• Incore Bank, Kyndryl and Google Cloud ran a multi-agent AI proof of concept for KYC onboarding, reported around September 2-4, 2026.
• The headline 99% figure is document-extraction accuracy, not the accuracy of the AI's risk-scoring output — a compliance officer still makes the final call.
• Incore Bank operates as B2B banking infrastructure for other Swiss financial institutions, so any onboarding-speed gain could pass through to the banks, fintechs and crypto platforms that rely on its rails.
• No party has disclosed a production rollout date, cost, or an independent audit of the risk-scoring agent's own accuracy.
• The pilot pairs Kyndryl's Agentic AI Framework with Google's Gemini models, splitting KYC work across separate agents for extraction, evidence-gathering, risk-flagging and audit-record generation.

Incore Bank, Kyndryl and Google Cloud ran a proof-of-concept multi-agent AI system that reportedly cut KYC customer onboarding from months to days and hit up to 99% accuracy — but that accuracy figure measures how well the system extracted data from documents, not whether its risk-flagging or scoring was correct. No production date, cost, or independent audit of the risk-scoring agent's accuracy has been disclosed, per Finextra and PYMNTS.

The 99% Number Is Extraction Accuracy, Not Risk Accuracy

Incore Bank, Kyndryl and Google Cloud say their KYC proof of concept hit up to 99% accuracy — but per PYMNTS' reporting, that figure measures how correctly the system pulled fields such as names, addresses and identification numbers off onboarding documents, not whether the AI's risk-factor flags or generated risk score were correct. The distinction matters: document-extraction accuracy has been a largely solved problem for years, while risk-scoring accuracy — the part regulators actually care about — was not disclosed in either the Finextra or PYMNTS accounts of the pilot.

Four Agents, One Workflow

The pilot splits KYC onboarding across specialized AI agents built on Kyndryl's Agentic AI Framework and Google's Gemini models, according to Finextra and PYMNTS: one agent extracts and verifies customer information, another gathers supporting evidence, a third flags risk factors, and a fourth produces an explainable risk score alongside an audit-ready record. Incore Bank frames the design as human-governed — a compliance officer still reviews and signs off on the agents' output rather than the system approving or rejecting a customer on its own, both outlets reported.

Why A Back-Office Vendor's Pilot Reaches Beyond Incore

Incore Bank is not a retail lender; it is a FINMA-regulated wholesale bank in Zurich that supplies custody, settlement and banking-as-a-service infrastructure to other Swiss financial institutions and crypto platforms. That structure means an onboarding process that reportedly shrinks from months to days at Incore would not just save Incore's own compliance desk time — it would shorten the queue for every downstream client bank, neobank or exchange that routes customer onboarding through Incore's rails, assuming the pilot reaches production.

What Hasn't Been Confirmed

Neither Incore Bank, Kyndryl nor Google Cloud has published a production deployment date, a cost figure, or an independent audit of the risk-scoring agent's error rate, based on the available reporting. Incore Bank CEO Mark Dambacher framed the project around 'trust, transparency' and regulatory rigor, according to PYMNTS, but no outlet reported a timeline for moving from proof of concept to live use. Until Incore or its partners disclose risk-decision accuracy and a rollout date, the months-to-days claim describes a lab result, not a production metric.