Leaving this course?

Lesson 1 of 12

The people who left, and why nobody counts them

An onboarding process produces three outcomes, and most institutions measure two of them. Accepted customers are counted because they generate revenue. Rejected applicants are counted because the decision was recorded. The third group — the people who started and stopped — leaves no record of itself at all.

Why that group is the subject of this course. A refusal is a decision somebody made, which can be reviewed, appealed and audited. An abandonment is a person who closed the application and did something else. Nothing in the file says why, nobody is accountable for it, and the only trace is a gap in a funnel report that may not exist.

Two courses in this catalogue already establish the principle that identity requirements exclude people: the remote identity proofing course has a lesson on inclusion without lowering the bar, and the Kenyan digital identity course has one on who gets shut out. This course assumes both and does something different — it treats the funnel as an operational system with numbers in it, thresholds that somebody chose, and a cost that lands on the business.

What the scale looks like. The World Bank's 2025 Global Findex reports that "Worldwide, 79 percent of adults have an account at a bank or similar financial institution, with a mobile money provider, or both, up from 74 percent in 2021" — real progress, against which "1.3 billion adults worldwide still lack financial accounts and are thus unable to benefit directly from the formal financial system."

And the reason is usually not the absence of identity. The same report finds that "Across low- and middle-income economies, 95 percent of adults have ID." Documentation is not the universal barrier it is assumed to be — but where it bites, it bites hard: "15 percent of adults without an account in Sub-Saharan Africa do not have a mobile money account because they lack the documentation they would need to open one."

One figure that shows how uneven it is. On the related problem of buying a SIM card, the report records that among adults without identification, "In the Republic of Congo, 58 percent of adults without ID experienced this challenge; in Guinea, 57 percent did; in Tanzania, 67 percent did; and in Uganda, 79 percent did."

Those numbers vary by a factor approaching two between neighbouring countries, which tells you the barrier is a function of local process rather than of technology.

The frame for the next eleven lessons. Every step in an onboarding journey has a pass rate, and every pass rate was set by somebody — a threshold, a document requirement, a data check, a rule about what counts as proof of address. Each of those decisions excludes a measurable number of legitimate people, and almost none of them was made with that number in front of the person deciding.

And one thing this course will not do. It will not argue that controls should be weakened. The question throughout is narrower and more useful: of the people your process refuses or exhausts, how many should have passed, and what would it cost to find out?

Terminology: the FATF standards say customer due diligence (CDD) and the phrase "KYC" appears nowhere in them. The industry and many national rule books say KYC, usually for the identity part of the same obligation. Where this course says one, read it as the other unless the context narrows it.

Lesson 2 of 12

Two metrics, defined, that most institutions do not have

You cannot manage a funnel you do not measure, and the two measures that matter most have standard definitions that are worth adopting verbatim rather than inventing.

They come from the digital identity guidelines. NIST's framework defines, among its performance metrics:

"Abandonment Rate (Overall): Percentage of unique users who start the identity proofing process but do not complete it without failing a process."

"Fail Rate (Overall): Percentage of unique users who start the identity proofing process but are unable to successfully complete all of the steps."

Read the first definition carefully, because the qualifier at the end is the whole point. Abandonment counts people who left without being failed. They were not refused. The process did not reject them. They stopped — and under any ordinary management reporting they are invisible, because no decision was recorded against them.

Why the distinction changes what you do. A high fail rate is a threshold problem: something in the process is rejecting people, and you can find it by looking at which step produced the refusal. A high abandonment rate is an experience problem: something is costing people more than the account is worth to them, and the cause may be a step that nobody considers difficult.

The two have different owners, too. Fail rates belong to risk and compliance, because they follow from control settings. Abandonment belongs to the product, because it follows from design. Which is precisely why neither function has historically owned the combined number.

Overall is not enough. Both metrics need to be broken down the same way, and the breakdown is where the findings are:

  • By step. Which stage lost them: document capture, the selfie, the data check, the questionnaire, the wait.
  • By attempt. Someone who failed capture three times and left is a different problem from someone who left at the first screen.
  • By channel and device. An older phone with a poor camera produces a different funnel from a current one.
  • By cohort, where it is lawful to measure: age, location, product. Lesson five explains why the aggregate figure can look acceptable while a group inside it is being refused at several times the rate.

And the guidelines name the harm in their own words. Among the risks to be considered in tailoring a process is "the possibility of legitimate users who are unable to access desired online services or experience sufficient friction or frustration with the identity system (and technology selection) that they abandon attempts to access the online service."

Friction and frustration, named as a risk alongside the fraud risks. That framing is worth borrowing when you need to argue for this work internally: exclusion is a risk the framework expects you to assess, not a commercial nicety competing against risk management.

Lesson 3 of 12

The false reject is a number you chose

Every biometric and matching step in an onboarding journey has two error rates that move in opposite directions, and the threshold that decides the trade is set by somebody — usually once, early, by a vendor's default.

The requirements are specific. NIST's identity proofing guidelines state performance thresholds for one-to-one comparison used to verify against a claimed identity: "False match rate: 1:10,000 or better False non-match rate: 1:100 or better".

Those two numbers are the whole subject of this lesson. A false match rate of one in ten thousand is a security requirement: at most one impostor in ten thousand attempts should be accepted. A false non-match rate of one in a hundred is an inclusion requirement: at most one genuine applicant in a hundred should be wrongly refused.

The second is a floor on how badly you may treat legitimate customers, written into a security standard. It is rarely read that way.

One in a hundred is not a small number. Run a hundred thousand applications a year at exactly that rate and a thousand genuine people are refused by the matcher alone — before document capture failures, before data mismatches, before the people who simply gave up. Those thousand do not know why they were refused, and most will not come back.

The threshold must be fixed, and that has a consequence. The guidelines require that "The biometric system SHALL be configured with a fixed threshold; it is not feasible to change the threshold for each demographic group."

This is worth understanding properly, because it is sometimes misread as indifference to fairness. It is the opposite. Tuning a threshold per group would mean applying a different standard of proof to different people, which is both unlawful in most places and operationally indefensible. So the threshold is fixed, and the obligation moves to the algorithm:

"Biometric verification technologies SHALL provide performance for applicants of different demographic types that is no more than 25 % worse than the performance for the overall population."

A bounded differential, at a fixed threshold. You cannot fix a biased matcher by adjusting it per group; you have to choose a matcher whose performance is within that bound for the population you actually serve.

Liveness has its own number. The same guidelines set a presentation attack requirement: the system "SHALL implement presentation attack detection (PAD) capabilities that meet the impostor attack presentation accept rate (IAPAR) performance metric of <0.07", with all testing "conformant to ISO/IEC 30107-3:2023".

Note what is specified and what is not. The attack acceptance rate is bounded; the rate at which genuine users are wrongly flagged as attacks is not. That second rate is where a liveness step quietly becomes an exclusion mechanism, and it is the number to demand from a vendor who has only quoted you the first.

The practical instruction. Find out what thresholds your journey actually runs at, who set them, when, and against which test population. In most institutions the answer to at least two of those four questions is nobody knows.

Lesson 4 of 12

The manual route is a requirement, not a courtesy

An automated process that refuses people needs a route for the people it refuses. In the identity proofing guidelines that route is mandatory in one case and expected in another.

The mandatory case. Where a one-to-many biometric search is used, the guidelines are categorical: a provider "SHALL NOT decline a user's enrollment without a manual review to confirm the automated search results".

A machine may not have the last word. The automated search can flag; a person has to confirm before an applicant is turned away. That is a design requirement with staffing consequences, and it is the single clearest statement in any standard that an unreviewed automated refusal is not acceptable.

The expected case. More broadly, the guidelines say providers "SHOULD offer trusted referee services for failures of automated verification processes (e.g., biometric comparisons)."

A trusted referee is a trained representative who can establish identity by other means when the automated path fails. The guidelines list among those who may need one "persons without a bank account or with limited credit history" — which is to say, exactly the population that an automated check built on financial records will fail.

Why this is the highest-value control in the whole course. Everything else reduces the number of wrongly refused applicants at the margin. A working manual route recovers them. It converts a refusal into a decision, which can be right, and it gives you the only reliable source of information about why your automated process is wrong — because the referee sees the cases the machine could not handle and can tell you what they have in common.

What makes a manual route real rather than nominal. Five things, and most institutions have one or two:

  • It is discoverable. Offered at the point of failure, in the flow, not described in a help article the refused applicant will never open.
  • It is staffed. A route with a two-week queue is a route nobody completes.
  • The staff can actually decide. A referee who may only re-run the same automated check adds a delay and changes nothing.
  • Its decisions are recorded as decisions, with reasons, so that the pattern across them is analysable.
  • And it is measured. How many were offered it, how many took it, how many passed, and how long it took.

The number that tells you whether it works. The proportion of applicants who pass the manual route. If it is very high, your automated threshold is too strict and the referee is correcting a systematic error. If it is very low, either the route is reaching the wrong people or the refusals were right. Either answer is actionable, and an institution that does not measure it has a manual route whose purpose is to exist rather than to work.

Lesson 5 of 12

Why the aggregate figure hides the problem

A funnel that performs acceptably on average can be refusing one group at several times the rate of another, and the aggregate number will not show it.

The evidence base. NIST's 2019 study of demographic effects in face recognition remains the clearest public measurement, and its headline finding is about asymmetry: "False positive rates often vary by one or two orders of magnitude (i.e., 10x, 100x). False negative effects vary by factors usually much less than 3."

Two different problems in one sentence. False positives — wrongly matching two different people — vary enormously between groups. False negatives — wrongly failing to match the same person — vary much less, by factors under three.

For an onboarding funnel the second is the one that excludes, and a factor of three is still a factor of three. The study also found that "false negatives are generally higher in people born in Africa and the Caribbean, the effect being stronger in older individuals", on border-crossing images, and that false positives were "between 2 and 5 times higher in women than men, the multiple varying with algorithm, country of origin and age."

This ground is covered more fully in the eKYC service and biometrics courses. What matters here is the operational consequence: a differential of three in the false non-match rate, applied to a population segment, is a differential of three in who gets refused.

And the dominant variable is not the algorithm. NIST's current programme page is explicit: "False negatives are strongly dependent on image quality, and poor photography of a face can induce a demographic effect."

Induce. The differential can be created by your capture flow rather than inherited from the matcher — bad lighting guidance, a cheap camera, a compression setting, a flow that does not tell people to remove a hat. An institution that attributes its demographic differential to the vendor, without first measuring its own capture quality, may be blaming the wrong component.

Which also means it is fixable without changing the matcher. Capture quality is a product problem with product solutions: on-device quality feedback before submission, a retry that explains what was wrong, and refusing a poor capture rather than scoring a poor face. Lesson eight returns to this.

The other kind of hidden exclusion. Biometrics are not the only step with a differential. Consider who fails each of the others: a data check against a credit record fails the young and the recently arrived; a proof-of-address requirement fails people in informal housing and those living with family; a document requirement fails anyone whose document has expired, which correlates with income; and a requirement to complete a journey in one sitting fails people on shared devices and metered data.

None of those is a biometric and all of them produce a differential.

So the instruction is the same as lesson two's. Break every stage's pass rate down by whatever cohorts you may lawfully measure, and look at the spread rather than the average. The institutions that discover a problem here are the ones that looked; the rest have the same problem and an aggregate number that reassures them.

Lesson 6 of 12

Tiering, with the numbers two regulators actually set

The standard answer to exclusion is a tiered approach: lower checks for a product with lower limits, raised when the customer wants more. It is a good answer, and it is usually discussed without numbers, which makes it hard to implement. Two regulators publish theirs.

India's small account. The Reserve Bank's know-your-customer directions for commercial banks define a limited account with explicit caps. Among its conditions: "the aggregate of all withdrawals and transfers in a month does not exceed Rs 10,000; and ... the balance at any point of time does not exceed Rs 50,000." The aggregate of credits in a year is capped at one lakh.

It is also time-bounded: "The account shall remain operational initially for a period of 12 months which the bank can extend for a further period of 12 months". So it is a route in, not a permanent category — twelve months, extendable once, by which time fuller verification is expected.

And it carries operational conditions that reveal the risk being managed: such accounts may be opened "only at Core Banking Solution (CBS) linked branches or in a branch where it is possible to manually monitor and ensure that it does not credit foreign remittances to the account." The concern is cross-border flows into a thinly verified account, and the control is a restriction on where it can be opened.

Nigeria's three tiers. The Central Bank of Nigeria's mobile money framework sets out limits by KYC level in a table: tier one at fifty thousand naira daily outflow with a cumulative balance limit of three hundred thousand; tier two at two hundred thousand and five hundred thousand; tier three at five million with an unlimited balance. The framework adds that "The Bank may determine the transaction limits from time to time", so the figures are a current setting rather than a permanent rule.

What the pattern teaches, independent of the figures. Three design properties worth copying:

  • The limit is the control. Lower verification is made safe by capping exposure, not by accepting more risk. The arithmetic is explicit and checkable.
  • The tier is a route, not a destination. Both regimes expect progression, and India's puts a clock on it.
  • The restrictions are specific to the risk. Not "enhanced monitoring" in the abstract, but a named prohibition on the flow that worries the regulator.

And a caveat from one national rule book. The Central Bank of Kenya's 2013 guidelines for mobile payment AML set a single identification rule rather than tiers: "Mobile payment accounts shall be opened using valid identification documents only", with acceptable documents listed, and investigation triggers at "a daily turnover of KShs 100,000" or "more than KShs 300,000 per week".

One rule, not a ladder. Whether later instruments have changed that was not established for this course, so a Kenyan institution should check the current position rather than assume a tiered regime exists. The international standards permit simplified measures where risk is lower; whether your own regulator has implemented them as tiers is a separate question with a local answer.

Lesson 7 of 12

What counts as proof when the usual proof does not exist

Most onboarding requirements assume a person holds a current government document and can evidence an address. The requirement is rarely stated that way, which is why the exceptions are not planned for.

A supervisor has said what the alternatives may be. The European Banking Authority's guidelines on managing financial crime risk while providing access to financial services require institutions to be specific rather than vague about this: "Institutions' policies and procedures should specify which alternative, independent documentation it can rely upon. This documentation may include, where permitted under national law, expired identity documents and documentation provided by an official authority, such as social services".

Two things in that sentence are unusual and useful. Expired identity documents are named as potentially acceptable — which is a direct contradiction of the near-universal rule that a document must be in date. And documentation provided by an official authority such as social services is named, which opens a route for people in care, in hostels, newly released from institutions or recently arrived.

The requirement is that the policy specifies which alternatives it accepts. A policy that is silent leaves a front-line officer to refuse, because refusing is the safe choice for an individual with no written authority to accept anything unusual.

What the UK conduct regulator expects. Its review of payment account access records the expectation plainly: "we expect all firms providing payment accounts to make sure customers understand what steps they need to take to provide acceptable forms of non-standard ID".

Note understand. Not that an alternative exists somewhere in a policy, but that the customer can find out what it is. And the regulator quotes its own rulebook on the underlying duty: firms should "do not unreasonably deny access to its services to potential customers who cannot reasonably be expected to produce detailed evidence of identity."

It goes further on the quality of the reasoning: "Policies that allow for the rejection of customers on grounds that appear spurious, speculative or unfairly biased against certain customers or groups of customers are unlikely to meet firms' obligations under the Duty."

Spurious, speculative or unfairly biased. That is a standard applied to the policy, not to the individual decision, which means a policy producing those outcomes is the finding.

And a right of access in one regime. European law on payment accounts gives consumers "legally resident in the Union, including consumers with no fixed address and asylum seekers" the right to open and use a payment account with basic features, and requires a decision "without undue delay and at the latest 10 business days after receiving a complete application."

Consumers with no fixed address and asylum seekers, named. And a ten-business-day clock. Its recital is blunt about the limits of the excuse: "a refusal can only be justified where the consumer does not comply with that legislation and not because the procedure to check compliance with the legislation is too burdensome or costly."

Too burdensome or costly is not a defence. That sentence is the whole argument of this course in a single clause of European law, and it predates most of the technology.

Lesson 8 of 12

Capture quality, and the friction you can remove for nothing

A large share of onboarding failures are not decisions about a person. They are a photograph that was never going to work, scored anyway.

The distinction that saves the most applicants. There are two different events in a failed biometric step: the system could not get a usable image, and the system got a usable image and concluded it did not match. The first is a product failure and is recoverable in the moment. The second is a decision.

Most funnels conflate them, which produces a refusal for a problem that could have been fixed by asking the person to move nearer a window. As the previous lessons established, image quality dominates the false negative rate, and poor photography can induce an apparent demographic effect that belongs to the camera rather than to the applicant.

So the first rule is: reject the capture, not the person. Assess quality on the device before submission, and if it is inadequate, say so and ask again — with a specific instruction. "Your face is too dark, move towards a light" is actionable. "Verification failed" is a dead end that the applicant reads as a judgement about them.

The second is: count attempts and act on them. A person on their fourth attempt is about to leave. That is the moment to offer the alternative route from lesson four, not after they have exhausted a fixed limit and been logged as a failure. An escalation triggered by attempt count is cheap to build and recovers exactly the applicants a strict threshold loses.

The third is accessibility, which has standards and is usually treated as a separate project. Three success criteria from the web content accessibility guidelines bear directly on an onboarding journey:

  • Redundant entry. "Information previously entered by or provided to the user that is required to be entered again in the same process is either: auto-populated, or available for the user to select." An onboarding flow that asks for the same address twice fails this, and every re-entry is an opportunity to abandon.
  • Target size. "The size of the target for pointer inputs is at least 24 by 24 CSS pixels, except when:" — a criterion that decides whether a person with a tremor or an older phone can complete a form at all.
  • Accessible authentication, which matters at the end of onboarding when the customer sets up their credentials, and which the authentication course covers.

And the fourth is the one nobody budgets for: the wait. A journey that pauses for a manual review, a data check or a document upload loses people during the pause unless it tells them what is happening and how long it will take. A status message is not a nicety; it is the difference between a customer waiting and a customer starting an application with a competitor.

Why this lesson is placed here. Everything in it is within the institution's own control, requires no change to risk appetite, no regulatory conversation and no threshold adjustment — and in most funnels it recovers more legitimate applicants than any change to the controls would.

Lesson 9 of 12

De-risking, as a supervisor now defines it

The institutional version of a false reject is not a threshold. It is a decision to stop serving a category of customer, and it has acquired a name and a definition.

The definition, from a supervisor. The European Banking Authority's guidelines define it as "a refusal to enter into or a decision to terminate business relationships with individual customers or categories of customers associated with higher ML/TF risk, or to refuse to carry out higher ML/TF risk transactions."

Note that the definition covers both individuals and categories, and both refusal and termination. It is deliberately wide, because the behaviour it describes takes several forms that institutions have historically described differently to themselves.

And the requirement. Institutions must ensure that the risk-based approach "does not result in the blanket refusal or termination of business relationships with entire categories of customers that they have assessed as presenting higher ML/TF risk."

Read that against how risk appetite statements are usually written. "We do not bank money service businesses" is a blanket position on a category. Under this guideline, a higher risk assessment for a category is the start of a differentiated response, not the justification for an exit.

The international standards say the same, and have for a decade. The FATF's guidance on the risk-based approach and de-risking addresses exactly this — that a risk-based approach means assessing and managing risk case by case, and that wholesale withdrawal from categories of customer is not what the standards require. That position is cited in several other courses in this catalogue, and the guidance is listed in the sources here.

A caution on sourcing, stated plainly: the FATF's website refuses automated access, so this course states the FATF's position in its own words from the catalogue's existing verified material rather than quoting text it could not retrieve. Where you need the exact wording — and for a policy document you will — go to the guidance itself.

What the newer European regulation adds. The anti-money laundering regulation, which applies from July 2027, is explicit that simplified due diligence is not an absence of due diligence: "That does not equate to an exemption or absence of customer due diligence measures. It rather consists of a simplified or reduced set of scrutiny measures, which should however address all components of the standard due diligence procedure."

All components, in reduced form. That is a materially different instruction from the common implementation, which drops components entirely.

And it names inclusion as a reason a product can be lower risk. Among the factors indicating potentially lower risk, the regulation lists "financial products or services that provide appropriately defined and limited services to certain types of customers, so as to increase access for financial inclusion purposes".

An inclusion product, with defined and limited services, is a recognised lower-risk category. That is the legal foundation for the tiering in lesson six, and it is worth quoting into an internal risk assessment, because it converts "we built a basic product to widen access" from a commercial argument into a risk-based one.

Lesson 10 of 12

Refusing somebody by machine

An onboarding decision taken by a system, with no person involved, is a regulated act in most places — and an automated refusal at onboarding is the clearest example of one.

The general right. The GDPR: "The data subject shall have the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her."

There is an exception for a decision "necessary for entering into, or performance of, a contract between the data subject and a data controller" — which an onboarding decision often is, so the exception frequently applies. But it is not a release. Where it applies, the controller must still "implement suitable measures to safeguard the data subject's rights and freedoms and legitimate interests, at least the right to obtain human intervention on the part of the controller, to express his or her point of view and to contest the decision."

Human intervention, a point of view, and a contest. Those three are the floor, and they describe exactly the manual route in lesson four — which means that route is not only good practice; in this situation it is the safeguard the law requires in exchange for automating the decision.

The sector-specific version. The anti-money laundering regulation addresses automated onboarding decisions directly. Any decision to enter or refuse to enter into or maintain a business relationship with a customer must be "subject to meaningful human intervention to ensure the accuracy and appropriateness of such a decision", and the customer "may obtain an explanation of the decision reached by the obliged entity, and may challenge that decision".

Meaningful human intervention. Not a review screen that an operator clicks through at four hundred cases a day — the adjective is doing work, and a process that cannot show what the human actually considered will struggle to defend it.

And one national law states both the right and the remedy. The Data Protection Act: "Every data subject has a right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning or significantly affects the data subject."

Where such a decision is taken, the controller or processor must "as soon as reasonably practicable, notify the data subject in writing that a decision has been taken based solely on automated processing", and the remedy available includes to "take a new decision that is not based solely on automated processing."

A new decision, not a review of the old one. The accompanying regulations add the obligations to "provide meaningful information about the logic involved", to "process personal data in a way that eliminates discriminatory effects and bias", and to "ensure that a data subject can obtain human intervention and express their point of view."

The operational conclusion, which ties the course together. An institution that automates onboarding refusals owes the refused applicant notification, an explanation, and a route to a decision made by a person. Most institutions provide none of the three — and the same process that would satisfy the law is the one that would recover the legitimate customers lesson three showed you are losing.

Lesson 11 of 12

What exclusion does downstream

Refusing a customer does not end the transaction. It moves it somewhere you cannot see, and the effect is measurable at the level of whole corridors.

The historical evidence. A World Bank survey conducted in the first half of the last decade recorded which customer types were most affected by banks withdrawing from correspondent relationships: "over 69 percent of the banking authorities responding to the survey indicated that money transfer operators and other remittance companies are most impacted, followed by small and medium domestic banks (44 percent) and small and medium exporters (26 percent)."

And from the money transfer operators' own side: "A significant portion of MTOs declared that the MTO principal (28 percent of the respondents) or its agents (45 percent of respondents) can no longer access banking services", with a further quarter of principals "currently unable to operate regularly through bank channels."

Those figures describe 2010 to 2014 and are cited here as history rather than as a current measurement. The reason they still matter is that they are the documented instance of the mechanism: controls applied categorically did not reduce the flows, they relocated them — and the correspondent banking course in this catalogue covers the institutional version in full.

Why over-cautious rules are chosen. A joint report by the central bank community and the World Bank identified the incentive honestly: "Many country regulatory authorities have therefore opted to adopt rules-based, somewhat restrictive AML/CFT regimes, rather than risk being found non-compliant with the FATF Recommendations."

That is a structural explanation rather than a criticism. A supervisor judged on compliance, by an assessment process, will prefer a rule that is easy to demonstrate over a judgement that is hard to defend. The same incentive operates inside an institution: a front-line officer judged on errors will refuse.

And the specific requirement that excludes most people is often not required at all. The same report notes that strict customer due diligence such as "requiring a formal address, which is not part of the FATF requirements" "may exclude them from the formal financial system".

Not part of the requirements. A proof-of-address rule that excludes people in informal housing is, in many cases, an institutional choice that has been mistaken for a standard — and that is worth checking against your own obligations, because it is the single most commonly cited barrier in the whole field.

What the evidence says tiering achieves. The successor report records an outcome rather than an intention: "Where implemented, tiered KYC regimes have produced positive outcomes on the target segments".

And one exclusion mechanism that no policy change fixes. The same report notes that "Certain populations, such as manual labourers with worn fingerprints, the elderly or individuals with disabilities, may have difficulty enrolling in or using ID systems that rely on certain types of biometrics", "which can lead to exclusion if no alternative options are in place."

If no alternative options are in place. Which returns the whole course to lesson four: the alternative route is not a concession. It is the control that makes the rest of the system defensible.

Lesson 12 of 12

The business case, and what to do on Monday

This work competes for resources against revenue projects, and it wins that argument on its own terms rather than on fairness alone.

The commercial argument, stated plainly. Every wrongly refused or exhausted applicant is an acquisition cost already spent and no revenue. The marketing that brought them, the capacity that served them and the vendor fee for the check that failed them are all sunk. A funnel losing a few per cent of legitimate applicants at each of five stages is losing a substantial share of the customers it paid to attract, and nobody has a number for it because, as lesson one established, the people who left were never counted.

So the first deliverable is not a policy change. It is a number.

What to measure first, in order of how quickly it can be produced:

  • Abandonment and fail rates, by stage, using the definitions in lesson two. Most institutions can produce this from existing logs within a week, and the result is usually a surprise to somebody senior.
  • Attempts before abandonment. How many people left on their third or fourth try at one step. That names the step to fix.
  • The pass rate of the manual route, if one exists. Lesson four explains how to read it.
  • The spread across cohorts, where lawful. The aggregate will look acceptable; the spread is the finding.
  • Reasons for refusal, grouped. If the largest category is "document capture failed", you have a product problem being recorded as a risk decision.

Then the order of work, cheapest and least contentious first:

  • Fix capture, per lesson eight. No risk appetite conversation, no regulatory engagement, and it usually recovers the most people.
  • Add the escalation on attempt count, routing people to a human before they give up rather than after they are logged as failures.
  • Write down the acceptable alternative documents, as the supervisor in lesson seven requires — specified, not left to an officer's discretion.
  • Make the manual route discoverable and staffed, which is the control that also satisfies the automated-decision safeguards in lesson ten.
  • Then look at thresholds, with the demographic differential bound from lesson three in hand, and at tiering, with the regulator-set examples from lesson six.

What to stop doing. Reporting the funnel as a conversion metric owned by product and a refusal metric owned by compliance, with nobody holding the combined number. That split is the reason this problem persists in institutions that are competent at both halves.

And the sentence to end on. Of the people your process refuses or exhausts, some should have passed. You do not currently know how many, and finding out costs less than the customers it will recover — which is why this is one of the few places in financial crime compliance where doing the right thing and doing the profitable thing are the same piece of work.

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: India.

Sources

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

  1. The Global Findex Database 2025: Connectivity and Financial Inclusion in the Digital Economy · World Bank
  2. SP 800-63-4: Digital Identity Guidelines, including the performance metrics in section 3.5.2 (July 2025) · US National Institute of Standards and Technology
  3. SP 800-63A-4: Identity Proofing and Enrollment, including the biometric performance thresholds and trusted referees · US National Institute of Standards and Technology
  4. NIST IR 8280: Face Recognition Vendor Test Part 3, Demographic Effects (December 2019) · US National Institute of Standards and Technology
  5. Face Recognition Technology Evaluation: Demographic Effects in Face Recognition · US National Institute of Standards and Technology
  6. EBA/GL/2023/04: Guidelines on policies and controls for the effective management of ML/TF risks when providing access to financial services (March 2023) · European Banking Authority
  7. Regulation (EU) 2024/1624 (the Anti-Money Laundering Regulation), Article 76 and Annex II · Official Journal of the European Union, via EUR-Lex
  8. Directive 2014/92/EU (Payment Accounts Directive), Article 16 · Official Journal of the European Union, via EUR-Lex
  9. Regulation (EU) 2016/679 (GDPR), Article 22 · Official Journal of the European Union, via EUR-Lex
  10. UK Payment Accounts Access and Closures: Update (September 2024) · UK Financial Conduct Authority
  11. FATF Guidance on the Risk-Based Approach and De-risking · Financial Action Task Force
  12. Guidance on financial inclusion and anti-money laundering and terrorist financing measures · Financial Action Task Force
  13. Payment aspects of financial inclusion (April 2016) · Committee on Payments and Market Infrastructures and the World Bank Group
  14. Payment aspects of financial inclusion in the fintech era (April 2020) · Committee on Payments and Market Infrastructures and the World Bank Group
  15. Withdrawal from Correspondent Banking, and the G20 Survey on De-Risking in the Remittance Market (November 2015, historical) · World Bank Group
  16. Reserve Bank of India (Commercial Banks - Know Your Customer) Directions, 2025 · Reserve Bank of India
  17. Regulatory Framework for Mobile Money Services in Nigeria (July 2021) · Central Bank of Nigeria
  18. National Payment System (AML Guidelines for the Provision of Mobile Payment Services) Guidelines, 2013 · Central Bank of Kenya
  19. Web Content Accessibility Guidelines (WCAG) 2.2, success criteria 2.5.8, 3.3.7 and 3.3.8 · World Wide Web Consortium
  20. Data Protection Act, 2019 (No. 24 of 2019), section 35 · Office of the Data Protection Commissioner, Kenya
  21. Data Protection (General) Regulations, 2021 (Legal Notice No. 263), regulation 22 · Office of the Data Protection Commissioner, Kenya