It can prepare the response; it must not send it. An AI agent can log a law enforcement request, check it cites a legal basis and scope, find the records and draft the response. Verifying that the request is genuine, judging the legal basis and releasing any data stay with a named person in compliance or legal.
Why are these requests risky to automate?
Firms receive requests from police, tax authorities, financial intelligence units and other agencies: production orders, information requests, account monitoring orders and informal enquiries. Each asks for customer data, and each puts the firm between two duties. It must cooperate with lawful requests, and it must not disclose personal data without a proper basis.
Two risks make this a person's decision. First, fake requests are a known fraud and stalking route: a forged letter on a police letterhead asking for a person's address or account details. Second, the legal basis matters. A court order may compel disclosure; an informal request may permit it but not require it, and the firm then has to decide whether disclosure is lawful and proportionate.
What evidence does the agent assemble?
- Request record: who sent it, how it arrived, the reference, the deadline and the stated legal basis.
- Authenticity checks: whether the sender's contact details match those held for that agency, and whether the request came through a known channel. An email arriving from outside is treated as an external claim until verified.
- Scope: exactly what is requested, about whom, and for what period.
- Records found: responsive data from each system, with the search terms and systems recorded.
- Out-of-scope data: anything found that the request does not cover, set aside.
- Linked matters: whether the subject has open AML cases, reports, complaints or other requests, including subject access requests.
- Draft response: a covering letter and the data package, in the format requested.
The full method is in how to respond to law enforcement information requests with an AI agent. Requests from financial intelligence units follow a similar pattern with some differences, covered in how to respond to financial intelligence unit requests with an AI agent.
What does a person decide?
| Decision | Why a person |
|---|---|
| Whether the request is genuine | A wrong answer gives a fraudster or stalker someone's data |
| Whether the legal basis is sufficient | A legal judgement |
| What to disclose | Must match the scope and be proportionate |
| Whether to tell the customer | Disclosure may prejudice an investigation, or may be required |
| Release of the response | Irreversible once sent |
| Conflicts with other requests | For example, a pending erasure request from the same person |
On the last point: if the subject has asked for their data to be erased, a law enforcement request usually means it must be kept. The agent flags the overlap; see can an AI agent handle a right to erasure request?
What does a response look like?
Take a digital bank. An email arrives from an address claiming to be a regional police fraud unit, asking for account statements and the registered address of a named customer, citing an investigation reference and asking for a reply within two days.
The agent logs it and checks the sender against the bank's register of agency contacts. The domain is correct, but the officer's name and phone number do not appear. It finds the customer, gathers six months of statements and the address, and drafts the response. It also notes that the customer reported domestic abuse to the bank some time ago and asked that their address be protected.
The specialist in the law enforcement liaison team calls the police force's published switchboard and finds there is no such officer. The request is fake. She refuses it, reports the attempt and adds a note to the customer's record. The agent had gathered everything, but nothing left the bank. That is why release is enforced at the moment of action, not by trust in the sender; see why enforcement must happen at the moment of action.
Which controls make it safe?
- Release held for a named person. No data leaves on the agent's decision.
- External requests cannot authorise disclosure. A request's own claims about its authority are never enough.
- Scope matching. The response contains only what the request covers.
- A request register. Every request, decision and disclosure is filed.
- Protected customers flagged. Records marked for protection are highlighted on every request.
The general principle is in our guide to AI compliance agents.
Questions people ask
Can the agent respond to urgent requests where life is at risk?
It can prioritise and alert the on-call specialist at once. The decision to disclose, even urgently, stays with a person.
Should the agent tell the customer about the request?
Not without a person's decision. Telling the customer can prejudice an investigation; some orders prohibit it.
Can the agent handle requests from overseas authorities?
It can log and prepare them. Overseas requests often need to come through formal channels, and a person decides whether this one has.