Assess one operations team at a time, starting where the first agent will work, such as payments exceptions or customer onboarding. Each person takes a short test on governed-agent concepts set in banking. The report scores the team by concept, names its gaps, and tells the bank which agent to start with and how far it may act.
Why does a bank need a team-level assessment?
Banks already assess models. Model risk teams validate the agent, its sources and its certificate. What they rarely assess is the people who will review the agent's held actions every day. A well-validated agent in front of a team that misreads its verdicts is a control failure waiting for a busy afternoon.
The assessment fills that gap. It asks whether the team understands depth, the decision chain, the four verdicts, provenance and the record, using banking situations. The head of the operations area owns it; the operational risk partner should see the result. For the concept in general, see what a readiness assessment is, and for the detail of the scoring, what a readiness assessment for AI agents measures.
Who should take it, and in what order?
One function per sitting, up to 25 people. Mixing payments, onboarding and lending in one cohort produces an average that describes nobody.
Start where the first agent will work. For most banks that is one of three queues: payment exceptions and repairs, onboarding document checks, or customer servicing requests such as address changes and standing orders. Include everyone who will approve held items, their team leads, and the supervisor who signs off exceptions above the bank's own limits.
The second line should observe, not sit with the cohort. Compliance and operational risk can take their own sitting later; their questions are different.
The steps
- Person (head of operations): names the function and the cohort, and agrees with the risk partner what a "ready" result means before anyone sits the test.
- Person (training lead): asks for the test set in banking, so questions involve held payments, identity documents and ledger entries.
- Person (each team member): sits the test individually, without preparation.
- Person (head of operations): reads the report by concept with the risk partner.
- Person (head of operations): picks the first agent and its starting depth. The default is a depth 2 agent that drafts for the team, whatever the business case says about depth 4.
- Person (training lead): trains the weakest concept first, starting with whoever approves the most items.
- Agent, depth 1: shadows the live queue, reading cases and producing the verdict it would have given, which the team compares with its own decisions.
- Person (cohort): retakes the weak concept before the agent drafts on live work.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Define ready | Nothing yet | None | The threshold, agreed with risk |
| Sit the test | Nothing yet | None | Individual answers |
| Read the report | Nothing yet | None | Interpreting by concept |
| Choose first agent | Named as candidate | None | Agent and starting depth |
| Train the gap | Runs in practice labs | Practice | Learning |
| Shadow | Reads and proposes verdicts | 1 | Real decisions |
| Reassess | Nothing new | 1 then 2 | Go-live decision |
What does the test usually find in banks?
Banks score well on depth. Staff already work with mandates, approval limits and four-eyes checks, so "depth 4 moves money; a person approves anything above the limit" feels familiar.
Withheld actions are the common confusion. Many staff assume that anything the bank can do, a sufficiently trusted agent can be allowed to do. They need to understand that some actions, such as reversing a settled payment on the bank's own rail, are withheld at the platform floor for every agent at every depth. That is not a setting the bank can change, and a team that expects to change it will raise the wrong escalations.
Provenance is the other gap. The core banking ledger is sealed; a customer's emailed instruction is external. Payments teams know emailed instructions are risky, but not always that an external claim cannot authorise a payment however convincing it looks. The test makes that concrete.
What does the report change?
The first agent. A team weak on provenance should not start with an agent that matches incoming instructions to accounts. A drafting agent for servicing letters is a safer start.
The order of training. If supervisors score below their teams, supervisors train first, because they decide what is escalated.
The conversation with risk. The report gives the risk partner evidence that people, not just the model, were assessed. Where the EU AI Act applies, Article 4 asks organisations using AI to take measures for adequate AI literacy among their staff, and a scored assessment is one way to show the starting point. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What should the bank keep as evidence?
The definition of ready, agreed before the test.
The report by concept, with the gaps named.
The decision on the first agent, with who made it and why.
The reassessment, showing the gap closed before live work.
What should the head of operations measure?
- Scores by concept for each function, compared over time.
- During shadow, how often the team and the agent reached different decisions, and who was right.
- How many approvers hold a current reviewer certificate before the agent drafts on live work.
Where does it go wrong in banks?
Assessing the whole bank at once. The result is a large, flat average and no decision.
Letting the business case set the depth. The case assumed depth 4 from day one. The report says the team is ready for depth 2. The report should win, just as evidence from certification sets the agent's own depth.
Leaving the second line out. If risk sees the result only after go-live, it cannot challenge the choice of first agent.
A fintech runs the same assessment with smaller, faster-changing teams; see the fintech version. Contact centres in banks face a different mix; compare the contact centre version. The report format is in the AI agent readiness assessment report template, and the topic guide is training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy offers a free readiness assessment for teams of up to 25: the team takes the Foundations test set in banking, and the report shows scores by concept, the gaps and the first agent to start with. Approvers can then take the Operator track, whose certificate a bank can require before granting approval rights at each depth.
Questions people ask
Does a strong result mean the team can approve depth 4 actions?
No. The assessment tests understanding. Approving money movement reliably is shown through supervised real decisions and certification.
Should contractors and offshore teams take it?
Yes, if they will approve held items. The bank's accountability does not change with who employs the reviewer.
How long before go-live should the assessment run?
Six to ten weeks, so there is time to train a weak concept and reassess.