Strategy is subtraction. It is the set of decisions about what a company will deliberately not do — which markets to skip, which customers to disappoint, which opportunities to let a competitor take.

Language models are structurally bad at subtraction. They are trained to be helpful and agreeable, which means they will find merit in almost any plan presented to them. Ask whether you should expand into a new market and you will receive a balanced, well-organised assessment that somehow makes the case for proceeding. Ask the opposite question about the same market and you will get an equally convincing case against.

That asymmetry is the central thing to understand about AI in strategic planning. The technology is exceptionally good at generating options and genuinely poor at eliminating them — and elimination is where strategy actually happens.

Where It Genuinely Helps

This is not an argument against using it. Used correctly, it removes weeks from a planning cycle.

Option generation. Producing twenty possible expansion paths, market entry approaches, or pricing structures in an hour. Most will be unsuitable. The value is in the three nobody in the room had considered, and in breaking the gravitational pull of the option leadership already favoured.

Research compression. Regulatory environments, market structures, competitor positions, and comparable case histories, assembled in a fraction of the time a team would need.

Structured stress-testing. Asking a system to construct the strongest possible case against your plan, then again from the perspective of each competitor who would be affected by it.

Assumption extraction. Feeding in a draft strategy and asking what it assumes to be true. Strategies fail on unexamined assumptions far more often than on flawed logic, and this is a task machines do well because they have no stake in protecting the plan.

Documentation. Turning a messy planning session into a coherent document that everyone can argue with precisely.

The Agreeableness Problem, and How to Work Around It

If you take one technique from this article, take this: never ask a model to evaluate your plan. Ask it to destroy it.

The framing determines the output. "What do you think of this expansion strategy?" produces a supportive analysis with polite caveats. "You are a competitor who wants this expansion to fail. What is your plan?" produces something considerably more useful.

Techniques that force genuine disagreement:

  • Assign an adversarial role. A competitor, a sceptical investor, a regulator, a customer who chose someone else.
  • Ask for the failure mode first. "This plan failed. It is eighteen months later. Write the post-mortem."
  • Request the strongest opposing case, explicitly, and forbid balance.
  • Vary the framing across separate sessions and look for the conclusions that survive all of them.
  • Never reveal your preference. Stating which option you favour contaminates every subsequent response.

That last point matters more than people expect. A model told which direction you are leaning will tend to support it, and the support will sound like independent analysis.

Market Entry: Separating the Knowable From the Unknowable

Expansion decisions are where teams most often mistake a well-formatted answer for a researched one.

Knowable from public sources, and worth automating:

  • Regulatory requirements, licensing regimes, and compliance obligations.
  • Market size estimates from published sources, with their methodology and limitations.
  • Competitor presence, pricing, and positioning in the target market.
  • Labour costs, corporate structures, and tax treatment.
  • Distribution and payment infrastructure.
  • Case histories of companies who entered and failed, which are more instructive than the successes.

Not knowable without going there:

  • Whether local buyers actually want your product in the form you sell it.
  • How purchasing decisions are made inside organisations in that culture.
  • Which relationships are prerequisites rather than advantages.
  • What the informal rules are — the practices everyone follows and nobody writes down.
  • Whether your brand means anything, or means something unfortunate.

The second list has ended more expansions than the first. A company can complete every piece of documented research flawlessly and still fail because it never learned that in this market, nothing moves without an introduction.

Scenario Planning, Done Properly

Scenario planning has a reputation for producing documents nobody reads, and it earns that reputation when done as an exercise in optimism, realism, and pessimism.

The version that works is different. It builds scenarios around the specific uncertainties that would actually change your decision:

  1. Identify the two or three variables that genuinely determine the outcome. Not ten — two or three.
  2. Build scenarios at the extremes of those variables, not around a mood.
  3. Write what you would do in each, concretely, including what you would stop doing.
  4. Define the observable signals that indicate which scenario is unfolding.
  5. Set review dates and assign an owner to watch each signal.

Machines are useful across all five steps and decisive in none. They can generate the scenarios and identify signals; they cannot tell you which two variables actually matter for your business, because that requires knowing things about your company that are not written down anywhere.

The Pre-Mortem Is the Highest-Value Technique

Borrowed from decision research and underused in practice: before committing, assume the initiative has already failed and write the explanation.

Run it twice. Once with your team, once with a model instructed to be uncharitable. Then compare the two lists.

The gap between them is diagnostic. Reasons the machine identified that your team did not are usually things everyone knew and nobody wanted to say. That is the most valuable output of the entire planning process, and it costs an hour.

Where It Fails Completely

Being clear about the boundary prevents expensive misuse:

  • Taste. Whether a product is good, whether a brand feels right, whether a market is worth being in for reasons beyond return on capital.
  • Timing. The difference between early and wrong is invisible in any dataset.
  • Conviction. Strategy requires committing resources under uncertainty. A model produces balanced assessments, and balance is the enemy of commitment.
  • Internal politics. Which executive will quietly undermine the plan, which team lacks the capability the plan assumes, which division will not cooperate. This determines outcomes more than analysis does, and none of it is in the data.
  • Relationships. In many markets, the deciding factor is who knows whom.

A Planning Cycle That Uses It Sensibly

  1. Frame the actual decision. Not "should we grow" but "should we enter this market, at this cost, by this date."
  2. Generate options broadly, including ones you expect to reject.
  3. Research the documented layer thoroughly and quickly.
  4. Identify what remains unknowable and design a way to learn it cheaply — a small test, a few conversations, a limited pilot.
  5. Stress-test adversarially before deciding, not after.
  6. Run a pre-mortem, twice.
  7. Decide with humans, and record why, so the reasoning can be examined later.
  8. Define kill criteria in advance, with a date and an owner.

Step eight is what most organisations skip, and it is the one that limits the damage when the decision proves wrong.

Execution Is Where Plans Die

Strategy documents fail more often in delivery than in design, and this is the part of planning that gets the least attention because it is the least intellectually satisfying.

A plan that survives contact with an organisation needs three things that no analysis produces: someone accountable, a sequence that respects capacity, and a mechanism for noticing early that it is not working.

Where the technology contributes to execution rather than to planning:

  • Dependency mapping. Laying out what must happen before what, and where a single team is on the critical path of four initiatives simultaneously.
  • Capacity honesty. Comparing what the plan requires against what the organisation actually has, which frequently reveals that three strategic priorities are competing for the same two engineers.
  • Progress synthesis. Pulling status from the systems where work happens, rather than from a monthly slide that someone assembled optimistically.
  • Early variance detection. Flagging when a milestone is slipping while there is still time to respond, instead of at the quarterly review.

The uncomfortable finding this usually surfaces is that the company has committed to more strategic initiatives than it can staff. That is a subtraction problem again, and it is the one leadership teams find hardest to face.

The Convergence Risk

A structural concern worth raising, because it applies to an entire industry rather than to any one company.

If every competitor in a market uses similar systems, trained on similar public information, asking similar questions, they will arrive at similar strategies. Analysis converges. Positioning converges. And the market fills with companies making the same reasonable, well-researched, indistinguishable decision.

Competitive advantage has always come from knowing or believing something others do not. As the cost of consensus analysis falls toward zero, the value of consensus analysis falls with it — and the premium on proprietary insight, direct customer contact, and willingness to act on an unpopular conviction goes up.

The Summary Worth Keeping

AI in strategic planning removes the research bottleneck, widens the set of options considered, and forces assumptions into the open. Those are real contributions and they are worth having.

What it cannot do is choose. Strategy is the act of committing to one path and accepting the cost of the paths abandoned, under uncertainty that no amount of analysis resolves.

The tools make the analysis cheaper. They make the judgement more valuable, not less — which is the opposite of what most people expected, and the most important thing to understand about them.