Swiss bank trial of agentic AI halves onboarding cycle with 99% data extraction accuracy

Incore Bank, Kyndryl and Google Cloud have completed a proof of concept using agentic AI to streamline KYC processes, promising improved accuracy and faster onboarding for institutional clients, though still in the testing phase.

Kyndryl, Incore Bank and Google Cloud say they have finished a Swiss proof of concept that used so-called agentic AI to handle large parts of know-your-customer work, with the partners claiming automated extraction accuracy of up to 99% and an onboarding cycle that could be cut from months to days. The work was announced by Kyndryl in Zurich on 31 August 2026, and Swiss coverage the following day presented it as a completed trial rather than a live, bank-wide deployment. (kyndryl.com)

What the 99% figure measures is narrower than the headline claim may suggest. PYMNTS reported that the number refers to how accurately the system pulled data from onboarding documents, not to whether the machine reached the correct overall risk decision 99% of the time. That distinction matters in compliance, where staff are not merely reading passports or company filings but reconciling information across internal systems and outside sources, investigating red flags and assembling a record a regulator can later inspect. PYMNTS added that only 20% of financial firms currently use AI for KYC and identity checks, although 65% plan to adopt or expand such tools within the next 12 months. (pymnts.com)

According to Kyndryl, the trial was built on Google’s Gemini models and Kyndryl’s Agentic AI Framework, with a semantic layer designed to sit over existing banking systems. Two elements recur across the reporting: machine-readable policy rules and separate guardrail agents. A short industry briefing described those guardrails in practical terms, saying they can refuse an action rather than merely suggest caution, which is central to the pitch that the system can operate inside a regulated workflow instead of as a free-form assistant. (kyndryl.com)

The workflow was split among specialised agents rather than one model doing everything. Reports said some agents extracted and checked customer information from structured records and unstructured documents, others gathered supporting evidence, and others identified risk factors before another produced an explainable score and an auditable record for review. PYMNTS noted that Google Cloud has separately described a comparable multi-agent KYC set-up in which one agent checks document consistency and validity while another cross-references outside sources, underscoring that the Incore exercise fits a wider move towards breaking compliance work into discrete, governed steps. (pymnts.com)

The companies have also been careful to frame the technology as support for compliance teams, not a substitute for them. Crowdfund Insider reported that final risk ratings and onboarding decisions remain with people, while the automation is meant to remove the most repetitive document review and evidence-gathering tasks. Kyndryl said the design also extends beyond account opening, with agents monitoring for changes in a customer’s circumstances so that KYC becomes an ongoing review process rather than a one-off hurdle at the start of the relationship. (crowdfundinsider.com)

That matters because Incore is not a consumer neobank experimenting at the margins. Independent reports describe it as a FINMA-regulated Swiss B2B transaction bank serving traditional and digital-asset markets. Crowdfund Insider located the bank in Schlieren, while Swiss coverage cast it as Zurich-based; both descriptions are consistent with a bank operating from the Zurich area. KI News Schweiz said Incore serves independent asset managers, financial intermediaries and companies, adding local context that helps explain why a bank of this type would want faster, more traceable onboarding for institutional and business clients rather than mass-market retail accounts. (amlobservatory.org)

Mark Dambacher, Incore Bank’s chief executive, said in Kyndryl’s announcement that innovation must go “hand in hand with trust, transparency and strong regulatory governance”. Jacqueline Wild, managing director of Kyndryl Alps, argued that agentic AI is valuable in banking only if it comes with “explainability, auditability, secure data access and human oversight”. Those remarks go to the heart of how the project is being sold: not as a faster chatbot for back-office work, but as a controlled orchestration layer that can show its workings and preserve an audit trail. (kyndryl.com)

For now, though, the story is still one of promise rather than proof at scale. Multiple outlets stressed that the exercise was a proof of concept, and the headline figures are vendor-reported. Even so, the trial is notable because it pushes the conversation beyond simple document-reading tools towards an end-to-end compliance process in which policy rules, evidence gathering, exception handling and human sign-off are all designed together. In a sector where many AI projects still stall at pilot stage, that governed, workflow-first approach may prove as significant as the extraction score itself. (amlobservatory.org)