Electronic KYC and Digital Onboarding: The Path and the One Step

Digital onboarding is the whole path from a person starting an application to a usable relationship. Electronic identity verification is one step inside that path and not another name for it. The step completion rates along the path multiply rather than average. Three steps each above 93 per cent therefore produce a whole of 86.0 per cent, and the path loses 1,400 of every 10,000 who start.

A path made of steps carries one arithmetic property that decides everything else about it: the whole is the product of the parts and never their average. Reported step by step, the month shows three healthy numbers in the nineties. Multiplied together, those numbers show how many people actually arrived. The two readings of the same month differ by more than nine percentage points, and the difference is made of people standing outside a door somebody believes is open.

What is digital onboarding, and where does electronic KYC sit inside it?

Digital Onboarding, the whole path from a first tap to a usable relationship

Digital onboardingThe whole path from a person starting an application to a usable relationship, made of several steps in a fixed order. is a path, not a check. The path starts when somebody taps a button on a handset and it finishes when that person has a working relationship with the institution: an account they can use, a loan decision they can act on, a mandate they can set up. Everything in between is a step, and the steps run in a fixed order because each one needs what the step before it produced.

Getting a new electricity connection in a new flat has the same shape, though nobody calls any part of it onboarding. A household fills in a form. Somebody checks that the name on the form matches the name on the ownership paper. Somebody takes a meter reading. Somebody records what the household agreed to pay for. And only after all of those does the supply come on. The electricity is not on when the name check passes; it is on when the last step finishes, and every step in between is a place the process can stop without anybody at the supply office noticing that a household is still sitting in the dark.

Sumeru Bank Limited, an invented lender, runs its retail loan front door in six numbered steps. The steps carry numbers rather than names: a number points at exactly one step and a name drifts between two.

Steps 2 and 3 together are the part everybody in the building has a name for. Steps 1, 4, 5 and 6 are the part nobody has a name for, and they carry two thirds of the path.

Why does calling the whole path by the name of one step cost a firm something?

Digital Onboarding vs eKYC, the two things that get merged and what the merge hides

Electronic identity verificationOne step inside the path, in which a claimed identity is checked against a record held somewhere else. is step 2 of the six, and almost everyone in the room calls it electronic know your customer (KYC). The step takes a claimed identity and puts it against a record held elsewhere. One bounded question with one bounded answer is genuinely the hardest step of the six to build. Being the hardest to build is exactly why it ends up lending its name to the whole path.

Here is the trap, and it is not a vocabulary quibble. A project that says it is building electronic KYC will scope, staff, test and report on step 2. A project that says it is building digital onboarding has to scope, staff, test and report on six steps, of which step 2 is one. The two projects sound identical in a status meeting and they produce completely different things: one produces a check that works, and the other produces a path a person can get to the end of. When the merge happens, the other five steps still get built. A path has to function at all. But nobody is accountable for how many people finish them.

Sumeru had exactly that merge. The monthly pack carried a detailed line for the identity match and a detailed line for the liveness check. The pack carried one line for everything else, and that line was a single number.

What are the six steps of one bank's path?

The table above lists them, and the useful way to read the list is by what each step is asking the applicant to hand over. Step 1 asks for a number they know by heart. Step 2 asks for an identity number and nothing else. Step 3 asks them to take a photograph of themselves. Step 4 asks them to find, photograph and upload documents. Step 5 asks them to read something and agree to it. Step 6 asks nothing at all. The system does the submitting.

Notice that the effort the applicant has to supply is not spread evenly across the six steps, and neither are the losses. Steps 1, 2 and 6 ask for almost nothing. Steps 3, 4 and 5 ask for a photograph, a document hunt and a piece of reading, and those three steps are where every single person the path lost in the month was lost. The pattern is not a coincidence and it is not a fact about applicants. The pattern is a fact about what each step demands.

Where did the month's applicants actually stop?

In month 6 at Sumeru Bank Limited, 10,000 applications entered step 1. A funnelThe count of people remaining at each step of a path, which is the only honest way to read completion. is the count of people still on the path at each step, and it is the only reading that shows where the losses fell rather than how many there were in total.

The red blocks from top to bottom are the month. Step 3 took 620, step 4 took 480, step 5 took 300, and 620 plus 480 plus 300 is 1,400. Ten thousand less those 1,400 is 8,600, the number that reaches the decision engine and the number the rest of the bank talks about. The 8,600 is a real population and so is the 1,400, and only one of them appears anywhere in the monthly pack.

Why does a step that loses nobody still send files to a person?

Step 2 is the strange one, and it is worth slowing down for. The identity match at Sumeru Bank Limited is a written rule of 9 lines that compares three fields across two sources, the application form and the identity record, at three lines a field: an exact match, a normalised match and a tolerance. In month 6 it found 189 files where the two sources did not agree. Not one of those 189 applications ended.

The 189 applications survived because a mismatch at step 2 is a routed fileA file sent to a person to resolve, which is work for the institution rather than a lost applicant. rather than a refusal. The file goes to a person at the exception desk, who looks at the two records and decides. The applicant carries on. The 189 are exception cause 5, and they sit inside the 3,010 files a person handled that month rather than inside the 1,400 the path lost.

The reason for the mismatches is ordinary rather than sinister, and the split says so plainly. Name accounts for 121 of the 189, being 64.0 per cent. Date of birth accounts for 44, being 23.3 per cent. Address accounts for 24, being 12.7 per cent. Names dominate for three ordinary reasons. Initials get expanded in one source and not the other. The given name and the surname arrive in a different order. And a transliteration made twice by two different people comes out twice in two different ways. A step can generate a great deal of work without costing a single applicant their application, and a report that counts only losses will show that step as costing nothing.

Why do step completion rates multiply rather than average?

Three steps in the month lost people. A step completion rateThe share of people entering one step who finish it, which is meaningful only beside the rates of the steps around it. is the share of people entering a step who finish it, so each one is measured against a different base. Step 3 took 9,380 out of 10,000, a rate of 93.8 per cent. Step 4 took 8,900 out of the 9,380 who reached it, a rate of 94.9 per cent. Step 5 took 8,600 out of the 8,900 who reached it, a rate of 96.6 per cent.

Now the arithmetic that decides everything. To reach the decision engine an applicant has to survive all three, and surviving all three is a chain: 0.938 times 0.949 times 0.966 is 0.860. Three step rates every one of which is above 93 per cent produce a whole of 86.0 per cent. The losses compoundThe property that a whole path is the product of its step rates rather than their average, so small losses in a row stack up. rather than sitting side by side.

The household version is a wedding invitation list. A family invites 200 people. Ninety per cent say yes, ninety per cent of those actually book a train, and ninety per cent of those actually board it. Every single number is ninety, and 146 people come to the wedding. Nobody in that chain did anything unreasonable, and the catering was for 180.

Why is the average of the three rates not a number that means anything?

The average of 93.8, 94.9 and 96.6 is 95.1 per cent. The average is arithmetically correct and it describes nothing that happened to anybody in month 6. Averaging step rates treats the three steps as alternatives, as though each applicant met one of them, when in fact every applicant met all three in a row. The only two honest summaries of a path are the product of the step rates and the count of people still standing at the end, and 95.1 per cent is neither.

The size of the error is the argument, so it is worth being precise about it. The average reads 95.1 and the whole path reads 86.0, a gap of 9.1 percentage points. On 10,000 applications that gap is 910 people. Somebody reading the average would budget, staff and plan for a front door losing about 500 people a month. The front door lost 1,400.

Which route an applicant took at step 4 decided whether they finished, so why did nobody see it?

The funnel above says 480 people were lost at document upload. The number 480 was the whole of what the monthly pack said about step 4, and it turns out to have been hiding the sharpest finding in the month.

Step 4 is not one step for everybody. Of the 9,380 applicants who reached it, 5,498 were on a consent driven rail, where the customer instructs a holder of their records to send named items to the bank. Another 2,996 were on the upload route, where the customer downloads a file themselves and uploads a picture of it. The remaining 886 were on a direct arrangement between two institutions under contract. 5,498 plus 2,996 plus 886 is 9,380.

Now split the 480. On the rail, 48 of 5,498 abandoned, being 0.87 per cent. On the upload route, 432 of 2,996 abandoned, being 14.4 per cent. On the direct arrangement, nobody abandoned. 48 plus 432 is 480. The upload route carried 29.0 per cent of the month's completed files and produced 90.0 per cent of the abandonment at step 4, an abandonment rate about sixteen times the rail's, and nobody at Sumeru Bank Limited saw it because abandonment was reported as one number.

Follow all three routes to the end of the path and they close back exactly. On the rail, 5,498 entered step 4, 48 left there and 204 more left at the consent step, leaving 5,246. On the upload route, 2,996 entered, 432 left at step 4 and 70 more left at consent, leaving 2,494. On the direct arrangement, 886 entered, none left at step 4 and 26 left at consent, leaving 860. The three closing counts are the recorded mix of completed files, and 5,246 plus 2,494 plus 860 is 8,600. The consent step losses of 204, 70 and 26 are arithmetic on those recorded counts rather than a separately measured split, and 204 plus 70 plus 26 is 300.

Merging two populations is the easiest mistake to make at this point, so the two have to be held apart. The 9,380 is the count of applicants who reached step 4 and the 8,600 is the count of files that completed the path, and no route share may be read against the wrong one. On the 9,380 the rail carries 58.6 per cent; on the 8,600 it carries 61.0 per cent. The two shares are statements about two different sets of people.

Who are the 1,400, and what happened to each group?

The 1,400 are not a rounding error and they are not one group. The 1,400 are three groups with three separate causes, and the reason to draw them apart is that a total cannot be fixed while a split can. 620 were refused by the liveness check at step 3. 480 abandoned at document upload. 300 could not complete the consent step. Each of those three numbers points at a different repair, and the single number 1,400 points at nothing.

The dignity of this matters and it is not decoration. Every one of the 1,400 is a person who wanted the loan enough to start an application on their own handset and to get partway through. Inside the 620 refused at step 3, a review by hand of 200 refusals found 31 genuine applicants, being 15.5 per cent, extrapolating to about 96 across the month, and all 31 shared one condition: a low light image on a low specification handset. Inside the 432 who left the upload route, the thing they were asked to do was find a document, photograph it and upload it on a connection they were paying for. None of that is a failure of effort.

What did the digital route remove, and what did it move somewhere nobody measures?

Before the digital path existed, onboarding at Sumeru Bank Limited happened in a branch. The branch route took 1 working day of elapsed timeHow long a customer waits from start to finish, as distinct from how many minutes of staff time the same file consumes. and an assumed 38 minutes of staff time a file. The month's 8,600 completed files at 38 minutes is 326,800 staff minutes, being 38.9 posts at an assumed working month of 8,400 minutes a person. The digital route removed every one of those 326,800 minutes. Put into money at the bank's assumed fully loaded cost of Rs 9,00,000/- a post a year, those 38.9 posts are about Rs 3,50,00,000/- a year, and that figure is arithmetic on two assumed numbers rather than a measured saving. The saving sits against a chain the same bank assumes costs Rs 65,00,000/- a year to run.

The saving is real. Beside it sits a second effect. The same change moved 1,400 applicants a month into a state where nobody at the bank spends a single minute on them, and no report the process produces has a line for those applicants. Minutes removed is a figure a system produces by itself, and applicants who never arrive is a figure somebody has to decide to go looking for, so the bank's own numbers record the 38.9 posts and not the 1,400. An uncounted applicantSomebody who left the path, whose absence appears in no report the process produces because nothing in it counts an absence. is not hidden by anybody. There is simply no field for them.

Where did the cost of the failures go, and who carries it now?

The comparison that closes the subject is arithmetic rather than an accusation. Restoring a route for the roughly 96 genuine applicants the liveness check refused costs an assumed 12 minutes a case at the exception desk. About 96 people a month at 12 minutes is 1,152 staff minutes a month, or 0.14 of a post at the assumed working month of 8,400 minutes. Set that against the 326,800 staff minutes the digital route removed and it is 0.35 per cent of them.

The reason the second route was not built is not that it was expensive; it is that the two numbers had never appeared side by side. One of them was in a business case, in bold, at the top. The other did not exist until somebody read 200 refusals by hand. A cost that nobody has calculated is not a small cost or a large one. A cost nobody has calculated is not in the conversation at all.

How should a completion rate be read, and what has to be printed beside it?

A completion rate is a statement about the people who finished and it is silent about everybody else. 86.0 per cent is true, useful and radically incomplete. Four things have to sit beside it before anybody can act on it, and none of the four costs anything to produce once the funnel exists.

Two firms can report 86.0 per cent completion and be in completely different positions. One of them can name the 620, the 480 and the 300, can say which route the 432 were on, and can put a cost against a repair for each. The other has a rate. The same headline covers a firm with a list of fixes and a firm with a number, and only the first one has anything it can do on Monday.