Segregation of duties is a control that divides a sensitive process between different people so that no single party can both carry out a transaction and conceal an error or fraud in it. The classic split separates four roles: authorising a transaction, recording it, holding the asset, and reconciling the records. Each role checks the others.
Why does segregation of duties matter once agents act?
An AI agent is a participant in the process, so it holds duties like anyone else. If one agent can create a supplier, code that supplier's invoice and schedule its payment, the split that was designed for people has collapsed into a single system. The agent does not need to be dishonest for that to matter. A manipulated input or a plain error now passes through every stage unchallenged.
There is a second, less obvious conflict. The person who writes an agent's brief and sets its bounds holds a kind of duty too. If that same person approves the agent's held actions, they are effectively both designing and authorising the work. Segregation has to cover who configures agents, not only what agents do, which is why it sits at the centre of AI agents for finance.
How does it work in practice?
Map every duty in the process to a party, then look for incompatible pairs held by the same one. Common incompatible pairs:
- Creating or changing a supplier, and approving payments to suppliers.
- Preparing a journal, and approving the same journal.
- Recording receipts, and reconciling the bank account.
- Setting an agent's bounds, and approving that agent's held actions.
Agents count as parties in the map. A segregation of duties matrix for finance agents gives a layout, and a general version for agents outside finance is the separation of duties matrix for AI agents.
What does that look like in an example?
A housing association uses one agent for accounts payable. It reads invoices, matches them and prepares the payment file. The agent also had permission to update supplier records so it could fix typos in addresses. During a review, the finance manager notices that permission included bank details. In principle, the agent could have changed a supplier's account and then paid it. The fix splits the capability: the agent may draft supplier changes, but a person in a different team confirms bank detail changes, and the payment run is approved by a third person.
How is segregation of duties different from the four-eyes principle?
The four-eyes principle requires two people on a single action: one prepares, one approves. Segregation of duties splits a whole process across people so that different stages are held by different hands. Four-eyes is one way to meet segregation at a given step, but a process can have four-eyes on every step and still fail segregation if the same two people do every step together. With agents, a person approving everything the agent drafts is four-eyes; making sure that person is not also the one who set the agent's bounds is segregation.
Which related terms should you know?
- Incompatible duties: pairs of tasks that should never sit with one party.
- Compensating control: a substitute check, such as a detailed review, used where duties cannot be split in a small team.
- Named approver: the identified person who signs off a held action.
The specific case of journals is covered in segregation of duties when an agent posts journals.
Questions people ask
Does an AI agent count as a separate person for segregation of duties?
For mapping, yes: it holds duties and they must be compatible. For accountability, no: an agent cannot answer for a control. Every duty an agent holds needs a named person who owns it, and that person's other duties must also be compatible.
How do small finance teams segregate duties with agents?
Small teams often cannot split every duty between people. Agents can help by taking the preparation duties, leaving people free to approve and reconcile. Where one person must still hold two duties, add a compensating review by someone outside finance, such as a director checking the payment run.