Segregation of Duties for SAP AI

Segregation of Duties (SoD), also called separation of duties, is one of the oldest and best-understood ideas in internal control. It is deeply familiar in SAP environments, where access governance is largely built around it. As AI agents begin to act inside the same business processes, the principle does not change — but it now has to be applied to non-human actors as well as people.

What Segregation of Duties is

SoD is an internal-control principle that prevents any single actor from controlling conflicting steps of a process. ISACA frames the core concept as ensuring that no employee or group is in a position both to perpetrate and to conceal errors or fraud in the normal course of their duties. In practice, business-critical duties are split across categories — typically authorisation, recording, custody of assets, and verification or reconciliation — so that the same actor does not hold an incompatible combination. The classic illustration is requiring two signatures on a cheque: the function is divided so that no one person can both initiate and approve the same transaction.

SoD in SAP access governance

Because SAP is a system of record for finance, procurement, supply chain and HR, SoD is central to how access is designed there. Roles are constructed so that incompatible activities — for example creating a vendor and approving a payment to it, or raising a purchase request and authorising the resulting order — cannot be performed by the same user. Where duties genuinely cannot be separated, compensating controls such as independent review, reconciliation and audit trails are put in place instead. This discipline is long-standing and well documented in the internal-control and audit literature.

Why AI agents must respect SoD

An AI agent acting in SAP can take or initiate actions inside these same processes. If it is not bound by the same SoD rules, it can become a path around them — combining steps that a human would be prevented from holding together, such as both recording and authorising a transaction on the same asset. The risk SoD exists to manage, that one actor can both cause and conceal an error or fraud, applies equally whether that actor is a person or an agent. So an agent's permitted actions must be assessed against the SoD model, not granted on convenience.

Mapping AI governance controls onto SoD

The two disciplines line up naturally:

• Treat each agent as an actor in the SoD analysis, classifying the duties its permitted actions represent (authorise, record, custody, verify) just as you would for a person.
• Constrain the actions an agent may take autonomously so it cannot hold an incompatible combination on the same asset or process.
• Where an agent must touch duties that would otherwise conflict, apply compensating controls — human authorisation of higher-impact steps, independent review, and a complete audit trail.
• Tie the agent's accountable owner to those duties, so there is a person answerable for the combination of actions the agent can perform.

Approached this way, AI governance does not replace SoD; it extends it to cover a new kind of actor, keeping the control intent intact as AI moves deeper into SAP processes.