The plan landed in a shared doc on a Tuesday afternoon: a hundred and forty lines, tasks nested three deep, durations, dependencies, named owners, a critical path that actually resolved. It had taken roughly ten minutes to produce. By any structural measure it was better than what the team would have built by hand — more complete, more internally consistent, and it caught two cross-workstream dependencies that a room of humans forgets every single time. Everyone read it. Nobody objected. We baselined it.
Six weeks later the program was behind, and when I went owner by owner to find out where, I kept getting a version of the same sentence: I didn't really look at my line. Not defiance, not laziness — these were good people who cared about the outcome. They had simply never been put in the position of saying a date out loud and having it held against them. The date had arrived pre-written and looking authoritative, and nobody argues with a document that looks like that.
The reflex diagnosis is that the plan was wrong. It wasn't. Every line was defensible and most of the durations turned out to be roughly right. The artifact was fine. The problem is that we had manufactured a plan without producing the thing that planning was actually for.
Planning was never about the plan
Ask anyone who has run a large program what the two-day planning workshop is for and they'll say it's to build the schedule. That's the cover story. The schedule is a by-product. Three other things happen in that room, and every one of them matters more than the document that comes out of it.
1. Commitments get extracted in public
When a workstream lead says "we can have the interface ready by the fourteenth" in front of eleven peers, something changes state. The date itself is a guess either way — that isn't the point. The point is that a specific human has attached their credibility to it in front of witnesses, and that's what makes them raise their hand early when it starts to slip. A date the machine assigned carries none of that weight. Nobody's reputation is standing on it, so nobody feels the tug.
2. Disagreement surfaces while it's still cheap
Half the value of a planning session is the argument in the middle of it: the moment two leads discover they have been assuming opposite things about a shared dependency. That disagreement exists whether or not you hold the session. The only question is whether you find it in week two, in a room, or in month four, in production — which is the same seam problem that breaks programs between teams, just caught earlier or later. An AI-drafted plan resolves the ambiguity silently: it picks one interpretation and writes it down cleanly. The conflict doesn't get settled. It gets buried under formatting.
3. Everyone leaves with the same model of why
A plan tells you the sequence. Sitting through the plan being built tells you why it's that sequence — which constraints are real, which are convention, and which one is load-bearing enough that moving it breaks four other things. That reasoning is what people draw on to make hundreds of small calls later without asking anyone. Hand someone a finished plan and you transfer the sequence and none of the reasoning, which is exactly how a lead optimises their own slice in a way that quietly wrecks somebody else's.
A plan is not a document. It's the residue of an argument a group of people had in a room. AI can produce the document in ten minutes — but the document was never the deliverable. The argument was.
How to tell you're carrying one
The tell I trust most is that the plan was accepted with no edits. Real plans get bruised — a date pushed, an owner swapped, a dependency contested, a line struck through with a reason attached. A hundred-and-forty-line plan that goes into baseline unmarked hasn't been reviewed. It's been received.
After that, the symptoms are consistent. Owners describe their own tasks using the plan's exact phrasing rather than their own, because they're reading rather than remembering. Nobody can explain why a given sequence is in that order, so nobody can tell you what happens if it changes. And slips show up in the status report instead of arriving as a heads-up from the person responsible — which is the expensive version, because by then the decision it forces is already sitting in a queue nobody is tracking.
Keep the draft. Keep the argument.
I'm not arguing for going back to whiteboards and sticky notes. The machine-built draft is genuinely better as a draft — more complete, faster to revise, and it doesn't get tired at hour six of a planning workshop. Use it. What has to move is the ritual. The machine drafts; the humans contest. The hour you saved on construction gets spent on the part that was always scarce, which is senior attention in the same room, disagreeing on the record.
The Army has a mechanism for exactly this, and it long predates anyone worrying about generated schedules. After an order is issued, subordinate leaders give a confirmation brief and then a backbrief — they restate, in their own words, what they understand their task to be and how they intend to accomplish it. The point is never to check whether they read the document. It's to make them build the plan in their own head and say it back, because a leader who can't articulate their piece unaided doesn't own it yet. That mechanism assumes the plan came from somewhere else. It works just as well when the somewhere else is a model.
In practice that means four things. Every owner backbriefs their slice with the document closed. Every workstream brings at least one contested line to the review, because a review that produces no friction didn't happen. Every assumption the model made gets logged as an assumption — with an owner and a review date — rather than sitting in the schedule dressed as a fact. And nothing gets baselined until the person who owns a date has said that date out loud, which is the same boundary as the one that decides who signs: you can delegate the drafting, you cannot delegate the commitment.
- Never baseline a plan its owners haven't spoken to. Silence in a plan review isn't agreement, it's an absence of data.
- Backbrief, don't circulate. Ask each owner to restate their slice and its constraints in their own words, with the document closed.
- Require one contested line per workstream. If nobody pushed back on anything, the review didn't happen.
- Log the model's assumptions as assumptions — with an owner and a review date — not as facts formatted like a schedule.
- The person who owns the date says the date, out loud, on the record. Attribution is what makes a slip get reported early instead of discovered late.
The bottom line
Generated plans are one of the clearest wins available on a program right now, and I'd rather start from one than from a blank page. But drafting capacity was never the constraint. The constraint was always getting a dozen busy, senior, mutually skeptical people to sit down and commit to the same version of the next six months — and that cost hasn't moved an inch. If you spend the time the machine gave you back on the argument you skipped, the draft is a gift. If you spend it getting to the next meeting faster, you've acquired a beautiful document and quietly traded away the only thing that ever made it a plan.