It depends on whether the hold follows a written rule. An AI agent can apply a short, pre-defined hold on one settlement batch when a trigger in the acquirer's risk policy fires, and then hand the case to a risk analyst. Setting a reserve, extending a hold, keeping funds against losses or releasing them stays with a named person.
Why is holding settlement a serious action?
For a small merchant, settlement is payroll and rent. Holding it for a week can do more damage than the fraud it was meant to prevent. For the acquirer, the hold is protection: if a merchant's customers start charging back, or the merchant disappears, the acquirer carries the loss.
Both interests are legitimate, which is why the decision belongs to a policy and a person, not to an agent's reading of a dashboard. The agent's value is speed and consistency: it sees the trigger as it fires, applies exactly the hold the policy allows, and makes sure a person looks at the case before the merchant is harmed further.
What conditions decide it?
The agent may apply a hold only when all of these are true:
- the merchant agreement allows the acquirer to delay settlement in the stated circumstances;
- the trigger is defined in the risk policy in measurable terms, such as a chargeback count above a set level or volume far outside the merchant's declared profile;
- the hold is limited to a set period and a set scope, such as one batch;
- a risk analyst is assigned to review the case before the hold period ends.
If the policy says "hold settlement where risk is elevated", that is not a trigger the agent can act on. Vague rules produce inconsistent holds, and inconsistent holds produce complaints.
What may the agent do, and what stays with a person?
| Action | The agent | Stays with a person |
|---|---|---|
| Monitoring chargebacks, refunds and volume | Reads them continuously, depth 1 | Nothing |
| Applying a policy-defined batch hold | Applies it within its time limit, depth 3 | Holds outside the written triggers |
| Drafting the notice to the merchant | Drafts it with the reason and timeline, depth 2 | Sending it where the reason is sensitive |
| Setting or raising a rolling reserve | Prepares the analysis | Deciding the reserve |
| Extending a hold | Never | The risk analyst |
| Releasing held funds | Prepares the release with the review notes | Approving the release |
| Keeping funds against losses | Never | Risk and legal teams |
A hold is easier to justify than many neighbouring actions because it can be undone, provided funds have not yet been paid out. Once settlement is final, the picture changes: see what settlement finality means.
An example from a payments company
Take a merchant selling event tickets that normally processes about 20,000 a week with few chargebacks. On a Tuesday, the agent sees 60,000 of sales in a day, many from new card numbers, followed by the first chargebacks citing goods not received. The risk policy defines this pattern as a trigger for a 48-hour hold on the next settlement batch.
The agent applies the hold, drafts a notice to the merchant explaining that settlement is delayed for review and when they will hear, and opens a case for the risk team with the volume figures and chargeback reasons. The analyst calls the merchant, learns the event has been cancelled and refunds are pending, and agrees a reserve to cover expected chargebacks. The agent did none of that judgement. It bought the analyst two days.
If the analyst suspected money laundering instead, the notice to the merchant would be reviewed first. Telling a merchant why funds are held can, in some cases, alert someone to a suspicion they must not learn about.
What controls make it safe?
- Triggers as code, from the policy library. The agent acts on sealed policy definitions, not on its own reading of what looks risky.
- A clock on every hold. Each hold has an expiry and a named reviewer. An unreviewed hold escalates to a manager before it expires.
- No release capability. Releasing funds is a separate capability the agent does not hold.
- Notices drafted, not sent, where reasons are sensitive. The agent drafts; a person checks wording when fraud or financial crime is suspected.
- Every hold filed. The verdict records the trigger, the bound that allowed it and the reviewer assigned.
Holds sit beside other merchant controls. Limits are covered in whether an AI agent can set a merchant's transaction limits, and the earliest controls start at merchant onboarding for card acquiring. The wider rules are in AI agents in banking: what they may and may not do.
Questions people ask
Can the agent hold all of a merchant's future settlements?
No. An open-ended hold is a decision about the relationship, not a routine control. The agent can recommend it with the evidence; a person decides and owns the conversation with the merchant.
What if the trigger fires on a merchant's busiest day of the year?
That is why a person reviews every hold quickly. A seasonal spike that matches the merchant's history should be released fast, and the policy should learn from the false trigger.
Can the merchant challenge the hold through the agent?
The agent can take the merchant's explanation and attach it to the case. The analyst considers it; the agent does not decide whether it is convincing.