What a fraud operation is actually for
Most fraud training teaches you what fraud is. This course assumes you know, and asks a different question: how do you run the function that deals with it?
A fraud operation exists to make a large number of decisions, quickly, with incomplete information, about other people's money. That sentence contains every tension in the job. Decisions must be fast, because the money is leaving. They must be consistent, because inconsistency is unfair to customers and invisible to management. They must be defensible, because some of them will be examined a year later by somebody who was not there. And there must be a lot of them, because the alternative is a queue that grows until it is abandoned.
Three distinct jobs sit inside one team, and conflating them is the commonest structural mistake:
- Prevention changes the odds before anything happens — controls, limits, friction placed where it buys the most.
- Detection finds what prevention missed, in time to act.
- Response decides, acts, recovers what can be recovered, and reports what must be reported.
These have different clocks and different measures of success. Prevention is judged over quarters, detection over hours, response over minutes. A team measured on one will quietly starve the others — a desk rewarded for closing alerts will not invest in the control that would have stopped the alert existing.
The other thing to settle early is what the function is not. It is not the owner of the loss. It advises on where the loss sits and what it would cost to move it, but the decision to accept a given level of fraud in exchange for a given level of customer friction is a business decision, made by whoever owns the product. A fraud function that believes it owns that trade-off will fight battles it cannot win; one that documents the trade-off clearly will win most of them.
That distinction has a practical shape. The function's deliverable to the business is not "less fraud"; it is a priced set of options. This control costs this much friction and removes this much loss; that one costs less and removes less. Presented that way, a decision to accept more loss is a decision somebody made on the record, which is both fairer to the fraud team and far more defensible afterwards than a vague shared understanding that the appetite was high.
It is also worth being honest about what the function cannot do. It cannot make a badly designed product safe. Where a payment journey has no way to warn a customer, or an onboarding flow collects nothing that could later distinguish a real applicant from a fabricated one, no amount of monitoring recovers that. The earliest and cheapest fraud work is done in product design, which is why a fraud function with no route into design is permanently repairing what it could have prevented.
One classification, applied consistently
Before measurement, taxonomy. You cannot manage what you count inconsistently, and fraud is counted inconsistently almost everywhere.
The Federal Reserve built the FraudClassifier model for exactly this reason — to address what it calls "the industrywide challenge of inconsistent classifications for fraud". Its design choice is the one worth copying. The model classifies fraud independently of payment type, payment channel or other payment characteristics, and it begins with a single question: who initiated the payment? From there it separates payments initiated by authorised parties from those initiated by unauthorised parties.
Start there and a great deal falls into place. "Card fraud", "wire fraud" and "mobile money fraud" are not fraud types at all; they are channels. The same scam appears in all three, and an organisation that classifies by channel will report three unrelated problems and fix none of them. Whereas "the customer was deceived into authorising it" and "somebody else authorised it" are genuinely different events, with different controls, different liabilities and different conversations with the customer.
The Federal Reserve also publishes a companion ScamClassifier model, and says FraudClassifier can be used on its own or applied before or after it. That is a useful hint about structure: classifying how the money moved and classifying how the person was manipulated are separate axes, and squashing them into one list produces categories that overlap.
Three rules make a taxonomy survive contact with a real desk:
- One case, one class. If a case can be two things, the definitions are wrong. Fix the definitions, do not let analysts choose.
- Attributes, not categories, for everything else. Amount, date, channel, product and payment type ride alongside the classification. The Federal Reserve makes this point directly: adding payment attributes to the classifications is what reveals trends.
- Changing a definition is a dated event. Every reported series has a break in it the day you change what a word means, and a chart without that break marked is a lie told by accident.
The FraudClassifier model is published by the Federal Reserve's FedPayments Improvement programme. Its detailed tree and definitions are available after registration, so this lesson teaches the published principle rather than reproducing the model.
The queue is the product
Everything a fraud operation does arrives as a queue: alerts, referrals, customer reports, requests from other banks. The queue is where the function succeeds or fails, and it is routinely treated as plumbing.
First in, first out is the wrong default. It feels fair and it is indefensible. Arrival order correlates with nothing that matters — not exposure, not recoverability, not customer harm. A queue worked in arrival order spends its best hour on a £40 card dispute while a £400,000 payment sits behind it, past the point where a freeze would have worked.
Work the queue by what is still at stake. That means ranking by something closer to: amount still recoverable, multiplied by the probability the alert is real, divided by the time left to act. Nobody computes that exactly, but every queue design is an approximation of it, and most approximations are better than arrival order.
Some structural choices that decide whether a queue is workable:
- Separate the reversible from the irreversible. An alert on a payment that has not left has a different clock from one on a payment that has. They should not share a queue, because the second one's deadline is real and the first one's is not.
- Give every queue an owner and a target, or it will be nobody's. A queue with no named owner grows.
- Design for the bad day, not the average day. Volumes are not smooth; they spike when a campaign hits. A queue sized for the mean is already failing when you most need it.
- A queue that is never empty is not a queue, it is a backlog with ambitions. If the oldest item is older than the window in which action is possible, that work has no value and is consuming the capacity of work that does. Say so, measure it, and either resource it or stop pretending to do it.
The last point is the one that gets resisted and the one that matters most. A backlog of alerts nobody will ever reach is worse than no alerts at all, because it creates a record showing the institution was told and did nothing.
Deciding against the clock
Fraud response is one of the few compliance activities where the value of an action decays by the hour, and there is good public evidence for how steeply.
The FBI's Internet Crime Complaint Center runs a process called the Financial Fraud Kill Chain, in which law enforcement works with financial institutions to freeze fraudulently induced transfers. Its 2025 report gives the arithmetic: 3,900 incidents, covering $1,163,919,846 of attempted theft, with $679,013,183 frozen at a 58 percent success rate — split into 3,574 domestic incidents with $507,042,623 frozen and 326 international incidents with $171,970,560 frozen. The team has run since 2018.
One operational detail in that report is worth more to a fraud desk than the headline. For domestic cases the team will expand the freeze request beyond the first recipient bank to "second hop" transactions — but only "if information is provided during the FFKC initiation". In other words, whether the money that has already moved on can be chased depends on what you put in the first request. A desk that sends a thin initial referral and follows up with detail an hour later has, in practice, given up the second hop.
For scale underneath those recovery figures, the same report records 1,008,597 complaints in 2025 and $20.877 billion in reported losses, a 26 percent increase on 2024, at an average reported loss of $20,699.
The United States Financial Crimes Enforcement Network reports the same shape from its own Rapid Response Program, which has assisted in recovering over $515 million with the assistance of 64 countries. Its advisory states the operational finding plainly: FinCEN "has had greater success in recovering funds when victims or financial institutions report BEC-unauthorized and fraudulently induced wire transfers to law enforcement within 24 hours".
Two independent authorities, the same conclusion. Recovery is a function of speed, and the window is measured in hours. That has concrete consequences for how a desk is built:
- Design the escalation path backwards from the deadline. If a freeze request must be out within hours, every step before it — detection, triage, verification, authorisation — has to fit inside that. Count the steps and the approvals; if they cannot fit, the path is decorative.
- Somebody must be able to act out of hours. Fraud does not observe business hours, and a 24-hour window that starts on Friday evening is gone by Monday.
- Authority has to sit low enough to be used. A control requiring a director's approval to request a freeze will not be exercised at 03:00, which means in practice it does not exist.
- Measure time-to-freeze, not just time-to-decision. The decision is not the outcome. The money moving back is.
One caution worth keeping: these figures describe the reporting populations of two American bodies, and they are attempts and freezes rather than final restitution. The direction is solid and the precise rate is not transferable.
Thresholds, and who is allowed to move them
Every automated fraud control has a number in it somewhere: a score above which a payment is held, a velocity above which a card is blocked, an amount above which a human looks. Those numbers are the most consequential and least governed objects in the function.
A threshold is not a technical setting. It is a priced decision about how much loss to accept in exchange for how much customer friction, and it should be treated with the seriousness of any other pricing decision. Move it one way and losses fall while more legitimate customers are stopped; move it the other and the complaints stop and the losses arrive. There is no setting that is simply correct, which is precisely why the reasoning has to be written down.
What a defensible threshold change looks like:
- A stated objective. Which quantity is being improved, and at whose expense.
- Evidence from the operating point you are actually at, not from a model's overall accuracy. What matters is the behaviour of the control at the number you are choosing — how many true events it catches there, and how many legitimate customers it stops there.
- A named owner who can authorise it, and a different named person who reviews it.
- A dated record of what changed, from what, to what, and why — because the first question in any later review is what the control was set to at the time, and an undocumented change makes that unanswerable.
- A rollback condition agreed in advance. What would tell you this was wrong, and how quickly would you see it.
The failure mode to watch for is the quiet drift: a threshold nudged a little every few weeks by whoever is on shift, each change reasonable, the cumulative effect never reviewed. After a year the control is nothing like the one that was approved and nobody can say when it changed. A single dated log of threshold changes, with the reason, prevents an entire category of argument.
Measuring the function honestly
Fraud functions are measured badly more often than they are measured well, and the bad measures share a feature: they are easy to collect and they reward activity rather than outcome.
Measures that flatter:
- Alert volume. A number that goes up when the rules get worse. It describes the sensitivity of a control, not the performance of a team.
- Cases closed. Closing a case is not an outcome; it is the end of an administrative process. A desk measured on closures will close things.
- Overall model accuracy. In a population where fraud is rare, a control that catches nothing is accurate almost all of the time. The number is true and useless.
- Fraud prevented. The hardest to resist, because it is the number everyone wants. It is an estimate of a counterfactual, and it is only honest when stated as one, with its assumptions written beside it.
Measures that bite:
- Loss rate per channel, per unit of volume, so that growth does not disguise deterioration. Absolute losses rising while the rate falls is a success; the reverse is a problem that absolute numbers hide.
- Catch rate and false-positive rate at the operating point, reported together. Either alone can be moved by moving the threshold, which is why neither means anything on its own.
- Time to decision, and time to freeze, as distributions rather than averages. The average hides the tail, and the tail is where the losses are.
- Reopened and overturned decisions, which measure whether the first decision was right.
- Queue age at the oldest item, which is the earliest honest signal that capacity has been exceeded.
One discipline matters more than the choice of metric: report the same number the same way every period, and mark every definition change on the chart. A fraud report whose definitions move is not a time series, and most of the apparent improvements in this field are definition changes nobody labelled.
Two reporting traps specific to fraud are worth naming. The first is the maturity lag: fraud is reported late, sometimes months late, so the most recent period always looks better than it will turn out to be. A chart that shows the current month beside settled months invites a conclusion the data cannot support. Show the lag, or hold the recent periods open.
The second is survivorship in the denominator. If a control blocks a cohort of applicants, the losses from that cohort never appear, so the control looks free. Its real cost is in the legitimate customers inside that cohort, who are now somebody else's customers, and that number is nowhere in the fraud report. It belongs in the conversation even when it cannot be measured precisely, because leaving it out systematically biases every decision towards more friction.
Quality assurance of decisions
If the function makes thousands of decisions, somebody has to check whether they were right — and "right" has to mean something other than "nobody complained".
Sample closed cases, not open ones. An open case under review behaves differently from a case closed under pressure on a busy afternoon. The population you care about is the decisions that stood.
What a reviewer should be looking for, in order of seriousness:
- Was the decision supported by what was on the file at the time? Not whether it turned out correct — that is hindsight — but whether a reasonable person with the same information would have reached it.
- Was anything obvious not checked? The commonest defect in a real case file is an unasked question, not a wrong answer.
- Is the reasoning recorded at all? A file whose decision is a status change with no note cannot be reviewed, which means it cannot be defended.
- Would a different analyst have decided the same way? This is the one worth measuring directly.
Calibration is the test that finds real problems. Give the same set of cases to several analysts independently and compare. Disagreement is not a performance failure; it is information about where the guidance is ambiguous. High disagreement on a category means the policy is unclear, not that somebody needs retraining, and fixing the policy fixes every future case at once.
Keep quality assurance structurally separate from the decision it reviews. A desk reviewing its own closures will converge on its own habits, and the point of the exercise is to find the habits.
Finally, treat customer complaints and overturned decisions as free quality data. They are a biased sample — you only hear from people motivated to push back — but they are the cheapest source of genuine false positives you will ever get, and a function that does not route them back into rule tuning is throwing evidence away.
One structural point about sampling. A fixed percentage of all closures sounds fair and wastes most of the effort, because the great majority of decisions are uncontroversial. Weight the sample: more of the decisions that were fast, more of those made by the newest analysts, more of those on the categories where calibration showed disagreement, and more of those where the outcome was severe for the customer. The point of a sample is to find defects, not to be representative.
And review the cases nobody worked. Alerts that aged out, referrals auto-closed by a rule, queues that were cleared in bulk at the end of a bad week — these never appear in a sample of decisions, because no decision was made. They are frequently where the real losses are, and a quality programme that only inspects completed work will never see them.
Investigating at depth
Triage decides whether something is worth looking at. Investigation builds the picture. The step most often skipped is the one between them: deciding what question the investigation is supposed to answer.
An investigation without a question expands forever. It accumulates screenshots, pulls another statement, requests another log, and ends when the analyst is tired rather than when the matter is resolved. Write the question at the top of the file — "did this customer authorise this payment", "is this account controlled by its holder", "are these twelve accounts one network" — and the scope follows from it.
Then build out along the dimensions that actually connect cases:
- Accounts and the money — where it came from, where it went, what it was converted into, how fast.
- Devices and sessions — what the activity was done from, and what else that thing has touched.
- Counterparties — who is on the other side, and who else they appear beside.
- Timing — the sequence, and the gaps. Time is the most under-used evidence in fraud work; the order in which things happened frequently settles a case that no single artefact settles.
- Behaviour against baseline — not "is this unusual" but "is this unusual for this customer".
Hold the difference between a hypothesis and a conclusion. A hypothesis is what you are testing; a conclusion is what the evidence supports. Fraud investigations go wrong when an early hypothesis quietly becomes the frame, and everything afterwards is read as confirming it. The discipline that helps is cheap: write down what you would expect to see if the hypothesis were wrong, then look for that.
And know when to stop. Stop when the question is answered, when the next step costs more than the answer is worth, or when further work cannot change the decision. Recording why you stopped is as important as recording what you found, because "we stopped here deliberately, for this reason" is a defensible position and silence is not.
Evidence that survives being looked at
A fraud case is read by people who were not there: a second-line reviewer, an ombudsman, an auditor, a regulator, occasionally a court. The file has to work for them, and most files do not.
Write contemporaneously. A note made at the time carries weight a reconstruction cannot. A file assembled afterwards from memory is weaker evidence even when every statement in it is true, and if it is presented as contemporaneous when it is not, that is worse than the original failure.
Record provenance for everything. Where each item came from, when it was obtained, who obtained it and what it is. A screenshot with no source and no timestamp proves very little. An export with the query that produced it, the system it came from and the time it was run proves a great deal.
Separate the three things a file contains, and label them: what was observed, what was inferred, and what was decided. Most weak case files blur them, so the reader cannot tell which parts would survive challenge.
Keep what the standard requires, for as long as it requires. The FATF's Recommendation 11 sets the international floor: records of transactions, records obtained through customer due diligence (CDD, which much of the industry calls KYC), account files and business correspondence, including "the results of any analysis undertaken", kept for at least five years after the relationship ends or the occasional transaction. Your investigation notes are the results of an analysis. National law frequently requires longer, and where it does, the longer period governs.
Two practical tests for a case summary. Could a colleague act on this without speaking to you — place a freeze, answer a complaint, file a report? And would you be content for it to be read aloud, in full, by somebody hostile? If the answer to either is no, the file is not finished.
Three habits that cost nothing and change the quality of a file completely:
- Timestamp in one zone, and say which. A case involving three countries and three local times is routinely reconstructed wrongly a year later. Pick coordinated universal time or a single named zone, state it on the file, and never mix.
- Record what you asked for and did not get. A request to another institution that went unanswered is evidence about the case and about the institution. A file that records only what arrived implies nothing was sought.
- Write the decision as a sentence with a subject. "Blocked" is a status. "I blocked the account because the three payments matched the pattern in alert 4471 and the customer could not be reached" is a decision, and only the second one can be reviewed.
And keep the opinion out of the observation. "The customer was evasive" is an inference presented as a fact; "the customer ended the call twice when asked about the counterparty" is the observation that supports it, and it is the one that survives being challenged.
Mules, and acting on the network
Fraud needs an exit. Stolen money has to land somewhere it can be moved on from, and that somewhere is usually an account in somebody's real name. Mule accounts are where fraud operations most clearly stop being about transactions and start being about networks.
The scale is documented. Europol's assessment of organised internet crime describes the ninth round of the European Money Mule Action, run across three months of 2023 by 26 countries with Europol, Eurojust, INTERPOL and the European Banking Federation. It identified 10,759 money mules and 474 recruiters, with 1,013 individuals arrested worldwide, and 2,822 banks and financial institutions collaborated with law enforcement alongside money transfer services, cryptocurrency exchanges, travel providers and identity verification companies.
Read that last figure again: nearly three thousand institutions, for one coordinated action. The implication is that mule networks are not visible from inside one bank. The accounts that look merely odd in your data look obviously coordinated in the aggregate, which is the argument for participating in industry information sharing rather than treating detection as purely internal.
Operationally, the difference between acting on an account and acting on a network:
- An account-level response closes one account and tells the holder why. The network re-routes within days.
- A network-level response looks for what the accounts have in common — the recruiter, the onward counterparty, the device, the timing pattern, the shared contact detail — and acts on that, accepting that some of the accounts are held by people who were themselves recruited and deceived.
Europol's assessment also notes the direction of travel worth planning for: fraud is the most frequently identified predicate offence involving illegal use of cryptocurrencies, and the misuse of virtual IBANs is flagged as a potentially growing threat because it puts distance between an account identifier and the institution that actually holds the funds.
The mule who is a victim is the hardest case in this area, and the honest position is that your operation will sometimes be unable to tell the difference quickly. Design the response so that being wrong is recoverable.
Recovery, and what is realistically possible
Recovery is the part of fraud operations most often promised and least often understood. Three different things get called it, and keeping them apart is the first job.
- Freezing stops the money moving further. It is fast, it is the only step with a real deadline, and it is the one that makes everything else possible.
- Recovering returns the money to where it came from. It depends on the freeze having worked, on the receiving institution's cooperation, and frequently on a law enforcement channel.
- Reimbursing makes the customer whole from the institution's own funds. It is a liability decision, not a recovery, and conflating it with recovery produces both bad accounting and bad customer conversations.
The public channels exist and have published results. The FBI's Financial Fraud Kill Chain froze $679,013,183 in 2025 across 3,900 incidents, at a 58 percent success rate — $507,042,623 of it domestically and $171,970,560 internationally. FinCEN's Rapid Response Program, working with the FBI, the Secret Service, Homeland Security Investigations, the Postal Inspection Service and counterpart financial intelligence units, has assisted in recovering over $515 million across 64 countries, with international requests coordinated through its Rapid Response Team.
What that means for a desk outside the United States: the mechanism you need is an equivalent domestic channel, identified before you need it. Find out now who to contact, through what route, with what information, and how long they take. A fraud operation discovering the answer during its first large incident has already lost the hours that mattered.
Set expectations honestly with customers, too. Recovery is not assured — FinCEN says so in terms — and a desk that implies otherwise converts a fraud loss into a complaint about the institution. "We have acted within the hour and here is what happens next" is both true and more reassuring than a promise nobody can keep.
What to have ready before the first large case, rather than during it:
- A named contact and a route at every institution you move money to most. Not a general enquiries address. The person or desk that can act on a freeze request out of hours.
- A template referral that is complete on the first send. The FBI's process will chase funds beyond the first recipient only on the information supplied at initiation, and other channels behave the same way, so a referral designed to be complete at the first attempt is worth more than a fast partial one.
- A decision already made about who may commit the institution. Recovery requests sometimes require an undertaking or an indemnity. If nobody knows who can give one, nobody gives one.
- A log of what each channel actually delivered. Over a year this tells you which routes work, which is the only way to prioritise effort honestly rather than by reputation.
The hand-off to anti-money laundering
Fraud and money laundering are not the same thing, and the relationship between them is simple to state: fraud is a predicate offence. The proceeds of a fraud are criminal property, and moving them is laundering. So a fraud case will often also be a reportable suspicion, and the two desks have to agree who reports it.
The duty itself is short. The FATF's Recommendation 20 requires that 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. The filing is called a suspicious transaction report (STR) in the FATF standards and a suspicious activity report (SAR) in the United States and the United Kingdom; it is the same filing, and a fraud team working across both will see both names. Note what it does not say: it fixes no number of days. That is a national choice, it is frequently very short, and you need to know both the deadline where you operate and whether it is counted in calendar or working days.
Recommendation 29 is the other half: countries should establish a financial intelligence unit as the national centre for receiving and analysing those reports. Its name differs everywhere, and knowing the name of yours is the minimum.
Then the rule that catches fraud teams out. Recommendation 21 requires that institutions and their staff be prohibited from disclosing the fact that a suspicious transaction report is being filed. A fraud investigator's instincts run the other way — you want to warn the customer, ask them a direct question, tell the receiving bank exactly why. Where a report has been or is being made, some of that is forbidden. The fraud conversation and the suspicion report have to be kept separate, and the people doing fraud work need to know where the line is before they trip over it.
The organisational failure to design against is the mutual assumption: the fraud desk assumes the AML desk filed because they could see the case, the AML desk assumes the fraud desk filed because they owned it, and nothing was filed. Prevent it with one unambiguous rule — a named owner for the reporting decision on every case type, written down, with the handover recorded on the file rather than agreed in a corridor.
What the law lets you do with the data
An investigation is itself a processing activity. You are gathering personal data about identified people, frequently without telling them, in order to make decisions that affect them. That is lawful in most places and it is not unconstrained, and fraud teams are often the least briefed function in the building on the point.
Four questions to be able to answer about your own operation:
- On what basis are you processing? There is normally a route — preventing fraud is a recognised purpose nearly everywhere — but "normally" is not an answer and the route differs by jurisdiction.
- How much are you keeping, and for how long? An investigation file accumulates far more than the decision needed. Every extra item is a liability if the file is breached, and a retention period that is simply "forever" is not a policy.
- Who can see it? A fraud case file is some of the most sensitive material an institution holds about a customer, including an accusation that may be wrong.
- What happens when the person asks? Many regimes restrict what must be disclosed where it would prejudice fraud prevention. Knowing which restriction applies, before the request arrives, is the difference between a considered refusal and an accidental disclosure.
The variation between jurisdictions is wider than teams expect, and two examples make the point:
- The right to a human decision is not universal. The European reference is Article 22 of the General Data Protection Regulation, which is the provision most national laws on this subject are modelled on. But India's Digital Personal Data Protection Act, 2023 confers rights to information, correction, erasure, grievance redressal and nomination and says nothing about automated decisions — "automated" appears in it only in the definitions. In Japan, the words "automated" and "profiling" do not appear in the Act on the Protection of Personal Information at all. If your automated blocking design assumes a human reviewer will be legally required, that assumption does not travel.
- Breach duties differ in shape, not just in hours. India requires intimation to the regulator and to every affected person, and fixes no number of hours in the Act. Japan's Article 26 requires a report to its Commission in situations prescribed by Commission Order, so the deadline is not in the statute at all.
The practical instruction is the same as everywhere else in this catalogue: find the rule where you operate, and do not assume the one you learned is the one that applies.
Running the function on the worst day
The last lesson is about capacity, people and change, because a fraud operation is judged on its worst day rather than its average one.
Staffing. Volume is spiky and the spikes are adversarial — a campaign arrives when it suits the attacker, frequently overnight or across a holiday. Staff to the peak you intend to survive, decide in advance what gets dropped when you are over capacity, and write that down. A documented degradation plan ("above this volume we stop doing X") is a professional position; discovering the answer live is not.
People. This work has a high emotional load. Analysts speak to people who have lost their savings, and they are measured on speed. Attrition is high, and every departure takes the pattern recognition that took a year to build. Two things help more than anything else: rotation away from the hardest queues, and feedback — most fraud analysts never learn how their cases ended, which is both demoralising and the reason the same mistakes repeat.
Change. Rules and models are changed constantly, by people under pressure, in production. Every change needs a before-and-after measurement, a named owner, a dated record and a way back. The most dangerous change is the small one made at speed during an incident, precisely because it is made under the conditions where review is skipped — so log it during the incident and review it after.
The incident itself. When a live campaign hits, three things matter in order: stop the bleeding (a blunt control is acceptable, a precise control that takes a day is not), preserve the evidence (the logs you will need are often the ones that rotate), and communicate honestly internally (an incident where the fraud desk understates the scale to avoid alarm is an incident that gets worse).
And afterwards, run the review. Not to assign blame, but to answer two questions: what control would have stopped this, and what did we learn that we must not relearn? Write the answers somewhere that will be read, because an operation that does not accumulate learning is running the same incident repeatedly and calling it bad luck.
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.
Your answers
Your certificate of completion
Enter your name as you want it to appear, then save the certificate as a PDF. In the print window, choose Save as PDF. A certificate is issued once per completion of the course.
Rate this course
How would you rate this course? Your rating is anonymous — just a number, with no name and no account.
Bring this to your team
Your colleagues can take this course too — it is free, needs no sign-up, and nothing they enter leaves their browser. If your institution wants the same ground covered as structured training for a team, with an instructor and your own cases, we run courses in a classroom or online for groups of ten or more.
Talk to us about training your team → See the paid courses →
Saolix does not record who takes this course, so it cannot verify these certificates. The certificate confirms completion of a free self-paced course and is not an accredited qualification.
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.
- Internet Crime Report 2025 · Internet Crime Complaint Center, Federal Bureau of Investigation (US)
- Updated Advisory on Email Compromise Fraud Schemes Targeting Vulnerable Business Processes (FIN-2019-A005) · Financial Crimes Enforcement Network (US)
- About the FraudClassifier Model · FedPayments Improvement, Federal Reserve
- Internet Organised Crime Threat Assessment (IOCTA) 2024 · Europol
- The FATF Recommendations (updated October 2025) · Financial Action Task Force
- The Digital Personal Data Protection Act, 2023 (No. 22 of 2023) · Ministry of Law and Justice, Government of India
- Act on the Protection of Personal Information (Act No. 57 of 2003), English text · Japanese Law Translation, Ministry of Justice, Japan



