Most AI companies pick a vertical because the sales cycle is short. We picked finance knowing it would be the opposite.
Almost every AI deployment we saw being celebrated in 2025 and 2026 was a customer support chatbot or some form of content summarizer. Useful, easily replaced, and priced accordingly. The work that carries real risk and cost inside a financial institution is nothing like that. It is a credit analyst reading four hundred pages of a CLO (collateralized loan obligation) indenture to extract terms and test covenant compliance. It is a compliance officer checking a UCITS portfolio against investment rules every day. It is trade surveillance, KYC (know your customer) remediation, counterparty onboarding, and the quiet, expensive back-office work that scales linearly with headcount.
That work is document-heavy, rule-bound, and repeated. It is the best-fit problem for AI agents that exists in any industry we looked at. And it sits on top of decades of accumulated data that nobody outside the firm can touch. The ability of AI to ingest this vast amount of data was the opportunity we embraced.
Adoption is not the problem
The stereotype of finance as a laggard is out of date. A 2026 survey of 150 financial services firms found a quarter running fully autonomous AI in production, more than double the cross-industry rate.
The same survey found 61% saying AI had fallen short of the return they expected. Grant Thornton’s 2026 work put a sharper number on the problem: only 18% of banking leaders were fully confident they could pass an independent review of their AI controls within 90 days, and half of banks cited governance and compliance barriers as a direct contributor to AI underperformance or failure.
Firms are deploying. Auditors are asking. The controls are not available, so projects stall in pilot, get scoped down to something harmless… or get switched off.
Bottlenecks are the opportunity
Four problems come up again and again. A better model solves none of them.
The rules have not caught up. In May 2026 the Federal Reserve confirmed that its revised model risk guidance does not cover generative or agentic AI. FINRA’s 2026 oversight report set out the gaps that leaves open. Until the guidance catches up, a firm has to prove its own controls. It cannot point to the vendor it bought from.
Trust is earned in a set order. A senior Standard Chartered executive said publicly last year that a large bank can afford to sit slightly behind the curve, because one public failure costs more than being first is worth. So these purchases turn on a harder test than whether the product works. It has to be explainable to a regulator.
These purchases turn on a harder test than whether the product works. It has to be explainable to a regulator.
Adoption spreads through peer networks. Deloitte’s EMEA survey found AI use at small banks rising from 22% to 52% in two years, after the larger institutions had established the pattern. One credible reference does more work here than any amount of marketing.
Unapproved tools are already inside the building. Reporting through 2026 found staff using them to draft suspicious activity report narratives and generate customer risk scores. None of it was logged. The firm already owns that problem, whether or not it ever signs a contract with us.
Finance has the highest ceiling and the highest cost of entry of any vertical we looked at. We think those are two descriptions of the same thing.
Part two sets out what we built as a result.
MisaLabs builds the enterprise AI control plane. Govern AI usage. Prove the ROI. In your environment.
Talk to us →