The Problem Is Fast Onboarding Without Proof
Kyndryl and Google Cloud announced a proof of concept with Incore Bank that used Gemini-powered agents to support Know Your Customer onboarding and risk assessment. The appealing promise is speed, but regulated onboarding does not become better merely because documents move faster.
KYC work has to preserve the reason for a decision. If an AI system extracts facts, checks sources, assigns risk, or recommends approval, the bank needs a durable trail showing what evidence was used and where human judgment entered.
Why KYC Agents Create Evidence Gaps
KYC agents may read structured forms, passports, corporate documents, sanctions data, beneficial ownership records, internal history, and external sources. They may also reconcile conflicts and produce a risk recommendation in a way that looks complete but hides intermediate uncertainty.
The gap appears when the compliance reviewer can see the answer but not the path. In a regulated workflow, explainability is not a presentation feature; it is how the organization proves that policy, evidence, and judgment were applied consistently.
The Cost Of Missing Audit Evidence
Missing evidence creates rework, delayed reviews, weak regulator responses, inconsistent customer treatment, and brittle model-risk governance. It can also erase the time savings that justified the automation because compliance staff must reconstruct what the system should have preserved.
A practical cost measure is cases processed times missing-evidence rate times reconstruction minutes. If 400 monthly cases have a 12 percent evidence gap and each gap takes 35 minutes to reconstruct, the process loses 28 hours before counting supervisory review.
How To Diagnose The Evidence Trail
Take ten completed AI-assisted onboarding cases and ask whether a reviewer can trace each risk factor back to a specific source, timestamp, rule, extraction confidence, and human decision. Include cases with clean approval, manual override, missing documents, and higher-risk customers.
The strongest diagnostic question is simple: could an independent reviewer reproduce the decision without rerunning the model? If the answer is no, the workflow is still depending on the AI session rather than a usable compliance record.
Options For Regulated AI Onboarding
The cautious option is to let AI summarize documents while humans perform all risk classification. That may be suitable when source quality is poor, rules are unsettled, or the institution lacks model validation capacity.
A more advanced option lets agents extract, compare, score, and prepare decision records under policy-as-code guardrails, human review, and continuous monitoring. The higher the automation level, the stronger the evidence trail must become.
Build The Audit Evidence Trail
The trail should include case identifier, document list, extraction fields, source references, validation checks, risk factors, policy rules triggered, model version, prompt or workflow version, reviewer decision, overrides, and monitoring triggers. Keep the record readable to compliance staff, not just engineers.
Separate facts from recommendations. A verified legal name, an unresolved ownership conflict, a risk-score explanation, and a human approval are different evidence types and should not collapse into one generated summary.
A Worked Example
A B2B bank onboards a corporate customer with three ownership layers and cross-border directors. The AI system extracts beneficial ownership data, flags one document mismatch, checks trusted external sources, and proposes enhanced review instead of straight approval.
The reviewer sees every source link, the mismatched field, the policy rule, the model confidence, and the final human note. The case may still take judgment, but the judgment is now attached to evidence rather than buried in email and screenshots.
Measures That Prove Control
Track onboarding cycle time, evidence completeness, extraction error rate, reviewer override rate, high-risk escalation accuracy, missing-document reopen rate, and monitoring-trigger follow-up. These measures should be split by customer type and risk tier.
Also measure audit response time. If a compliance manager can answer why a decision was made in minutes instead of hours, the AI system is improving the control environment instead of merely accelerating intake.
The Next Step This Week
Before expanding KYC AI, define the minimum evidence bundle for one customer type. Name the fields that must be sourced, the rules that must be visible, the human decision points, and the documents that must be retained.
Then run one historical case through the bundle and identify what the current process cannot prove. SynHy would treat that gap list as the build specification for the next controlled automation step.
Sources And Methodology
This article was triggered by Kyndryl's announcement that it worked with Incore Bank and Google Cloud on agentic AI customer onboarding and by Finextra's coverage of the proof of concept. Reported performance claims are treated as project-specific claims, not general benchmarks.
The audit-trail framework is SynHy original analysis informed by FinCEN's customer due diligence rule background, FATF's digital identity guidance, and the Financial Stability Board's AI financial stability report. Institutions should adapt it to their regulators, counsel, and model-risk requirements.