The Strategy Nobody Decided
A strategy that cannot answer is undecided, and no amount of fluency, human or synthetic, will make it deliverable.
The Tuesday the Document Arrived
It landed mid-morning: a forty-two-page transformation strategy, generated overnight from a folder of board minutes and a one-line prompt, and by every surface measure it was the best strategy document the programme had ever received. The prose was even. The logic connected. Every section a strategy is supposed to have was present and correctly labelled. The delivery leads read it in the meeting room, and for a few minutes the mood was something close to relief.
Then someone asked the only question that ever matters on the delivery side. The document named three priorities. If the quarter goes badly and two of them can be funded but not the third, which one gives?
The room went quiet, and it was the particular quiet of people realising the document could not answer — because no one in it had ever been made to choose. It listed three priorities the way a menu lists three desserts: as options, generously, without the faintest suggestion that you cannot have all of them. A human author, forced to write that sentence at midnight, would have felt the cost of the third priority in their own hands and either cut it or defended it. The machine felt nothing, so it kept everything.
That meeting is the whole argument of this piece, and I want to make the case plainly. The danger of the generated strategy document is not that it is fake. It is that it is undecided. And the people who discover this first, always, are the ones downstream who have to turn the document into work.
What a strategy document was for, before it was cheap
We have spent two years now — since the tools became good enough to trust with a first draft — talking about authenticity as though it were a question of style. Can you tell it was written by a machine? Does it have that flat, faintly over-fluent register, the slop we have all learned to recognise? This is the wrong question, and the delivery side has always known it is the wrong question, because a strategy document was never primarily a piece of writing.
It was a record of arguments that had already happened.
When a leadership team wrote a strategy the slow way, the document was the visible residue of an invisible process: the priorities that survived the meeting where three others were killed, the number someone had to stand behind, the sequence that reflected a real judgement about what had to come first. The delivery team could not see the arguments, but they could feel their shape in the document — the way you can feel, in a well-made joint, that someone measured twice. A strategy that had been fought over had a certain economy to it. It committed to fewer things, because commitment is subtractive, and the writing was where the subtraction happened.
A strategy document was never the strategy. It was the scar tissue left by the act of deciding — and delivery teams read the scars, not the prose.
This is why the delivery side developed, over years, an almost forensic skill: reading a strategy not for what it said but for what it had cost. A document with seven equal priorities told you the leadership had not yet had the fight. A document that went quiet exactly where the hard trade-off lived told you where the programme would later tear. The prose was almost beside the point. What mattered was the pattern of pressure underneath it.
What generation removes
Here is the mechanism, and it is worth being precise, because the loss is not obvious and it is not the loss most people worry about.
Generation does not remove intelligence from the document. If anything it adds a kind of intelligence — the models are genuinely good at completeness, at structure, at anticipating the section you forgot. What generation removes is friction, and friction was doing far more work than we admitted.
The old way was expensive in a specific direction: it made comprehensiveness costly and commitment cheap. Writing a fourth priority by hand, after the first three, is painful enough that you ask whether you really need it. So people cut. The physical cost of the writing acted as a governor on scope, and the scarcity of effort forced a hierarchy. Generation inverts the economics exactly. It makes comprehensiveness free and commitment expensive — because to commit, in a generated document, you now have to go back in and deliberately delete the reasonable-sounding alternatives the machine has helpfully supplied. Almost nobody does. Deleting a fluent, plausible paragraph that took no effort to produce feels like vandalism. So the priorities multiply, the caveats accrete, and the document swells toward the shape of everything-considered — which is the precise opposite of strategy.
I have watched a delivery team try to plan against one of these documents, and the failure mode is consistent. They cannot find the load-bearing wall. In a hand-made strategy there is usually one sentence the whole thing rests on, and you build the plan around it. In the generated version every sentence is equally weighted and equally reasonable, so the team ends up doing the deciding themselves, six months downstream, with none of the authority and all of the risk. The strategy did not get written by a machine. It got deferred by a machine — onto the people least equipped to carry it, and least mandated to.
The strongest version of the other side
I want to take the opposing case seriously, because it is not weak.
The honest objection runs like this. The strategy document was always theatre. Most were never fought over; they were assembled the night before the board meeting, padded to look thorough, and the fabled “record of real arguments” is nostalgia for a rigour that rarely existed. If the document was already performance, then a machine that produces better performance faster is a straightforward gain. And there is a second, sharper point: the tools do not only generate, they interrogate. A well-used model will pressure-test a strategy, surface the trade-off a tired executive missed, play the sceptic at two in the morning when no colleague is awake to. Used well, generation could raise the floor on rigour, not lower it.
Both points are true, and neither rescues the generated document, for the same underlying reason.
The value was never in the artefact. It was in the deciding, and the artefact was merely the place the deciding was forced to happen. Yes, plenty of old strategy documents were theatre — but the good ones were load-bearing precisely because writing them made someone choose. Remove the friction and you do not only lose the bad theatre; you lose the mechanism that occasionally forced the real thing. And yes, a model can interrogate a strategy — but interrogation you are free to decline costs nothing, and a challenge with no consequence changes no decision. The executive who lets the machine play sceptic and then accepts all of its balanced options has not been through the fight. They have watched a re-enactment of one.
“A challenge you can dismiss with a keystroke is not a challenge. It is a rehearsal — and the delivery team can always tell the difference between a decision and a rehearsal of one.”
Authenticity is the wrong test; decidedness is the right one
So let me put down the question we keep asking — did a machine write this? — because on the delivery side we do not care, and never did. A strategy typed by a human who decided nothing is worthless. A strategy generated by a machine and then genuinely owned, cut, and committed to by someone who will answer for it is sound. The signature is irrelevant. The commitment is everything.
The test that actually matters is whether the document can survive contact with a single question, asked in good faith: what did you decide not to do? A strategy that can answer — that names the roads not taken and the cost of not taking them — has a person behind it, whatever tool held the pen. A strategy that cannot answer is undecided, and no amount of fluency, human or synthetic, will make it deliverable.
That gives the delivery side something concrete to demand, and I think we should demand it without embarrassment:
- Ask, of any strategy, what it explicitly rules out — and treat the absence of a clear answer as an unfinished document, not a finished one.
- Ask who will answer for the central number, by name, in the room — not who wrote the paragraph it sits in.
- Treat comprehensiveness as a warning sign rather than a virtue: the more a strategy covers, the less likely it is that anyone has chosen.
None of this is anti-tool. I draft with these models daily, and they are extraordinary. The point is narrower, and I suspect permanent: generation collapses the cost of producing the document, and in doing so it quietly removes the one thing the document was ever really for. If we want the deciding back, we now have to add the friction deliberately — to make someone sit with the generated draft and cut it until it commits — because the process will no longer impose that discipline for free.
The Tuesday, revisited
The team in that meeting did, in the end, get their answer. They sent the document back with one instruction: pick the priority that gives. It took the leadership a fortnight, two genuinely uncomfortable conversations, and the deletion of eleven pages. What came back was shorter, plainer, and in places less articulate than the machine’s version — and it was the first strategy that programme had received in years that the delivery team could actually build against.
That is the whole of it. The machine can write the document. It cannot do the deciding, and it never could — because deciding was never a writing problem. It was only ever the writing that fooled us into thinking it was.