The higher the stakes, the less of it is automated.

Almost every conversation about AI in planning opens on the same question, and it is the wrong one. What gets asked is which model, which platform, which vendor. What decides the outcome is which decisions you are willing to let something else make, and on what evidence you would be willing to let it.

Those are different questions, because automation is not a capability you buy. It is a capability you point. A planning function that knows who owns the demand number, where a trade-off gets settled and what the escalation path looks like can hand a great deal of its work to a machine safely. A planning function without those things can hand over exactly the same work and get the wrong answer faster, more consistently, and with more confidence attached to it than any person would have had.

So the useful framing is not how intelligent the system is. It is how much of the decision it has been handed, and whether anybody would notice if it were wrong.

Planning decisions are not one kind of thing. They run from reading a record in the ERP to committing the business to a capacity investment, and those two have almost nothing in common except the word decision.

Automation reaches all of them, and this is the part usually drawn wrong. Diagrams of this tend to taper toward the top, implying the senior decisions are somehow out of reach. They are not. An executive call is supported by automation as thoroughly as data cleansing is, and often more valuably, because the analysis behind a capacity decision is precisely the kind of work a machine does faster and more consistently than a committee. What changes as you go up is not whether automation helps. It is how much of the decision it is allowed to hold.

Read the two tracks rather than the rows. Automation support widens toward the data at the bottom, where the volume is and the rules hold. Human judgement widens toward the decisions at the top, where the stakes are and the rules run out. Every row is full width, because every row gets help.

Neither track is numbered. A percentage of automation per level does not exist anywhere in a real business, and putting ticks on that axis would be the first invented claim on the page.

Stripped of the category noise, the work divides into a small number of jobs. Some of them are arithmetic done faster than a person can do it. Some of them are attention, which is the scarcer thing.

The first group is generation: forecasts produced, demand read from signals closer to the customer than a shipment record, inventory positions optimised against a service target. This is the work most people mean by AI in planning, and it is the part where the technology is furthest ahead of the average deployment.

The second group is attention, and it is the one that changes a planner day to day. Risk detection, exception management and scenario simulation are all the same underlying job: looking at more than a person can look at, and saying which of it matters. A planner who spends the morning finding the twelve things that moved is not doing planning. Handing that search to a machine does not replace the planner, it returns the part of the role that needed them in the first place.

The third is recommendation, and this is where the care belongs. A recommended decision is still a decision, and the question of who owns it does not disappear because a system proposed it.

In the Planning Excellence Framework automation sits last, underneath nothing and on top of everything. That ordering is the argument, not a sequencing convenience.

Automation multiplies whatever it is pointed at. Pointed at a process with clear ownership, a settled cadence and a decision framework people actually apply, it multiplies all of that and the return is real. Pointed at a process nobody follows, it multiplies that instead: the workaround becomes the system of record, the undocumented override becomes a rule, and the thing that was quietly wrong becomes quickly and repeatably wrong.

This is why digital enablement sits sixth and automation seventh rather than either coming first. Instrumentation applied to a process that exists makes it faster. Instrumentation applied to a process that does not exist digitises the gap.

When the capability underneath is real, this is the shape of what goes on top of it. None of it is exotic and none of it is the point on its own.

The two that matter most are the least impressive to demonstrate. Exception intelligence decides what a planner sees first, which is the difference between a system that saves time and one that adds a screen to check. And decision governance decides how much any of it is allowed to do without asking, which is the difference between automation you can defend and automation you have to explain.

The question is never whether a planning function should be automated. It is how far along the dial each individual decision should sit: advise, recommend, act with review, act and report. Different decisions belong at different settings, and the setting should follow the cost of being wrong rather than the confidence of the model.

Two conditions make the dial safe to turn. The decision has a named owner even when a system is making it, so there is somebody accountable for the setting rather than for the individual call. And every action is auditable after the fact, because the first time an automated decision goes badly, the only thing that matters is whether you can reconstruct why it was made.

Get those two right and autonomy becomes a normal engineering decision. Get them wrong and every increase in autonomy is an increase in exposure that nobody has agreed to carry.