Independence between the first and second lines can fail quietly when both use agents. If the same agent, the same sources or the same bounds serve the team doing the work and the team checking it, the check inherits the work's assumptions. Separation now has to cover agents, sources and configuration, not just reporting lines.
Why does separation matter in the first place?
The second line of defence exists to give an independent view of whether the business is managing its risks and meeting its obligations. Its value depends entirely on that independence. A check performed by people who share the first line's assumptions, data and incentives tends to confirm the first line's conclusions, whatever they are.
Firms have spent years building that independence into structures: separate reporting lines, separate budgets, separate access rights. Agents introduce a new way to undermine it, one that does not show up on an organisation chart.
How do agents erode the separation?
In several ways, most of them unintended and most of them driven by efficiency.
A shared agent. The same agent drafts customer communications for the first line and reviews them for the second line. It is unlikely to find fault with its own reasoning.
A shared brief. Two agents exist, but the second line's was cloned from the first line's and shares most of its instructions. The reviewer was told to look at the work the same way the worker was told to do it.
Shared sources. Both agents read the same product sheets and the same interpretation of policy. If the first line's source contains an error, the second line's check will treat the error as the standard.
A shared configuration owner. One platform team configures both lines' agents. A single change can alter how the work is done and how it is checked at the same time.
A shared model. Both agents run on the same model and may share blind spots, missing the same kind of problem in the same way.
None of these appears in a governance diagram. All of them weaken the second line's independence.
What has to be separate?
Not everything. Some sharing is sensible and some is unavoidable. The test is whether the shared element could cause the second line to miss a problem the first line created.
| Element | Should it be separate? | Why |
|---|---|---|
| The agent itself | Yes | An agent cannot independently review its own work |
| The brief | Yes, written by the second line | The checking perspective must not be inherited from the doing perspective |
| Interpretive sources | Yes where possible | The second line should check against the rule, not the first line's reading of it |
| Sealed systems of record | Can be shared | The ledger or policy library is the common truth both lines check against |
| Configuration ownership | Yes | Changes to the second line's agents should need the second line's approval |
| The underlying model | Ideally different; if shared, mitigate | Shared models can share blind spots |
| Bounds and limits | Separate sets | The second line's agents should have read access the first line does not need, and no authority to change first-line work |
The second-line separation checklist for compliance agents works through each of these in more detail.
Is a shared model acceptable?
This is the hardest question and deserves a straight answer. Running separate models for each line is costly and sometimes impractical. Sharing a model is often acceptable if everything above it is separate: a different brief written by the second line, different interpretive sources, different bounds and separate certification on the second line's own cases.
The residual risk is correlated blind spots: a model that consistently misreads a certain kind of document will misread it in both lines. Mitigate that by including in the second line's certification set cases specifically designed to catch the first line's typical errors, and by having people review a sample of the second line's conclusions. If the second line keeps agreeing with the first line on cases where people disagree, the shared model is the first place to look.
Who may change the second line's agents?
Only the second line, or a function independent of the first line acting on its approval. This includes the brief, the sources, the bounds and the certification set. If a first-line team, or a platform team that reports into the business, can alter how the second line's agents work, the second line is not independent, whatever the organisation chart says.
The same rule applies to the records the second line keeps. A breach log, for example, should not be editable by the first line. The question of whether an AI agent can keep a compliance breach log has a yes answer, provided the log belongs to the second line and is append-only.
What evidence shows separation is real?
A supervisor or internal auditor will want to see it, not be told it. Useful evidence includes a register of each line's agents with their owners, a record of every change to second-line agents and who approved it, proof that second-line briefs and sources were authored independently, and certification results for second-line agents on their own case sets.
It also helps to show disagreements. A second line that never disagrees with the first is either reviewing an exceptionally good business or not reviewing independently. Recorded disagreements, and how they were resolved, are some of the strongest evidence of real independence.
Where Pryme fits
In Pryme, each agent has its own brief, sources, bounds and certificate, scoped to a role. Second-line agents can be given read access across first-line records without any authority to change them, and changes to their configuration require approval from their own owner. This supports the role set out in AI compliance agents: second line, never the last word.
Questions people ask
Can both lines use the same agent platform?
Yes. The platform is infrastructure, like a shared email system. Independence depends on what sits on top: separate agents, briefs, sources, bounds and owners.
Should the third line also have its own agents?
If internal audit uses agents, the same principle applies with more force. Audit's agents should be independent of both the first and second lines.
Does this double the cost of agents?
It adds cost, mainly in writing separate briefs and certifying separately. That cost is the price of keeping a second line that means something.