Sanctions hits: explaining a match, not clearing it

An agent can explain a sanctions hit: why the name matched, which identifiers agree, which differ, and what further information would settle it. It should not clear the hit. Discounting a potential match releases a payment or a relationship that may be prohibited, and that decision stays with a trained analyst who answers for it.

What does a good match explanation contain?

Most sanctions alerts are false positives, and analysts spend much of their day working out why. A screening tool flags that "A. Rahman" on a payment resembles a listed name. The analyst then has to compare dates of birth, nationalities, addresses, known aliases and any other identifiers, check whether the transaction context makes sense, and write down their reasoning.

An agent can do the comparison and write it up. A useful explanation includes:

1. The listed entry that triggered the alert, with the list name and version.
2. What matched: the name elements and how closely, plus any other agreeing identifiers.
3. What differs: each identifier that does not agree, and how.
4. What is unknown: identifiers the firm does not hold for its customer or counterparty.
5. Context: the customer's history, the payment's purpose, the counterparties involved.
6. What would settle it: the specific further information that would confirm or rule out the match.

A sanctions match explanation template gives these fields a consistent shape, which makes explanations faster to read and easier to audit.

Why must clearing stay with an analyst?

Because clearing is a decision with legal weight. Dealing with a sanctioned person or entity is prohibited in many jurisdictions, and the consequences of getting it wrong fall on the firm and sometimes on individuals. When an analyst clears a hit, they are deciding that the evidence shows the customer is not the listed party. That decision needs someone who can be asked to account for it.

There is also an asymmetry in the errors. A false positive wrongly held costs the customer a delay. A true positive wrongly cleared may mean the firm has dealt with a sanctioned party. No rate of correct clearances on routine cases offsets one wrong clearance of a real match. That asymmetry is exactly why the decision should rest with an accountable person, however good the agent's explanations become. The short answer to whether an AI agent can clear a sanctions alert is no.

Doesn't the explanation bias the analyst?

It can, and this is the real risk to manage. An explanation that leads with differences and buries agreements nudges towards clearing. One written in confident prose makes thin evidence feel conclusive. Analysts working through a long queue will be tempted to read the summary and accept its implied conclusion.

Three design choices reduce the pull. First, present agreements and differences with equal prominence, in a fixed structure, so the format does not tilt the reader. Second, have the agent state what it does not know as clearly as what it does. Third, do not let the agent write a recommendation such as "likely false positive" unless the firm has deliberately decided to allow that and monitors whether analysts simply follow it.

Sampling helps too. If analysts clear nearly every alert the agent presents as low-similarity, check a sample of those clearances independently. If the clearances hold up, the explanations are useful. If they do not, the explanations are steering.

How should held payments be handled?

A payment with an open sanctions hit should be held, not refused or released. The hold sits with a named analyst until they decide.

The full process is set out in how to review payment sanctions hits before release with an AI agent, and the integration side in how to connect an AI agent to your sanctions screening tool.

What happens with true matches?

Rarely, but it happens. When an explanation points towards a real match, the agent's job is to make that as clear as possible, as quickly as possible, and stop. Freezing, reporting to the authorities and deciding what to say to the customer are decisions for the analyst and the responsible officer, made under the firm's procedures and the law of the jurisdiction. The agent should have no contact with the customer about the matter.

How do you know the explanations are any good?

Measure them against analyst decisions. For each alert, compare the agent's explanation with what the analyst found. Did the agent miss an identifier? Did it misread a date? Did it present a difference that was not one? Track how often analysts correct an explanation and why. A falling correction rate on a growing sample is evidence the agent's explanations can be trusted as a starting point.

Certification should include known true matches, near misses and false positives, labelled by experienced analysts. The broader picture of what agents can and cannot do here is in whether an AI agent can do sanctions screening.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

What should you do differently?

Use an agent to make every alert arrive with a clear, structured explanation. Use the time saved to give analysts more attention for the alerts that matter, not to process more alerts per hour with less scrutiny. Keep clearing with people, and watch for explanations that start to do the deciding for them. This is the role of AI compliance agents as second line, never the last word.