The short version

A video KYC system is designed, specified and audited around one adversary: somebody pretending to be someone else. That is the right thing to design against and it is not what makes most sessions fail. Sessions fail because the upstream connection cannot carry evidential-quality video, because a laminated identity card reflects the ceiling light straight into the camera, because the room is dim, and because the customer did not have their documents on the table when the agent appeared. Four of the five stages where people drop out are physical problems. The controls are aimed at the fifth.

Teams building V-CIP flows spend most of their design effort on the part a regulator will examine, which is correct and also incomplete. The regulatory requirements decide whether the system is permissible. They have very little to do with whether a given customer gets to the end of it.

We have written separately about what the RBI's rules actually require you to build. This is about the other half: the funnel, and where real people fall out of it.

The adversary you designed for is rare

Liveness detection, randomised prompts, face matching against the document, geotagging, IP checks — these exist to defeat a deliberate impostor. They are necessary and they work.

But consider the population passing through the flow. The overwhelming majority are ordinary customers trying to open an account, on their own phone, in their own home, holding their own documents. When one of them fails, almost never is it because they were pretending to be someone else. It is because something physical went wrong, and the system has no way of distinguishing "this image is unusable" from "this image is suspicious" except by refusing both.

That is the same observation we keep arriving at in computer vision work: four of the five variables are physical. Video KYC is a vision system with a compliance obligation stapled to it, and it inherits every physical constraint vision systems have.

WHERE SESSIONS ACTUALLY END STAGE WHAT GOES WRONG CAUSE BEFORE THE CALL the queue, the moment documents not to hand called at a bad time physical THE CONNECTION evidential quality required upstream bandwidth mobile network handover physical DOCUMENT CAPTURE the biggest single loss glare off the laminate autofocus hunting, shaky hands physical FACE MATCH, LIVENESS compared against what, and how old dim rooms, backlight glasses, headwear, old photo physical THE AGENT'S DECISION where the tension is resolved threshold set tight to avoid a false accept policy
The controls were designed for the bottom row; the volume leaves from the four above it.

Bandwidth, in the direction nobody tests

An ordinary video call degrades gracefully. Resolution drops, the picture softens, people carry on talking. A KYC session cannot do that, because a soft picture is useless for a face match and the recording has to stand up to later inspection.

Two things make this harder than it looks. The constraint is on the upstream connection, which is the weaker direction on virtually every consumer line and the one nobody checks. And a customer walking around their home during the session will hand off between access points or cells, which is invisible on a phone call and fatal to a continuous recording.

The practical mitigations are unglamorous and effective: test the upstream before starting rather than after, tell the customer to stay put in plain language, and detect degradation early enough to pause rather than letting the session fail at minute four.

Glare is the single biggest loss

If you fix one thing, fix this. Identity documents are laminated. Laminate is reflective. A person holding a card up to a phone, indoors, under a ceiling light, will produce a bright band across the card — and because they are holding it to face the camera, that band falls almost exactly where the printed details are.

Underneath it sit two more problems. Phone cameras hunt for focus at close range, so the moment the card is steady enough to read is often the moment the lens decides to re-focus. And holding a card at arm's length while watching a screen is genuinely awkward, so the image moves.

What helps: capture video rather than relying on one still, and select the sharpest frame automatically; give live feedback while the document is in frame rather than a verdict afterwards; and tell people to tilt the card slightly away from the light, which is a one-line instruction that measurably converts. What does not help is a generic "image unclear, please try again", which tells a customer nothing about what to change.

The threshold is a commercial decision wearing a technical costume

Here is where the physical problems become a policy problem.

A face match produces a score, and somebody chooses where to cut. Set it tight and impersonation gets harder — and more honest customers in dim rooms, in glasses, with a ten-year-old passport photograph, get refused. Set it loose and the reverse.

The two errors are not symmetric and they are not equally visible. A false accept is a regulatory event with a named owner and a file. A false reject is a person who gives up, is never counted, and appears in no report at all. So the incentive runs one way, and the cost of running it that way is paid silently by the acquisition budget.

Worth stating plainly in design: that threshold is a commercial decision. It should be set by people who can see both costs, reviewed against actual outcomes, and not left as a default that arrived with a vendor SDK.

You get one good attempt

Product plans routinely assume a customer who fails will simply try again. Most will not, and the ones who do are rarely in a better mood or a better-lit room.

A retry means queueing again, repeating something that has already made them feel foolish, and trusting a process that has already let them down once. In a regulated flow it also costs you another agent slot, so a low first-pass rate is expensive on both sides of the transaction.

Designing on the assumption of one good attempt changes the priorities: front-load everything that can be checked before the agent joins — connection quality, document present, lighting adequate, face detectable — so that the session itself starts only when it can plausibly succeed. Rejecting someone at second zero with a clear reason is a far better experience than failing them at minute five.

Instrument the stage, not the outcome

Almost every V-CIP deployment reports a completion rate. Very few can say where the incompletions happened, and that single number cannot be acted on, because the five stages have five different remedies.

Record which stage each session ended at. Then the data tells you what to do: a cluster at the connection stage is a pre-flight check, a cluster at document capture is a guidance and frame-selection problem, a cluster at face match is a threshold conversation. Without it, teams tend to improve whichever part they find most interesting, which is usually the model.

Questions we get asked

Why do video KYC sessions fail so often?

Mostly for physical reasons rather than fraudulent ones. Upstream bandwidth on a mobile connection, glare from the laminate on an identity card, poor indoor lighting, a camera that will not hold focus at close range, and a customer who does not have their documents to hand. The controls are built around detecting impersonation, and impersonation is a small fraction of what actually stops a session completing.

Why does bandwidth matter more for video KYC than for a video call?

Because an ordinary call degrades gracefully and a KYC session cannot. A blurry face is fine for conversation and useless for a face match, and the recording has to be of evidential quality for later audit. The constraint is also upstream rather than downstream, which is the direction consumer connections are weakest and the direction speed tests are least likely to be checked in.

What causes document capture to fail?

Glare is the dominant cause. Identity cards are laminated, laminate reflects, and a customer holding one up to a phone under a ceiling light will produce a bright band across exactly the area that matters. Add autofocus hunting at close range and hands that are not steady, and the result is a document the agent cannot verify and the customer does not understand why.

Does a stricter liveness threshold make onboarding safer?

It makes impersonation harder and it makes honest customers fail more often, and the second effect is much larger in volume. A false accept is a regulatory event; a false reject is a lost customer who rarely returns. Both are real costs, and a threshold set only against the first will quietly spend a great deal of the second without anyone attributing it.

How many times will a customer retry a failed video KYC?

Far fewer than product teams assume, and the second attempt is where most of the remaining drop-off happens. A retry costs the customer another wait in a queue and a repeat of something that has already embarrassed them. Design on the basis that you have one good attempt and a weak second, and spend the effort on making the first succeed.

What should be instrumented in a video KYC flow?

The stage at which each session ended, not just whether it completed. A single completion rate cannot tell you whether to fix the network guidance, the document capture or the agent decision threshold, and those have entirely different remedies. Per-stage drop-off is the one piece of instrumentation that makes the rest of the work targetable.

Related service Video KYC development V-CIP systems built to RBI's requirements

If you have a V-CIP flow live and a completion rate you are not happy with, per-stage instrumentation is the first thing worth adding — it usually takes days and it tells you which of five quite different problems you actually have. It is where we start on video KYC engagements.