From pilot to production: AI inside governed workflows

Most financial institutions run AI pilots; far fewer run AI in day-to-day operations. MIT's 2025 research found only around one custom tool in twenty reaches production. The gap is rarely the technology.

The paper sets out the three gates that decide the outcome — data the institution can prove, model risk it can validate, and a named owner accountable for the result — alongside the two questions banks ask first: what does this cost at scale, and where does our data go. The same workload can run at a fraction of its unengineered cost through caching, prompt design and matching the model to the task, and three deployment patterns keep customer data under the institution's control, with personal data masked before the model ever sees it.

The full paper covers the use cases that pay first — screening, surveillance and client onboarding, drawn from production deliveries in regulated institutions — plus the EU AI Act timetable and the governance supervisors now expect. Sources include MIT NANDA and Lenovo/IDC.