How a Mid-Sized Malaysian Bank Cut KYC Time 60% with Digital Workers

An anonymised case study: a Klang Valley universal bank deploys a Digital Worker for KYC document review, cuts average processing time from 18 minutes to 7, and finds the harder problem was governance, not technology.

The client

A mid-sized Klang Valley universal bank — roughly 4,500 employees, retail and commercial banking, mature digital channels. Customer-onboarding KYC operations sat in a 40-person team processing ~9,000 new customer applications per month. Average end-to-end document review time per application: 18 minutes.

The client engaged us with a clear ask: assess whether AI could meaningfully shorten that 18-minute window without compromising regulatory adherence, false-rejection rate, or audit traceability.

The constraint set

• BNM regulatory expectations on KYC documentation evidencing — every decision must be auditable end-to-end.
• Customer photographs and identity documents are personal data under PDPA — cross-border inference was off the table.
• The existing team had to remain — workforce displacement was not acceptable to leadership or to the workforce union.
• Implementation budget: medium six figures (MYR). Anything more required board approval, which would have pushed the project past the regulatory cycle window.

What we built

A Digital Worker on the Teragrid Ai platform, deployed in the client's on-premise environment to satisfy the data-residency constraint. It performs three discrete tasks on each application: (1) extracts structured fields from uploaded identity documents (NRIC, passport, supporting documents), (2) cross-references those fields against the application form and flags discrepancies, (3) drafts the disposition recommendation — proceed, request clarification, or escalate to senior reviewer.

The Worker does not make the decision. A human reviewer reviews and confirms (or overrides) the draft recommendation. Every interaction is logged with full audit trail. Override rate is the headline metric we now track.

The 90-day numbers

• Average processing time: 18 minutes → 7 minutes per application (61% reduction).
• Throughput per FTE: 18 applications/day → 46 applications/day.
• Override rate: 11% in the first month, 6% by month three, 4% by month six (after model tuning on the disagreement cases).
• False rejection rate: unchanged from baseline. Confirmed via parallel manual sampling for the first 60 days.
• Customer onboarding NPS: +14 points (faster decisions, fewer "we need additional documents" cycles).

What surprised us

Two things, neither of them technical.

First: the existing team adopted the tool faster than leadership expected. The narrative going in was "AI is replacing your work." The reality was "AI is doing the part of your work that you hated, freeing you to do the part you valued." The team — most of whom had been hand-typing data from document scans for years — preferred reviewing draft recommendations over manually extracting fields.

Second: governance was the harder problem than technology. Building the Worker took 11 weeks. Getting the disposition framework approved by compliance, the audit trail format signed off by internal audit, and the override escalation policy ratified by the operations committee took 16 weeks — running in parallel. The technology was ready before the organisation was.

I thought the hard work would be the AI. The hard work was getting four different oversight committees to agree on what "good" looked like. Once they did, the Worker just slotted in.
— Head of Onboarding Operations, the client

What we would do differently

Two things, with the benefit of hindsight.

1. Engage compliance, audit, and operations governance in week 1, not week 8. The 16-week governance cycle was the bottleneck. Front-loading it would have shortened the overall timeline by ~6 weeks.
2. Define the override-rate target before launch, not after. We launched without an agreed acceptable override rate, then had to negotiate it under pressure when the first month came in at 11%. Setting 5% as the acceptance target before go-live would have removed ambiguity.

What this client did next

Three things, all on the same Teragrid Ai platform: (1) extended the Worker to handle KYC document refresh cycles, not just new onboarding; (2) deployed a second Digital Worker for transaction monitoring (AML) triage; (3) began the data work to support a third Worker on collections-letter generation. The compounding pattern — second and third Workers ship in 6 weeks instead of 11 — is the real long-term ROI.

Discuss your own KYC or AML use case →