Skip to content
Tarka
Hands holding a contactless card to a payment terminal

62%,

A GenAI analyst assistant built to pass model risk review

A GenAI analyst assistant for a digital bank.

The context

The bank's onboarding grew faster than its compliance team. Around 60% of new applications were flagged for manual KYC review, mostly for name mismatches, unclear documents or thin address histories. Analysts spent much of each review gathering evidence from five internal systems before making a decision.

Any model touching KYC decisions had to pass the bank's model risk review, which had already rejected one vendor tool for lack of explainability. The assistant could not decide anything on its own: every outcome needed a named analyst, a recorded rationale and evidence an auditor could follow.

A cash machine glowing in a dark doorway at night

What we did

  1. 1

    Evaluation set first

    We built an evaluation set of 3,000 past KYC cases with the analysts' final decisions and reasons. Every prompt, retrieval change and model version was scored against it before release, and the reports went to model risk unedited.

  2. 2

    Evidence retrieval

    The assistant pulls applicant records, document checks and sanctions screening results into a single case summary, citing the source of each fact. Low-risk mismatches, such as transliterated names, are resolved by deterministic rules rather than the language model.

  3. 3

    Human review in the loop

    Analysts see a recommendation, the evidence and a confidence level, and they approve, amend or reject it. Cases above the risk threshold always go to full manual review. Every action is logged with the analyst's name and the model version.

  4. 4

    Named owners and monitoring

    Each model and prompt set has a named owner in the bank. We set up drift and accuracy monitoring against a fresh weekly sample, with a documented rollback path that model risk signed off before go-live.

The result

62%.

Fewer manual KYC reviews

Cases needing a full manual KYC review fell 62% against the six months before launch, measured on the bank's own case management data. The assistant passed model risk review at the first submission, with evaluation reports, audit trails and named owners already in place.

Analysts now spend their time on the higher-risk cases that need judgement. Average time to clear a flagged application dropped from two days to under one, and the audit trail answers most compliance queries without anyone reopening the case.

past cases in the evaluation set
3,000
submission passed model risk review
1st
of decisions signed off by a named analyst
100%
“Their GenAI work cleared our model risk review first time: evaluation reports, audit trails, named owners. Our risk team asked for nothing extra.”
Portrait of Tom HendricksTom HendricksCTO, a digital bank

Have a number you need to move?

Tell us what it is. We will tell you what we would measure, and how long a discovery would take.

Talk to Our Team