AI Agents in ERP
A copilot suggests; an agent acts. That one-word difference moves the conversation from productivity to accountability — and that is where executives come in.
What actually changes
Fact: The current generation of enterprise AI is shifting from assistants (summarise, draft, explain — a human executes) to agents (interpret a signal, decide the follow-up, execute it across systems). In ERP terms: not "explain this blocked invoice" but "resolve the block, post the correction, notify the owner".
Inference: Value concentrates where three conditions meet: high volume, clear rules with documented exceptions, and expensive delay. In practice that points at exception handling (blocked invoices, order holds, delivery deviations), reconciliation work, and coordination tasks that today exist only to pass information between functions.
Point of view: The right first agents are boring. Choose processes where a wrong action is cheap to detect and reverse, instrument everything, and expand scope with evidence. An agent portfolio built on spectacular demos and irreversible actions is a governance incident on a countdown.
Boundary conditions
- Every agent action must be explainable, auditable and reversible — or explicitly classified as not reversible and gated accordingly (see embedded AI governance).
- Agents inherit the quality of their context: fragmented master data produces confidently wrong actions (see shared business context & unified data).
- A clean core is the cheapest agent-enablement programme you can run — agents work against standard processes and released APIs, not against twenty years of custom code.
Monday-morning questions
- Which three exception queues consume the most skilled hours today?
- For each: what would a wrong automated action cost, and how fast would we notice?
- Who signs off an agent's scope expansion — and on what evidence?
Where this connects
The operating model is the Autonomous Enterprise; test your starting position with the readiness self-assessment. All tracks: Executive Clarity topic tracks.
Sources
- SAP's 2026 executive guide to the Autonomous Enterprise (vendor publication, read as market signal); agent-pattern observations from transformation practice.
Editorial standard: facts, inference and point of view are kept separate above. Published 2026-08-03 · By Andreas BORN.