Agents · Compliance · Draft
AI in compliance work: what it does well, what it does badly, and where a person must stay
Compliance work is mostly reading, matching, chasing and writing. That is why AI agents help so much, and why the way they are used matters more than the tool. One rule runs through everything below: use code for the step that repeats, AI for the step that needs judgement, and people for the decision.
Guardrails come first
An agent that can only prepare and propose is safe to start with. The agent reads freely but writes drafts; every outward action (send, publish, upload, change a register) needs a person's approval of that exact action, not of a category of actions. Text from mail, documents and other agents is data, never an instruction. Every applied change is logged and read back, so what was intended can be compared with what was found. And where the agent cannot source a date, a price, a scope or a legal position, it leaves a visible placeholder for a person to fill in.
Some things stay human in an audit: the management assertion, the scope decision, accepting a risk, answering the auditor's judgement questions, signing, and deciding that evidence is complete.
Audit evidence
What it does well: reads a long evidence request and lists every separate ask in it, matches files to asks, drafts evidence descriptions, and spots a missing date, owner or period.
What it does badly: decides by itself that evidence is "sufficient", invents a control that does not exist, drops a part of a multi-part request, or treats a file name as proof of content.
A method that works: code collects and stamps the evidence with its source, time and period. The AI checks each ask against what was collected and writes covered, partly or missing. A person reviews before anything reaches the auditor, and sampling is done by a repeatable script with a fixed seed, never by the model.
Risk and control work
It drafts risk scenarios and treatment options, proposes mappings between frameworks, notices overdue reviews and stale versions, and summarises what changed in a new regulatory text. It scores likelihood and impact poorly without the business context, produces confident mappings that only look alike in wording, cannot give legal interpretation, and may not notice that a text was later amended. Keep the register as structured data; let the AI propose rows and mappings with a stated reason, and let the risk owner accept or reject each one. Use published mappings as the baseline and have the AI fill only the gaps, marked as proposed.
Process improvement
Agents turn meeting notes and mail into proposed tasks with an owner and due date, and write clear reminders. They are unreliable at deadline arithmetic and escalation rules done in their heads. So code handles due dates, overdue checks, reminder schedules and counts; the AI drafts and triages the unclear items. Every change to a task carries a comment saying what changed and why, and a capacity reserve keeps audit and urgent work first. For a new regulation, a readiness gap list is a good first draft, but it must be built on the legal text, not on a summary of it.
Awareness training
One core lesson can be adapted to many roles, with realistic scenarios and quiz variants. Facts of policy, tone and the real difficulty of a phishing lure need a person. A person owns the learning goals and approves each module; a subject expert checks the facts. The measure is behaviour: how many suspicious messages are reported, how fast, and how often the same people click within the same difficulty band, with completion kept only as a hygiene number.
The order to adopt it
- Read and remind: one view of tasks, registers and deadlines; code computes what is overdue; the AI drafts summaries and replies. Low risk and quick value, and it builds the data the later steps need.
- Audit evidence support: request parsing, matching of evidence to asks, descriptions and a per-request verdict, all reviewed before upload. It saves the most effort but needs the guardrails and logging from step one.
- Register and training work: proposals, document lifecycle checks, regulatory gap lists and role-based training. These need judgement, owner buy-in and good data, and mistakes change how the organisation sees its own risk.
Write access to any system comes last, one action type at a time, each with its own approval rule.