account number that had been retired in a system migration two years earlier, which meant it was excluding nothing and alerting on every internal transfer — accounting for a third of total alert volume on its own.

Third: what are we meant to be watching for? The risk assessment named fourteen typologies. The coverage matrix, built over four days, mapped scenarios to six of them. Trade-based laundering, third-party funding and virtual asset exposure had no scenario at all. Structuring had five.

What was done, in order. The broken exclusion was fixed, which removed a third of the alerts in a week and freed the capacity to do everything else. The two scenarios depending on the 11% field were suspended and recorded as a data gap with an owner, rather than left running as apparent coverage. The three unfireable thresholds were re-set from the distribution of actual activity. Four of the five structuring scenarios were consolidated into one.

Then the gaps. Two new scenarios for third-party funding, written from the data the bank actually held, after establishing that the payer name was captured on 94% of deposits. Trade-based laundering was recorded as an accepted gap for that year, with reasoning, because the bank's trade exposure was small and the data to support a scenario did not exist.

And the loop. Closure reasons were counted for the first time. Two reasons accounted for 61% of all closures, and both became exclusions. Below-the-line testing was run on the cash scenario and found three cases in a sample of fifty that would have been escalated, so the threshold came down and the volume was absorbed by the capacity the exclusion fix had released.

Where it ended. Alert volume fell by about half; cases rose; reports rose to fifty-three the following year, of which twenty-two came from monitoring. Nothing about the software changed. What changed was that somebody asked which scenario had ever found anything.

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.

National detail, by market

This course teaches the international standard and one worked example. The national detail for a market sits in its own pack, so it can be kept current without reopening the course: Japan.

Sources

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

  1. The FATF Recommendations (Recommendations 20 and 21) · Financial Action Task Force
  2. Proceeds of Crime and Anti-Money Laundering Act, 2009 (No. 9 of 2009) · Kenya Law
  3. Proceeds of Crime and Anti-Money Laundering Regulations, 2023 (L.N. 153 of 2023) · Kenya Law
  4. Guidance to Reporting Institutions on Suspicious Transaction and Activity Reporting (2025) · Financial Reporting Centre
  5. Money Laundering and Terrorist Financing Trends and Typologies Report (2025) · Financial Reporting Centre
  6. goAML Registration Guide · Financial Reporting Centre
  7. Mobile payments statistics · Central Bank of Kenya