Why Your AI-Written Strategy Isn't a Strategy (And Your Team Already Knows)

I am aggressively pro-AI.

I’ll argue with anyone that most SMEs are too slow, too cautious, and too over-planned in how they adopt it. My standing view on digitalisation is simple: find the enablers in your team, let them tinker, and accept that nine out of ten experiments will fail because the tenth will change everything. The leaders who win over the next decade are the ones who let their people break things with these tools, not the ones still waiting for a perfect rollout plan from a consultancy. Most of the advice I give comes down to the same sentence: run more experiments, sooner, with fewer approvals.

Now here’s the part I haven’t said loudly enough.

There is exactly one place I watch this approach destroy value instead of create it. It’s when SME leadership stops using AI as a tool and starts using it as a leader. When strategy, communication, and direction quietly get outsourced to a text box. The damage isn’t loud — it’s slow, invisible, and very hard to reverse once your team has seen it happen.

It doesn’t look reckless from the inside. It looks productive. The leader is shipping comms faster, producing more documents, sounding more “strategic” on paper. The calendar looks healthier. The inbox moves. What’s breaking underneath is the part nobody tracks on a dashboard — the trust of the team, the clarity of direction, the conviction of the voice at the top. By the time those show up in performance, they are very expensive to rebuild.

The biggest mistake an SME leader can make right now is to delegate leadership to ChatGPT.

Not because AI is dangerous. Because of what AI is — and isn’t — by design.

AI doesn’t give you the best answer. It gives you the most probable one.

Those are not the same thing. They are almost never the same thing.

Picture ten voices in a room arguing about a hard decision. One of them has the perfect answer. Eight of them have something good and something bad mixed into their take. The last one is right for a reason nobody else has seen yet. An LLM trained on that room will not pick the perfect voice, and it will not pick the outlier. It will produce something that sounds like all ten of them averaged together — some good, some bad, edges smoothed off, conviction sanded down. That’s not a flaw to be patched in the next model release. That is the mechanism itself: stochastic pattern-matching over the most probable next word. The median of opinion is mediocrity. That’s what the machine is trained to produce.

Now look at what a strategy is. A good strategy is almost always the outlier decision — the move the market isn’t already making, the one most competitors would reject, the bet everyone else thinks is crazy until the quarter it isn’t. That’s literally where competitive advantage lives. You don’t beat your competitors by picking the averaged-out opinion of ten people in a room. You beat them by seeing something nobody else sees and betting on it with conviction.

This is also why the “1-in-10” framing works in product experimentation but breaks the moment you try to apply it to strategy-by-ChatGPT. When you run ten experiments, one wins big and the other nine are cheap lessons. When you ask ChatGPT for a strategy ten times, all ten answers live in the same averaged-out zip code. You are not sampling ten different bets. You are sampling ten rephrasings of the same safe middle. The variance you need for a real 1-in-10 isn’t there, because the generator was built to suppress exactly that variance.

So the moment you ask ChatGPT what your strategy should be, you are asking a machine engineered to average ten voices to produce the one voice no one in your market has heard yet. It cannot. It is not built for that. It is built for the opposite of that.

Remember this every time you hit send on a strategy prompt. The machine is doing exactly what it was designed to do. The problem is what you asked it for.

Strategy is not a document. It’s a set of decisions made in context.

The context is everything ChatGPT doesn’t have.

It doesn’t know your three biggest customers by name, or which of their edge cases are about to eat a quarter of your roadmap. It doesn’t know that your strongest engineer is burning out quietly, or that your new VP of sales won’t survive another bad quarter politically. It doesn’t know which team member will sabotage a plan they weren’t consulted on, or which one you need to win over early for the plan to stick. It doesn’t know what cash looks like at the bottom of your P&L, what morale actually feels like at Monday stand-ups, or what your technical debt will cost you the moment you try to move fast in the wrong direction.

That’s the visible context. The invisible context is worse.

It doesn’t evaluate the cost of the change a strategy triggers, hidden or otherwise. It doesn’t weigh “this pivot looks smart on paper, but it will consume six months of engineering time, destroy a product line our largest customer depends on, and the two people who understand that product line will leave over it.” It doesn’t weigh what happens to morale when you reorg the team three months after the last reorg. It doesn’t weigh the political cost of going around your COO, or the opportunity cost of saying yes to this market and no to the one you were actually winning. Those are the calculations strategy is made of — and ChatGPT doesn’t do them. Not because it’s a bad model. Because it does not do estimates at all. It does language that sounds like estimates.

There is a difference between those two things, and it is the difference between a strategy and a paper napkin.

This is also why “ChatGPT wrote our strategy” documents almost always read like they could belong to any company. They could. They were written from outside the company, about no specific customers, with no specific constraints, by a process that had no access to the one thing strategy actually lives on.

A good advisor tells you what you don’t want to hear. ChatGPT is architecturally incapable of that.

This isn’t a personality flaw. It’s a design property.

A peer-reviewed 2026 study in Science (readable summary) tested ChatGPT, Claude, Gemini, and Llama against scenarios where the accurate answer conflicted with what the user wanted to hear. All four consistently sided with the user. Worse, participants rated those sycophantic answers as more helpful and more trustworthy — even when the model was steering them wrong.

The mechanism is simple to observe. You tell it no. You tell it that doesn’t work. You tell it you disagree. It softens. It concedes. It reframes. Over a few iterations of negative feedback, it drifts toward whatever position lets the conversation end on a pleasant note. The technical term is sycophancy. The functional description is simpler: an advisor who eventually tells you what you want to hear.

An advisor who eventually tells you what you want to hear is not an advisor. It’s a mirror with better vocabulary.

You can watch this happen in your own chat window. Ask ChatGPT to critique a strategy you already believe in. Push back twice on anything it says against your position. Watch the third answer. You will see your own conviction reflected back at you, polished, with a few extra bullet points. You did not get challenged. You got validated. And you paid for that validation by mistaking it for advice.

Now imagine this loop running for six months across every important decision a leader makes. Each of those conversations ends with the leader feeling sharper, more confident, more right. None of them involved anyone who would actually push back. The leader is not getting smarter — they are getting more certain, which is a very different thing, and a much more dangerous one.

There’s a second problem stacked on top of that one, and it matters at least as much. Leadership is multi-stakeholder by definition. A good consultant doesn’t brief the CEO and walk out. They work the organisation. They pressure-test the plan against the COO. They flag the political landmine with the VP of product before it blows up. They get the skeptical senior engineer on board in a 1:1 before the all-hands. They care whether the strategy can actually land in a room full of humans who have to execute it — because a strategy that doesn’t land is a PowerPoint, not a strategy.

ChatGPT does none of this. Not because it’s lazy. Because it has one user in every conversation, and it is optimised to please that one user. The architecture is one-to-one. Leadership is many-to-many. The gap between those two things is where companies fall apart.

So when an SME leader asks ChatGPT “what should we do,” they get a machine that is both too agreeable to actually disagree with them and too context-free to understand the people the strategy has to survive.

That is not an advisor. That is a tool pretending to be one.

And the pretending is where the real damage starts.

Once you know the pattern, you can see it everywhere.

The tells aren’t subtle. They’re just rarely named out loud.

The communication structure that becomes an organisational chart. A leader asks ChatGPT for help on a communication plan, and what comes back is a flat pyramid with titles. No information flow. No meeting cadence. No actual problem solved. Shape without substance — because the tool returned the most probable artifact associated with the phrase “communication structure,” not the one that would fix the real communication breakdown in the room.

The voice-output conflict. What the leader says in meetings is sharp, specific, and unmistakably theirs. What comes out in writing is smoother, flatter, and could belong to anyone. Their team notices this before they can put words to it. Something is off. They don’t know exactly what. But the trust delta shows up in small ways — fewer questions in all-hands, more eye-rolls in side channels, a quiet deprioritisation of “strategic updates” in everyone’s inbox.

The strategy document that could belong to any company. No named customers. No specific constraints. No tradeoffs stated out loud. Three to five goals that sound reasonable but commit to nothing. The document reads like it was written by someone who has never met the company — because it was.

The framework name-drops without conviction. OKRs. Jobs-to-be-done. North-star metrics. Flywheels. All the right words are in the deck. The underlying thinking isn’t. The deck sounds like it knows what it’s doing, but the decisions it’s supposed to drive never quite get made.

The trust drip. Nobody sends a Slack message saying “I think our CEO is running the company through ChatGPT.” But employees notice the flatness, register the voice conflicts, internalise the generics — and something quiet changes in how they listen in meetings. A CEO steered by consultants is already a bad look. A CEO steered by AI is worse. At least a consultant has a reputation, a contract, skin in the game, and a face to trust or distrust. AI has none of those. So when a team figures out where the voice is actually coming from, there is nobody to correct, nobody to replace, and nobody to trust less. The trust just quietly leaves.

That last one is the most expensive. It is also the hardest to see coming.

The fix: treat ChatGPT like a Junior Assistant.

Here’s the mental model that puts everything back in place.

A junior assistant is useful. They’re fast. They’re cheap. They occasionally produce something genuinely sharp. But you also know — from the second they walk into the role — that they are probably overlooking something, probably wrong on at least one important point, and definitely unable to back up their recommendations with lived experience. So you use them. You just don’t hand them the wheel.

You delegate output to them. First drafts. Summaries. Reformatting. Translating a strategy you already decided into comms for 120 people. Turning a bullet list into a prose memo. Generating three versions of a paragraph so you can pick the best one. Pulling together meeting notes. Stress-testing a sentence. Legwork.

You do not delegate outcome. You don’t ask them what the company should do. You don’t ask them whether to hire, fire, pivot, or hold. You don’t ask them to define the strategy you are supposed to own. You don’t hand them the conversation with your top engineer. You don’t let them draft the quarterly vision from scratch.

The cleanest test I know is to ask yourself one question before every prompt: am I asking for output, or am I asking for outcome? If the answer is output, go wild. AI will make you five times faster on the work you already know how to judge. If the answer is outcome, close the tab. That decision is yours.

There’s a grey zone worth naming: pressure-testing. “Here’s my strategy, what am I missing?” is still output territory, but only if you remain the one weighing the response. The moment you adopt what it said without internalising it, you’ve slid into outcome. The line is not about the prompt. It’s about who’s making the decision after the answer lands.

And notice what all of this actually requires of you: expertise. The leaders who win with AI are not the ones who prompt the hardest. They are the ones who know enough about the problem to spot what the AI got wrong and steer it toward the good parts. Prompting is the skill everyone is talking about right now. Quality assurance is the skill that actually matters. The expert steers. The AI produces. Reversed, it’s a Junior Assistant running the company — and you already know how that goes.

This is also why the “use AI or get left behind” narrative is incomplete. You won’t get left behind because you didn’t use AI. You’ll get left behind because the person who did use AI brought expertise to it, steered it, corrected it, and shipped five times faster than you did. AI without expertise is a loud, confused junior. AI with expertise is a multiplier. The distinction is everything.

So what does this mean if you’re an SME leader reading this?

There’s a competitive edge buried in all of this, and most leaders are missing it.

If ChatGPT drifts toward the averaged answer, then the leader who refuses to outsource their strategic thinking to it is the one picking the outlier voice — the right answer, the conviction call, the move their competitors’ AI would never suggest. While one CEO is getting smoother-sounding communication out of a machine that appeases them, the other is out-deciding them on the decisions that actually compound over quarters.

The practical ask is small. Before every important prompt, pause for three seconds and decide whether you’re asking for output or outcome. Protect the outcome decisions like they’re the only thing on your job description — because in an AI-native company, they are. Everything else is delegable. That one isn’t.

Use AI. Use it aggressively. Use it everywhere there is output to produce and expertise to steer with.

Just don’t use it to lead. That part is still yours — and the next few years are going to reward the leaders who know the difference.