AI governance

Automation collapses segregation of duties unless you design against it

A control framework separates who creates a vendor from who pays one, because concentrating both is how money leaves quietly. Automation reconcentrates those duties by default — one agent doing four jobs a policy manual spent a decade splitting apart — and almost nobody is checking.

Review your SoD

Tell us your size and current controls. We will map where automation would create a conflict.

1 / 3
Agents are actors, not toolsSame matrix as peopleNo credential inheritance
DutyWho mayConflicts with
Create a vendorAP clerk · AP agentCannot also release payment
Code and post a billAP agent · AP clerkCannot also approve above threshold
Approve a billController · dept headCannot also create the vendor
Release a paymentCFO · controllerHuman only, at any amount
Prepare an accrualController agentCannot also post it
Post an adjusting entryControllerCannot also sign the close
Sign off the closeCFOHuman only, at any level
Change authority levelsAdministratorCannot also transact
Agents are subject to the same matrix as people. An agent holding two conflicting duties is the same control failure as a person holding them.

Six ways it breaks

Most of these are invisible until an audit.

One agent, many duties

The obvious failure. An agent that creates vendors, codes bills, approves them, and schedules payment holds four duties a control framework deliberately separates.

Agents acting as a user

Where an agent inherits the credentials of whoever triggered it, the audit trail shows a person doing something software did — and every SoD assertion built on that trail is wrong.

Policy written by the actor

If the person who grants an agent authority also writes the rules defining what that authority permits, they hold an unreviewed ability to automate postings.

Automation invisible in the log

If agent actions are logged separately from human ones, or not at all, an auditor cannot test the population and has to assume the worst about it.

Escalation by retry

An agent that can re-attempt a rejected action under different framing has an effective authority higher than its stated one. Ours cannot retry a policy rejection.

Self-granted scope

Any path by which an agent modifies its own permissions collapses the whole model, which is why that capability is absent rather than merely restricted.

Agents are actors, not tools

The single decision that determines whether segregation survives automation is whether an agent has its own identity. Treat it as a tool that a person wields, and it inherits that person’s permissions and appears in the log as them — at which point your audit trail is describing something that did not happen.

Treat it as an actor, and it holds its own identity, its own permission set, and its own entries in the same trail as everyone else. Every conflict rule that applies to a person then applies to it without anyone writing a second framework.

If your audit trail says a person coded four hundred invoices on a Tuesday, either they had a remarkable day or your agents are wearing their credentials.

The conflicts we enforce by construction

  • Vendor creation and payment release. No actor, human or agent, can hold both. This is the classic fraud path and it is enforced rather than documented.
  • Preparation and posting of the same entry. The Controller Agent prepares accruals and cannot post them, which is why we advise against raising it to Level 2.
  • Transacting and authority administration. Whoever grants authority levels cannot also operate as a transacting user in the same entity.
  • Policy authorship and authority grant. Controllers write policies; administrators grant levels. Different permissions, usually different people.
  • Posting and sign-off. The person who posts adjusting entries is not the person who signs the close.

What your auditor should ask us

Four questions, and we would rather you ask them than take this page on trust. Does each agent have its own identity, or does it act as the triggering user? Can an agent hold two conflicting duties under any configuration? Can an agent modify its own permissions or retry a rejected action? And can you produce, for any period, the complete list of actions taken by automation and the authority under which each was permitted?

The last one is the practical test. A vendor who cannot produce that list has not thought about this, whatever their documentation says.

This is getting asked more

Audit firms are becoming specific about automated activity inside financial systems. The questions arriving now are what the software was permitted to do, whether that changed during the period, and how you would detect it acting outside scope — all of which are answerable here and unanswerable in most systems.

Questions

What auditors ask.

Do agents act as the user who triggered them?
No. Each agent holds its own identity and permission set, and its actions are logged as its own. Credential inheritance would make every SoD assertion built on the trail unreliable.
Can one agent hold conflicting duties?
No configuration permits it. The conflict matrix applies to agents identically to people, and a grant that would create a conflict is rejected when you attempt it.
Can an agent retry a rejected action?
No. A policy rejection is final for that proposal. An agent able to re-attempt under different framing would have an effective authority higher than its stated one.
How do we evidence this to an auditor?
Export the authority matrix as configured for each month of the period, the change log for every grant, and the population of agent actions with the policy version that permitted each. Those three together are what a controls test needs.
What if we are too small for formal SoD?
Many companies at this size genuinely cannot separate every duty across three people. The honest answer is compensating controls plus a complete trail — and automation actually helps, because an agent holding one duty is easier to constrain than a person holding four out of necessity.

Have your auditor test this.

We will walk your audit team through the conflict matrix and the automation record, without a salesperson present.