Danai Data Book a discovery call

Blog In the news

80% of UK financial firms already run AI agents, say the Bank of England and FCA. How to choose your first AI agent use case: a 7-point checklist

Herman Wa · · 3 min read

Illustration

The news. On 26 October 2026 the Bank of England and the FCA published their fourth survey of AI in UK financial services. 86% of responding firms use AI. Of the 105 firms that completed the full survey, 80% have AI agents in use, mostly for decision support, and a further 14% plan to within a year. Only 2% let agents act with broad autonomy. More than half find it hard to apply model risk management to systems bought from vendors. These are self-reported figures from one sector, but the direction is clear.

The firms furthest ahead started with agents that support a decision a person still makes, and kept the risky actions for later. That is the right shape for a first use case in any sector.

The fastest way to fail with AI agents is to start too big. "Transform the finance function" is not a use case. "Decide each quarter which suppliers to put on watch" is.

Here is how we help clients pick a first workflow that will show results in weeks.

Look for the meeting that keeps coming back

Every organisation has a decision that gets revisited again and again: the pricing review that reopens every quarter, the go/no-go on a new market that has been "nearly decided" three times, the supplier shortlist that never quite settles.

Those recurring decisions are ideal first use cases. They already have evidence, stakeholders and a cost in time. An agent team can prepare the ground, surface the disagreements early and leave a record, so the human meeting becomes shorter and better.

The checklist

A good first use case scores well on most of these:

  1. It is a decision, not a document. "Choose between three options" beats "write a report".
  2. The evidence exists. There are documents, data or reports the agents can read. Agents cannot reason from nothing.
  3. It has several angles. Cost, risk, customers, operations. If only one angle matters, you do not need a team.
  4. It recurs. Monthly or quarterly is ideal. The team's memory and standing orders improve with each run.
  5. Someone owns it. A named person who will read the minutes, sign off rule changes and tell us what was wrong.
  6. The data risk is manageable. For a first project, pick something where the data controls are straightforward. Save the most sensitive material for when the controls have been proven in your environment.
  7. Success is visible. Less time in meetings, fewer reopened decisions, or a clearer record for the board.

Examples by function

These are illustrative starting points, not promises:

  • Strategy: a quarterly review of which markets or segments to prioritise, argued by market, finance and operations agents.
  • Finance: investment cases with stated confidence and a record of who disagreed with which assumption.
  • Procurement: supplier risk reviews combining financial health, delivery record and sustainability data.
  • Legal and compliance: policy reviews where every point is sourced and every change waits for sign-off.
  • Product: roadmap trade-offs weighed by agents representing customers, engineering effort and cost.
  • Property: portfolio reviews of energy performance, compliance deadlines and capital spend.

What to avoid first time

  • Customer-facing automation. Start internal, where a wrong answer is caught by your own people.
  • Anything with no owner. Without a person reading the minutes, nothing improves.
  • The most sensitive data you hold. Prove the controls on something lower-risk first.

How the Discovery Sprint works

Our Discovery Sprint takes two weeks. We map one decision-heavy workflow with you, design the agent team (the roles, the rules and what people decide), and end with a working session on one of your own questions. You see real minutes on a real decision before committing to a full build.

If you already have a decision in mind, bring it to a thirty-minute call. We will tell you plainly whether it is a good first candidate.

Sources

Read next