No, not as the decision-maker. A granted permit or licence is acted on almost at once: people build, trade or hire in reliance on it, and taking it back is a separate legal process, not an undo. An agent can check completeness, verify documents against official records, and draft a recommendation. A named official makes the grant.
Why is a grant so hard to reverse?
Most agent decisions can be weighed by asking what happens if they are wrong. For a permit or licence, the answer is that the applicant relies on it. A builder starts work under a planning consent. A business opens under a trading licence. A professional begins practising. Revoking the grant later involves notice, reasons, often a right of appeal, and sometimes compensation. The harm of a wrong grant falls on third parties too: neighbours, customers, the public.
Refusals are not easy either. A refused applicant is entitled to reasons, and public bodies face added duties of fairness and transparency. Under the EU AI Act, some public-sector uses of AI are classed as high-risk, which brings further obligations on those deploying them.
Both directions, then, are decisions with significant effects. Under GDPR, decisions about individuals based solely on automated processing with legal or similarly significant effects are restricted. A permit decision fits that description closely.
What can an agent do in the process?
A great deal of the work around the decision.
| Step | Agent | Person |
|---|---|---|
| Check the application is complete | Yes, depth 3: return incomplete ones with a list of what is missing | Exceptions |
| Verify documents against official registers | Yes, depth 1 to 3 | Resolves mismatches |
| Check eligibility rules that are purely factual | Yes, and record the result | Reviews borderline cases |
| Gather consultation responses | Yes | None |
| Draft a recommendation with reasons | Yes, depth 2 | Reads and decides |
| Grant or refuse | No | The named official |
| Issue the permit or the refusal letter | Prepare it | Signs and sends |
The pattern is that the agent does the assembling and checking, and the official does the judging. That is also what makes the official's decision better: they see a complete file with sources marked, instead of spending their time chasing documents.
How should the agent treat what applicants say?
Much of an application is the applicant's own declaration: their experience, their premises, their plans. Those are claims, not verified facts. The agent should mark each one by source: confirmed against a register, supported by a signed document, or declared by the applicant only. The official then sees at a glance which parts of the case rest on the applicant's word.
This mirrors how lenders handle applications, covered in how to review loan applications for application fraud with an AI agent.
What does this look like in practice?
Take a city licensing team handling applications for street trading licences. An agent receives each application, checks every required document is present, verifies the applicant's identity and business registration against official records, confirms the pitch is in a permitted zone, and collects comments from the highways and environmental health teams.
It then drafts a recommendation: "complete; identity and registration verified; pitch permitted; one objection from highways on access, attached; applicant declares five years' trading experience, not verified." A licensing officer reads the file, phones the highways team about the objection, and grants the licence with a condition on trading hours. The agent prepares the licence document; the officer signs it. Every step, including the officer's decision and reasons, is in one record.
Is this different from approving a claim or declining a loan?
The structure is similar: the agent prepares, a person decides the significant outcome. What differs is who bears the cost of a wrong decision. With an insurance claim, a wrong approval mainly costs the insurer. With a loan decline, it mainly costs the applicant. With a permit, it can cost the public. That is the strongest reason for keeping the decision human. The wider argument, that autonomy is a question about reversal, is in autonomy in regulated industries.
Which controls should a public body put in place?
- Name the official accountable for each decision type.
- Mark every fact in the file with its source.
- Keep the agent's recommendation and the official's decision separate in the record, with reasons for both.
- Test the agent on past applications before use and recertify regularly; how to approve a new AI agent for production covers the sign-off.
- Tell applicants where AI was used and how to ask for a person.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Could an agent ever grant routine, low-risk permits?
Some bodies may decide that very narrow, rule-based permits can be issued automatically, as already happens with simple online forms. That is a policy decision for the body and its legal advisers, not a default.
Can the agent refuse incomplete applications?
It can return them as incomplete with a list of what is missing, which is not a refusal on the merits. Final refusals stay with the official.
How do we explain an agent's role to applicants?
Say plainly which steps the agent performed, that an official made the decision, and how to ask questions or appeal.