AI in banks and insurers: from pilot to production under audit

Field notes from Strategus. October 2026.

Banks and insurers have no shortage of AI pilots. Few of them reach production, and the reason is rarely the model. It is the path through risk, compliance, security and audit. Here is what it takes to get through it.

Why financial AI pilots stall

• The pilot was built on data that production can't use
• Risk and compliance see the system for the first time at the end
• Nobody can explain how an answer was produced
• No owner in the business once the pilot team moves on

Principle 1: Start with a workflow, not a model

Pick one workflow with a clear owner and a measurable outcome: time to answer a customer, time to prepare a proposal, hours spent on document review. A narrow workflow with a number attached gets approved. A general AI assistant does not.

Principle 2: Bring risk and compliance in during week one

Show them the data flow, where the model runs, what it can and can't see, and how a person reviews its output. Agree on the approval criteria before the build. Their questions are much cheaper to answer on a whiteboard than after go-live.

Principle 3: Keep data where it lives

Run the system in your own cloud account, private cloud or on-prem environment. Customer data should not leave your controls to make the system work. This one decision removes a large share of the security review.

Principle 4: Make every answer traceable

Log every question, the documents retrieved, the model version and the answer. Show sources to the user. When an auditor asks why the system said something, you should be able to replay it.

Principle 5: Keep a person in the loop where it matters

Use AI to draft, search and summarize. Keep a person accountable for decisions that affect customers, credit or claims. Design the review step into the workflow, not as an afterthought.

Principle 6: Measure before and after

Record the baseline before you build: handling time, error rate, backlog. Measure the same numbers after go-live. That is the evidence that funds the next workflow.

Where results come fastest

• Internal knowledge assistants for operations and service teams
• Drafting proposals and client summaries from CRM and product data
• Document intake and review for onboarding and claims
• Customer service assistants that suggest the next best action

A realistic timeline

With access and approvals planned from week one, a first workflow can be in production within 5 weeks of kickoff, and rolled out across teams in about 6.

Checklist before you start

• One workflow, one owner, one number to improve
• Risk and compliance contacts named in week one
• Agreed deployment environment
• Logging and audit requirements
• Baseline measurement
• Review step designed into the workflow