How to run customer risk rating reviews with an AI agent

The agent prepares the review pack; an analyst sets the rating. For each customer due for review, it refreshes the risk factors, compares actual activity with what was expected, re-runs your rating matrix and proposes a rating with reasons. A named analyst confirms or changes it, and any downgrade from high risk goes to a senior approver.

What is a customer risk rating review, and who owns it?

A customer risk rating places each customer in a band, typically low, medium or high, based on who they are, where they operate, what products they use and how they transact. The band decides how much due diligence applies and how often the customer is looked at again. A review checks whether the band is still right.

Reviews run on a cycle set by risk: high-risk customers more often, low-risk less often. They also run when something changes, which is the subject of trigger-event KYC reviews with an AI agent. The financial crime compliance function owns the methodology and the customer risk rating matrix. Analysts in the first or second line run the reviews, depending on the firm.

The agent does the assembly. It does not own the rating, which follows our wider position on AI compliance agents.

What are the steps, and what does the agent prepare?

1. Agent, depth 1: builds the review queue from customers whose review date has arrived, ordered by current band and how overdue they are.
2. Agent, depth 1: refreshes the factors. Ownership and control, occupation or business type, countries involved, products held, delivery channel. Sealed data from the core system is used as is; the customer's own statements are marked as declared.
3. Agent, depth 1: compares activity with expectation. Volumes, values, counterparties and countries against what the customer said at onboarding or last review. Differences are listed, not explained away.
4. Agent, depth 1: re-runs screening results for sanctions, politically exposed persons and adverse media, and attaches any open alerts. It does not clear alerts here.
5. Agent, depth 2: applies the matrix and proposes a rating. Each factor's score, the total, the resulting band, and a short narrative of what changed since the last review.
6. Agent, depth 2: drafts information requests where data is missing or stale, for the analyst to send.
7. Person, analyst: sets the rating. Confirms the proposal or overrides it with a written reason.
8. Person, senior approver: approves downgrades from high and any override that lowers the matrix result.
9. Agent, depth 3: schedules the next review from the approved band, and updates the due diligence requirements that follow from it.

Which controls matter, and what evidence does each leave?

The rating is a person's act. The agent's role scope stops at proposing. Evidence: the analyst's name on each final rating.

Downgrades are harder than upgrades. Moving a customer down reduces scrutiny, so a lower band, or an override below the matrix result, is held for a senior approver. Evidence: held verdicts and approvals on every downgrade.

Matrix version on every proposal. If the matrix changes, you can find every review scored under the old one. Evidence: version reference on each pack.

Declared facts are labelled. A customer's statement of expected turnover is not evidence of turnover. Evidence: provenance on each factor.

Overrides carry reasons. Evidence: written reason against each override, reviewable as a set.

What should you measure?

• Overrides of the proposed rating, by direction and analyst. A cluster of downgrades by one person is worth a look.
• Reviews completed on time, by band. High-risk overdue reviews matter most.
• Stale data: factors older than your policy allows at the time of review.
• Activity mismatches found, and how many led to a rating change or a suspicious activity referral.
• Information requests sent and answered, and how long customers take to reply.

Where does it usually go wrong?

Backlog pressure lowers ratings. A team behind on high-risk reviews is tempted to rebase customers to medium. The held step for downgrades exists for exactly this.

Expected activity was never recorded. If onboarding captured no expectation, the comparison at step 3 is empty. The agent should report the gap rather than infer an expectation from current behaviour.

Screening is treated as done. An open alert attached to the pack is not cleared by the review. Keep alert clearance in its own process.

The matrix is gamed by data entry. Changing a business type code can move a customer a whole band. Watch factor changes made shortly before a review.

Rating feeds back into nothing. The approved band must update monitoring thresholds and review frequency. Check that it did.

Where Pryme fits

Pryme files each review pack with the provenance of every factor, the matrix version and the agent's proposed band. The customer's own limits can send any downgrade to a held verdict for a named approver. Whether an agent could ever set the rating alone is answered in can an AI agent decide a customer's risk rating?