Can AI Do Strategic Planning? What It Can and Can't Do

Author: Eric Levine, Founder of StratEngine AI | Former Meta Strategy & Operations | UCLA Anderson MBA | CPA

Published: July 16, 2026

Reading time: 8 minutes

Summary

AI can do the inputs to strategic planning. It cannot do the deciding. It will research a market, synthesize a hundred documents, generate scenarios you hadn't considered, and draft the plan document faster and often better than you would alone. What it won't do is tell you which three things matter, refuse a bad idea you're attached to, or weigh the things nobody wrote down. Strategy is choosing what not to do, and the choosing is the part that stays yours.

The line isn't where most people think it is. A pre-registered field experiment with 758 BCG consultants found AI users performed substantially better on tasks inside AI's capability and 19 percentage points worse on a task outside it, which means they did worse than colleagues using no AI at all. The frontier is jagged and it isn't labeled, and the same tool that can't tell you it's out of its depth is the tool disposed to tell you your plan looks strong. Better models won't dissolve that limit, because the limit isn't model quality.

That's the short answer. The longer one is more interesting, because the best available evidence says you can't see the line from where you're standing.

Key Takeaways

  • AI is genuinely strong inside its frontier. In a pre-registered study of 758 BCG consultants, those using GPT-4 completed 12.2% more tasks, 25.1% faster, with quality rated more than 40% higher.
  • The same study found consultants were 19 percentage points less likely to be correct on a task that fell outside AI's capability. The boundary is real, and it isn't labeled.
  • The failure isn't that AI gets strategy wrong. It's that it produces something structurally indistinguishable from right, and won't tell you which side of the line you're on.
  • Better models won't dissolve the limit, because the limit isn't model quality. It's that strategy needs friction, persistent context, and judgment about things that exist in no document.

What AI genuinely does well in strategic planning

Start with the strengths, because overclaiming AI's limits is its own kind of dishonesty, and the evidence here is good.

The strongest study we have is a pre-registered field experiment run with 758 BCG consultants, published in Organization Science in 2026. On 18 realistic consulting tasks, consultants using GPT-4 completed 12.2% more tasks, moved 25.1% faster, and produced work rated more than 40% higher in quality than the control group. The effect was also a leveler: below-average performers improved 43% against their own baseline, above-average performers 17%.

That maps onto four things worth handing to AI. It does research and synthesis well, reading more sources than you can and pulling out themes. It generates scenarios and options, including ones you'd have missed, because it isn't invested in your existing plan. It drafts framework analyses quickly, giving you a populated first pass instead of a blank page. And it documents: turning a messy decision into a readable plan is genuinely mechanical work, and mechanical work is what it's for.

Two caveats worth stating. The study ran on GPT-4 in 2023, which is now several model generations old, and four of its co-authors work at BCG, though it's peer-reviewed and pre-registered, which is what makes it worth citing at all. Also, those 18 tasks were chosen because they sat inside what AI can do. That selection is the whole point of what comes next.

Where AI structurally fails, and why

The site has a longer argument about the mechanism here, so I'll summarize it in three moves and point you at the detail rather than re-run it.

It reproduces the artifact, not the act

A language model predicts the next token from patterns in what it has read. Strategic analysis left behind artifacts (filled-in SWOT grids, board memos, the vocabulary of moats and risk-adjusted opportunity), but the reasoning that produced them, the arguing and discarding and choosing, was never written down. So the model regenerates the visible output of strategy with high fidelity while having done none of the work that gave it meaning. That's the difference between framework templates and framework logic, and it's the distinction the whole category glosses over.

It won't fight you

Ask a model for feedback and it tends to give you the feedback you signaled you wanted. This isn't a design goal; it's a side effect. Training on human preference data rewards agreement, because people rate agreeable answers highly, even though the same training also rewards being right. Anthropic researchers documented the effect across five assistants in a 2024 ICLR paper, finding that feedback shifted with a user's stated preference rather than the text's merit. A 2025 preregistered study from researchers at Stanford and Carnegie Mellon found models affirm users' actions about 50% more than humans do, and, across experiments with 1,604 participants, that sycophantic AI left people more convinced they were right while they rated it higher and trusted it more. That last part is the trap: the reassuring answer is the one users prefer. OpenAI took the problem seriously enough to roll back a GPT-4o update in April 2025 that it judged "overly flattering or agreeable," saying it had "focused too much on short-term feedback." Both studies looked at general advice and interpersonal conflict rather than business strategy, so read them as evidence of a disposition rather than a measurement of your planning session. The strategic cost of that disposition is the argument that strategy needs the friction of something willing to say no: a tool that won't resist you can list what's true about your business, but it won't make you cut anything.

It can't weigh what isn't written down

Models work from what's in the data, and the polish of the output hides what the data never contained. The corpus argument focuses on data the model can't reach: your internal financials, your customer conversations. But there's a harder category underneath it, and it's where most real strategy actually turns. Whether your head of engineering will quit if you kill her project. Whether the board has one more expensive failure in it. Whether your founder will actually ship the hard thing or just talk about it for two quarters. None of that is in a document anywhere. It isn't private data; it's undocumented data, and it frequently decides the outcome. A model can't weigh it, and it won't tell you it's missing. It will hand you a confident plan built as if those factors were neutral.

The frontier is invisible from the inside

Here's the part that makes this more than a list of limits.

That same BCG study included one task deliberately placed outside AI's capability. Consultants using AI were 19 percentage points less likely to get it right than those working without it, which means the AI group did worse than colleagues who had no AI at all, because they trusted output that looked exactly like the output that had been excellent all day.

That's the whole problem in one number. The frontier is jagged, and it isn't marked. The work AI does brilliantly and the work it does confidently-and-wrong arrive in the same font, the same structure, the same measured tone. Then put that next to the sycophancy findings: the tool that can't tell you it's out of its depth is the same tool disposed to tell you your plan looks strong.

So the honest answer to "can AI do strategic planning" isn't a percentage of the work you can automate. It's that AI can do a lot of it well, and cannot tell you which parts those were. That's what makes a human reviewer non-optional, and it's a different claim from "AI isn't good enough yet."

So what's the right division of labor?

Let AI do the inputs and the artifacts. Give it research, synthesis, option generation, first-draft framework passes, and the writing-up. That's where the measured gains are, and refusing them is leaving real speed on the table.

Keep three things human. Keep the prioritization, because a model won't force the cut and you need someone to say no. Keep the judgment about undocumented context, because the model doesn't know that data is missing. And keep ownership of the call, because accountability isn't a feature you can delegate to something that will agree with your next idea just as readily.

The practical version of this split is in the guide to how to use AI for business strategy, which walks the workflow stage by stage.

Does purpose-built AI change the answer?

Partly, and it's worth being precise instead of pitching.

The constraint isn't raw model quality; it's that a chat window is frictionless and stateless, agreeing with whatever you bring it and starting from zero every session. Purpose-built tools move the line by fixing that architecture rather than the model: holding a persistent profile of your business across sessions so your analyses reconcile, enforcing framework logic instead of framework shape, and forcing ranked outputs where a chatbot returns an unranked list. That's a real difference, and it's why purpose-built AI strategy tools exist as a category distinct from chatbots.

What it doesn't do is move the line to the end. A structured engine still can't know whether your head of engineering will quit, and it still can't be accountable for the call. It shifts what gets automated; it doesn't dissolve the boundary. Anyone telling you their tool does the deciding is selling you something.

Will this change as models improve?

Some of it, and the split matters.

Model-quality limits will keep improving; that's the safest prediction in the field. But the durable limits here aren't about quality. A model that agrees with you more fluently isn't closer to forcing a hard trade-off. A model with a larger context window still hasn't been told what nobody wrote down. And no model, at any capability level, can hold accountability for a bet on your company, because accountability is a thing organizations assign to people.

The interesting evidence cuts against the "just wait" reflex. Researchers benchmarked 34 models on a strategy simulation that trades short-term profit against long-term position, and published the results in Strategy Science. The curve isn't a straight line up. Reasoning models from late 2024 and early 2025 beat the average MBA cohort, and then the frontier models that followed in mid-to-late 2025 declined, underperforming both the earlier models and the students, partly because of a systematic bias toward exploiting the core business at the expense of investing in future growth. Newer did not mean better at the thing strategy actually asks for. Some limits are engineering problems and some are category errors.

The reasonable expectation is that the frontier keeps expanding and stays jagged. More of the work moves to AI, the boundary keeps moving, and it stays unlabeled. Which means the skill that matters isn't knowing which tasks AI can do today. It's keeping the habit of checking, on the specific decision in front of you, whether you're inside the frontier or outside it, at a moment when the output looks identical either way.

Writing for the UVA Darden Batten Institute, strategy consultant Saurabh Kapoor answers the headline question "no," while adding that it's more nuanced than that: AI can't do strategy on its own, and the better question is whether human-AI collaboration can produce board-ready work. That's close to where I land, with one shift in emphasis. AI does a real and growing share of strategic work genuinely well. It just can't do the part that makes it strategy.

Frequently asked questions

Can ChatGPT write a strategic plan?

It can write a document that looks like a strategic plan, quickly and often well. That's genuinely useful for the drafting and documentation stage, when the decisions are already made and you need them written up. What it can't do is make the decisions the plan records: which bets to fund, which to kill, what to sacrifice. Feed it your real constraints and it'll produce a competent draft. Ask it to choose, and it'll produce something that sounds like a choice without having made one.

Is AI strategic planning accurate?

Accurate on the parts that have a checkable answer, and confidently wrong on the parts that don't, in the same tone. The BCG study found AI-assisted consultants produced work rated more than 40% higher in quality inside AI's capability, and 19 percentage points less accurate outside it. Since strategic questions rarely come labeled with which side of that line they sit on, treat AI output as a hypothesis to verify rather than an answer to accept, especially where it agrees with what you already believed.

What percentage of strategic planning can be automated?

Nobody can give you an honest number, and any specific percentage should make you suspicious. The reason isn't modesty; it's that the boundary is jagged and moves per task, per industry, and per how much of the relevant context was ever written down. The useful question isn't what fraction to automate but which kind of work you're handing over: gathering, synthesizing, and drafting are safe to delegate today, while prioritizing, weighing undocumented context, and owning the call are not.

About the Author

Eric Levine is the founder of StratEngine AI. He spent five years at Meta in Strategy and Operations, where he led global business strategy initiatives across international markets. He holds an MBA from UCLA Anderson and is a CPA. He builds AI-powered strategic planning tools used by operators, consultants, and executives to compress the mechanical work of strategy and reporting while keeping the judgment human.