Can an AI agent submit regulatory returns?

No, not the submission. An AI agent can compile a regulatory return from source systems, run the validation rules, reconcile figures to the general ledger, explain movements and draft the commentary. Approving and submitting the return stay with a named person, because the firm answers for every figure and a submitted return stays on the record.

What is a regulatory return?

A regulatory return is a periodic report a firm must send its supervisor: capital and liquidity data, client money and safeguarding figures, complaints data, financial crime data, conduct data and more, depending on the firm and the market. Each has a template, definitions, validation rules and a deadline.

This is not the same as handling product returns or payment returns. If you are looking for those, see how to handle returns requests with an AI agent or how to handle payment returns with an AI agent.

Regulatory returns are high-stakes for an unglamorous reason. Supervisors use them to monitor firms, and a wrong figure can trigger questions, a re-submission or, if it hides a problem, enforcement. Late returns can be a breach on their own.

What evidence does the agent assemble?

1. Source extracts: the data for each field, pulled from the ledger, the core system, the complaints system or wherever the definition points.
2. Mapping: how each field was populated, showing the source, the filters and any calculation.
3. Validation: the regulator's validation rules and your own internal checks, run and passed or failed.
4. Reconciliation: totals reconciled to the general ledger and to other returns that share figures.
5. Movements: changes against the previous period, with an explanation for each one above your threshold.
6. Definition changes: any change to the template or definitions since the last return, and how it was applied.
7. Draft commentary and a sign-off pack for the approver.

The sixth item is where returns often go wrong. Templates and definitions change, and a return compiled the old way passes its validations but answers the wrong question. The agent should check each return against the current template, drawing on your regulatory change management process; see also can an AI agent track regulatory change?

What decision does a person keep?

Submission is irreversible in practice. A resubmission is possible, but the first version stays on the regulator's record, and so does the fact that it was wrong.

What does a return look like when the agent prepares it?

Take a consumer credit firm preparing its half-yearly complaints return. The agent pulls complaints from the complaints system, maps them to the return's product and cause categories, and runs the validations. All pass.

But its movement analysis shows complaints about one product down by half against the previous period, while customer contacts about that product are up. The agent finds the cause: a change in the complaints system six weeks earlier made one cause category default to "enquiry", so some complaints were not recorded as complaints. The validations could not catch this, because the return was internally consistent.

The return owner confirms the problem, has the team re-classify the affected cases, and approves the corrected return. Compliance decides the misclassification is a breach of complaint handling rules to be logged, and submits the return. The agent's movement analysis is what found it.

Which controls make agent preparation safe?

• No submission capability. The agent has no access to the regulator's portal.
• Lineage for every figure. Each field links back to its source and calculation.
• Movements explained, not just checked. Validation passing is not the same as the return being right.
• Template version checked. The agent confirms the current template and definitions on every run.
• Approval recorded. The approver's sign-off sits with the pack, so an inspector can see what they reviewed; see can an AI agent help prepare for a regulatory inspection?

The general principle is in our guide to AI compliance agents.