Compliance, done by an agent,certified by a human.

Some of the better opportunities in enterprise AI sit in markets nobody calls AI-first. Compliance is one. The work is high in volume, rule-bound and repetitive, and the cost of getting it wrong is carried by a person rather than a program.

The tasks are the same in every programme: mapping regulations to controls, finding the gaps, keeping control descriptions current, collecting evidence, and watching for changes to the rules. The volume is high, the rules are explicit, and the judgement needed per item is small. That is the profile machines handle well.

It is also the profile where a mistake travels. A control mapped to the wrong clause reaches an audit file. Evidence that expired last quarter reaches a customer’s questionnaire. By the time anyone reads it, it has been relied on, and the vendor answers for it rather than the model that drafted it.

Regulators have been circling the same point all year. The European Union began enforcing its AI transparency rules on 2 August: a chatbot must tell people it is not a person, and AI-generated content must carry a visible label and a machine-readable mark. Singapore published a governance framework for agentic AI in January and revised it in May. Six agencies across five countries, among them CISA, the NSA and Britain’s National Cyber Security Centre, issued a joint guide on adopting agentic AI in April. NIST opened an agent standards initiative in February. None of it limits what an agent may do. Each asks who authorised the action and whether a record survives it.

The assurance standards go further. ISO/IEC 42006, published last year, sets out what a body must satisfy to audit and certify an AI management system against ISO/IEC 42001. It is written about the competence of the certifier, not the technology being certified. Buyers run a rougher version of the same test: who reviewed this answer, on what date, against which document.

Which is where the product question sits. The split that works is by task rather than by system. Retrieval, drafting, cross-referencing and monitoring belong to the agent. Judgement and attestation stay with a named person, and that second list is short enough to defend properly.

In practice it means agent output arriving with its provenance attached: the source it drew on, the version it read, and the name of the person who accepted it. Where that holds, a compliance team’s week moves from producing the work to checking it.