AI agents in S&OP should work like a new manager with written approval limits: they decide routine supply and demand changes within set thresholds and escalate everything else to a named person.
The software is ready to act. The question most companies have not answered is what it is allowed to decide.
Every autonomy setting is an approval limit in disguise
Vendor material for 2026 describes “touchless” S&OP, in which agents handle routine balancing on their own. One typical example has the system approve an overtime shift automatically when a demand spike stays below five per cent, with people receiving only a notification.
Ask who chose five per cent. That threshold decides overtime cost, which is a finance matter, and delivery performance, which is a promise the commercial team made. If the planning team set it while configuring the tool, an operational parameter now carries a policy no executive signed.
Every company already knows how to handle this for people. A delegation of authority matrix sets approval limits by amount and by type of decision. An agent is a new kind of approver, and it belongs in the same matrix. The principle is the one behind any well-designed role: define its decision rights before you appoint anyone to it.
Good planners already work to limits nobody wrote down
When I worked in supply chain functions, the best planners knew exactly when to ring the plant manager and when to move an order quietly themselves. None of that was in a procedure. It lived in experience, and it was usually right.
An agent has to work from something. Either somebody writes those limits down deliberately, or the agent infers them from planning history, including the habits nobody would choose to keep. The first option is a management decision. The second is an accident with good documentation.
An agent cannot be accountable. Someone has to own its limits.
The agent’s limits fit on four lines
The limits need not be long. They need to be written in money and service terms rather than technical ones:
- Approves alone: routine rebalancing where quantity, cost and service impact all stay inside agreed tolerances.
- Recommends: changes that cross a tolerance but stay within the agreed plan; a named planner approves them.
- Escalates: anything that touches customer commitments, margin or the agreed supply plan; it goes to the S&OP process owner.
- Never touches: prices, allocation between key customers, and anything the executive S&OP meeting has decided.
Take the overtime example again. Under these four lines, the agent may approve the extra shift if its cost stays within the weekly overtime budget and no customer delivery moves. If the shift would push the plant over budget, the agent recommends it and a planner decides. If the only alternative is a late delivery to a key account, the agent escalates.
Four lines like these turn a configuration screen into a management decision. They also tell auditors, finance and the plant who is accountable when the agent gets it wrong.
Autonomous planning is still mostly an aspiration
The cautious route is also the realistic one. BCG’s 2026 research on supply chain planning finds that fully autonomous planning remains an aspiration. Most of today’s value comes from forecasting, exception management and workflow automation. Companies that try to leap ahead with AI alone tend to struggle, while those that add it to stable planning foundations see more lasting gains.
That argues for starting narrow. Give the agent tight limits, watch what it does, and widen them as the evidence builds. That is how a sensible company treats a new manager in the first months.
The S&OP forum reviews the limits, not the agent
Once a month, the S&OP meeting should see one page on the agent’s work: what it approved, what it escalated, where planners overrode it, and what those decisions cost. On that evidence, the forum widens a limit, narrows one or leaves them as they are, and records the choice. Over a few months, that page becomes the evidence for widening the limits, or for taking them back.
It is the same discipline I ask for in my own contracts, where a written annex lists what I decide, recommend and escalate. Authority that is written down can be reviewed. Authority that lives in a configuration file cannot.
Switch the agent on after the limits are written, not before.