Representative engagement
The system assembles periodic KYC review files from core banking data and client documents. It drafts the narrative with cited evidence and flags what is missing. Relationship managers review and sign.

- Client
- A Swiss private bank (client withheld under NDA)
- Sector
- Private Banking & Wealth
- Duration
- 2-week assessment, 10 weeks to production
- Services
- AI Strategy · Agentic Systems · AI Operations
01Challenge
Periodic KYC reviews for high-risk relationships took relationship managers and compliance officers several hours each. Information was scattered across the core banking system, a document management system and years of correspondence, much of it in French and German.
A backlog had formed. Reviews were completed close to their deadlines and quality varied between bankers. The bank wanted better files in less time. This had to work under FINMA supervision, with all client data kept in Switzerland.
02Approach
We started with a two-week assessment with compliance, front office and IT to agree the target file structure, the evidence standard for each section and the data residency architecture. The bank’s risk function reviewed the design before any build began.
The system was built on bank-controlled infrastructure in Switzerland. We used a model deployment that keeps data in-country, with no retention and no training. Each sentence in the drafted narrative cites the document or record it comes from. The reviewer sees the source side by side.
03The system
For each review due, agents gather client data and documents, check them against the bank’s KYC checklist, draft the source-of-wealth and activity narrative with citations, and list missing or expired evidence with a draft request to the client.
Components
- 01Connectors to core banking and document management, with inherited entitlements
- 02Checklist engine encoding the bank’s KYC policy
- 03Narrative drafting with sentence-level citations
- 04Missing-evidence detection and client request drafts in FR, DE and EN
- 05Reviewer workspace with side-by-side sources and sign-off
- 06Hosting on bank-controlled Swiss infrastructure, no data retention by model providers
Stack
- Swiss-hosted LLM deployment
- Bank on-premises Kubernetes
- Hybrid search
- Core banking API
- Document management connector
04Results
A Swiss private bank (client withheld under NDA)
- −55%
- Time per periodic review
- 4.1 → 4.6
- Average file quality score in compliance sampling (out of 5)
- 100%
- Narrative statements linked to source evidence
- 0
- Client data processed outside Switzerland
Representative engagement. Client withheld under NDA. Details are anonymised and combined so that no client can be identified; metrics are those measured on the engagement.
Related work
Next step
Tell us which process takes your team the most time. In a thirty-minute call, we can tell you if the sprint is a good start.


