The pilot worked. Then it met real data.
Most agentic AI pilots in banking and insurance don’t fail because the model misbehaves. They fail because your estate runs on knowledge nobody wrote down. The analyst knew what status code seven meant. The pricing manager waited a day for a reliable balance file. Agents remove that human layer, and what’s left is your architecture, exposed.
In this ebook, Rav Hayer, Head of UK/I and EMEA BFSI at Thoughtworks, sets out what that means for the people who design the estate:
- Why semantics, entitlements and evidence are now production infrastructure.
- Where privacy and security controls must sit in the request path.
- How to scale in thin slices without stalling on a platform program.
Correct data isn’t enough anymore
You spent a decade making data correct. Now it has to make sense to something that can’t ask a colleague. Give an agent two accurate fields, both called customer status, and it will pick one, confidently, and carry on. It won’t raise a ticket. You’ll find out later, from a customer or a supervisor.
Meaning comes first
Eight shifts in data architecture, from schemas to versioned semantics and from lineage to the evidence behind each decision.
Gates, not principles
Seven guardrails ordered by where they act in the request lifecycle, including the one most firms are missing.
Privacy blind spots
Six failure points in an agentic estate, from purpose creep to deletion requests that never reach the vector index.
Thin slices, not programs
What to do in the next 90 days, the next 12 months and beyond, with each stage paid for by the use cases inside it.
The hard questions
A dozen tests of where your firm really stands, such as: if you had to change model provider in 90 days, what breaks?
Build AI that works on the estate you actually have
Pick one use case. Build the meaning, permissions and evidence it needs. See which parts of your estate were being held together by people.We help banks and insurers rebuild their core systems and data so agents can interpret them, and rewire for agents with governance that runs in the request path, not in a committee. Agent/works™, our enterprise platform for governing and running autonomous AI agents, brings that discipline to every agent you put into production.