Most banks now have AI running somewhere. Fewer can say what it returned. Evident's review of bank AI use cases found that 38% of those announced in Q1 2026 reported an outcome (source). The rest went live without a number attached.
I do not think this is a technology problem. It is an ordering problem. The number gets discussed after launch, when it should be agreed before the first sprint.
What I did on a chatbot build
I sat on the build committee for a consumer bank's AI chatbot, and my part was its success KPIs. Before launch we defined the target use cases, the digital adoption target and the customer satisfaction target. We also wrote the customer-experience framework that steered pre-launch design and rollout readiness.
That work changed the build. Once a use case has a target, weak use cases drop out early, and the team argues about customers instead of features.
Three rules I now hold
- Name the number and its owner before the build. If nobody owns the result, nobody will report it.
- Pick a number a customer can move. Adoption, conversion and satisfaction tell you whether the AI helped someone. Model accuracy alone does not.
- Agree what would make you stop. A pilot with no stopping rule becomes a permanent pilot.
When I later deployed an AI-driven personalised offer engine, it delivered a 60% conversion lift across targeted engagement journeys. I can state that figure because measurement was part of the design. A lift is only a lift against something.
If your AI is live and the return is unclear, go back one step. Ask what the number was meant to be, and who agreed to it.