Summarising a new rule is the easy part of regulatory change, and agents do it well. The real work is deciding what the rule means for your firm: which products, processes, policies and controls it touches, what has to change and by when. That interpretation is a judgement a named person must own, with the agent preparing the ground.
What can an agent do well in regulatory change?
Quite a lot of the volume. Regulatory publications arrive constantly: final rules, consultations, guidance, speeches, supervisory statements, enforcement notices. Keeping up is a reading problem, and reading is something agents do quickly and without tiring.
Useful tasks include watching the sources the firm has chosen to track, summarising each publication in plain language, comparing a new version against the previous one to show exactly what changed, tagging the publication by topic and jurisdiction, and matching it against the firm's policy library to suggest which internal documents might be affected. An agent can also draft the first entry in a regulatory change log so nothing arrives unrecorded.
These are depth 1 and depth 2 tasks: reading, and drafting for a person. They save the compliance team hours of triage. The question of whether an AI agent can track regulatory change is answered yes, as long as tracking is what you mean.
Why is interpretation the real job?
Because the same rule means different things to different firms. A new requirement on how consumer credit affordability is assessed may reshape a lender's whole underwriting process and barely touch a firm that offers credit only to businesses. A change to outsourcing expectations may be a large project for a firm that relies on many third parties and a paperwork update for one that relies on few.
Deciding which is which needs knowledge the regulator's text does not contain: what the firm does, how it does it, what controls already exist, what its risk appetite is and how its supervisor has engaged with it before. It also needs judgement about proportionality and timing. Those are decisions, and they carry accountability.
Regulatory change management as a discipline exists because reading was never the hard part. The hard part was always turning text into owned actions.
Where do agent summaries mislead?
Summaries are compressed, and compression loses things that matter in regulation. The common traps are worth knowing:
- Status. A consultation summarised in the same tone as a final rule can prompt work on proposals that may change.
- Scope. Rules often apply to some firm types, products or customers and not others. A summary that drops the scope clause makes everything look relevant.
- Transitional provisions. Dates and phased requirements are frequently in schedules or annexes a summary skips.
- Guidance versus rule. Whether something is a binding requirement or an expectation changes how a firm must respond.
- Jurisdiction. A publication from one regulator may not apply to a branch or subsidiary supervised by another.
An agent can be briefed to state each of these explicitly in every summary, and certified on whether it does. That is a better control than hoping readers notice what is missing.
How should the work be split?
The split follows the line between preparing and deciding.
| Stage | The agent | Depth | Stays with a person |
|---|---|---|---|
| Monitor chosen sources | Detects new publications | 1 | Choosing which sources to track |
| Summarise and compare | Plain summary, changes against prior version, status and scope stated | 1 | Checking summaries on a sample |
| Log the change | Drafts the log entry | 2 | Accepting the entry |
| Map to policies and controls | Suggests affected documents and processes | 2 | Confirming the mapping |
| Assess impact | Drafts an impact assessment | 2 | Deciding applicability and impact |
| Plan and assign actions | Drafts an action plan | 2 | Approving actions, owners and dates |
| Update policies | Drafts revised wording | 2 | Approving and publishing the new version |
The middle rows are where firms are most tempted to let the agent decide. An agent-drafted regulatory change impact assessment is a helpful starting point. Signing it is a different act. The method for doing it well is covered in how to assess the impact of regulatory change with an AI agent.
What should the record show?
For each change: what was published, when the firm became aware of it, the agent's summary and mapping, the person's assessment and the reasoning, the actions decided with owners and dates, and the evidence each action was completed. The record should make clear which parts the agent prepared and which a person decided.
That separation matters when a supervisor asks how the firm responded to a rule. "The agent summarised it" is not an answer. "The agent summarised it, our head of compliance assessed it as applicable to two products for these reasons, and these three changes were made" is.
What should you measure?
Measure the parts that show whether the agent's reading is reliable: how often people correct its status and scope statements, how often its suggested mapping misses an affected policy, and how long publications wait between the agent's summary and a person's assessment. The last figure is often the most revealing. If the agent has made triage fast but assessments still queue for weeks, the bottleneck was never reading.
Where Pryme fits
In Pryme, a regulatory change agent can be certified at depth 2 on the firm's past changes, scored on whether summaries state status, scope and dates correctly. The firm's policy library is a sealed source, so suggested mappings point to the version in force. Impact decisions are held for a named person and filed in the verdict record, in line with the view of AI compliance agents as second line, never the last word.
Questions people ask
Can the agent decide that a rule does not apply to us?
It can draft that conclusion with reasons. A person should confirm it, because a wrong "not applicable" is one of the costliest errors in regulatory change and the hardest to spot later.
How do we stop the agent missing publications?
Define the sources explicitly, check periodically that each one is still being read, and record a nil return when a source has nothing new. Silence and failure look the same unless both are logged.
Should the agent read regulator websites directly?
It can, but content fetched from a website has external provenance. Where possible, confirm important texts against the official published version before relying on them for an impact decision.