Cheap Thinking Makes Strategy Harder

She has run the plan the same way for twenty years, and until recently it worked.

You watched the industry. You went to the conferences. You saw which competitors were hiring and which were quiet. You made the team ten per cent bigger and aimed for eleven or twelve per cent more revenue, and away you went — year on year, more profitable, stronger. It was not naive. It was a method, and it was calibrated against a world that changed slowly enough for calibration to mean something.

That method has stopped working, and the people using it are not the reason. What broke is not the discipline of the planners. It is the assumption underneath the discipline: that the number of things that could plausibly happen to your business is small enough to think about.

Most explanations for the current disorientation point at speed. Everything is faster; the tools change every quarter; keep up. That explanation is popular because it is comforting — speed is something you can respond to by running harder. It is also wrong, or at least it is the smaller half of the story.

AI makes thinking cheaper at exactly the same time that it makes strategy harder. Those are not contradictory. They are causally connected.

That is the claim this piece defends. Not that AI creates uncertainty — every technology does — but that the specific thing making strategy harder is the same thing your AI programme is celebrating internally. The cost of thinking fell. Everyone’s cost of thinking fell. And the consequence of universally cheap cognition is not convergence on the right answer. It is an explosion in the number of answers that are now affordable to try.

Generic cognition is the floor — not thinking

Start with the cost collapse, because the size of it is easy to underestimate and easy to overstate, and the argument needs it stated precisely.

Epoch AI measured the price of reaching a fixed level of capability across six benchmarks over three years. The price of matching GPT-4’s performance on PhD-level science questions fell roughly 40× per year; across all the milestones they measured, the decline ranged from 9× to 900× per year, with the fastest drops occurring most recently — and Epoch itself notes it is unclear whether those persist.¹ Gartner’s forecast is that inference on a trillion-parameter model will cost providers more than 90 per cent less in 2030 than in 2025.²

Two things matter in that paragraph. The first is the magnitude: a capability that cost real money to deploy in 2023 is now a rounding error, and the direction is not in dispute. The second is the unevenness — 9× to 900× is not a uniform wave, it is a series of specific tasks becoming free at different times. Cheapness arrives task by task, which is exactly why it is hard to plan against. You cannot schedule it.

The tempting conclusion is that thinking has become table stakes. If your business is the floor, you are out of business.

That conclusion is directionally right and one word short. It is not thinking that has become the floor. It is generic cognition — reproducible analysis, competent drafting, standard research, first-pass expertise, the work that has a known shape before anyone starts it. What has not become cheap is knowing which analysis is worth running, recognising that a structural break has occurred, and accepting responsibility for the decision that follows. Our own Great Reset argument is that every process designed on the assumption that thinking is expensive is now built on a false premise, and that what stays scarce migrates towards trust, taste, judgment, accountability and relationships. Nothing here contradicts that. This piece is about what happens after the premise breaks — when everyone else’s premise breaks at the same moment.

The Fog has two directions, and only one of them is about speed

We named this condition the AI Fog: the simultaneous compression of the credible forecast horizon and expansion of the plausible solution space. The visible distance shrinks; the map enlarges. Less time, more possibility, less clarity.

The first term is well covered. Half of CEO planning effort now sits inside horizons of less than a year, up from 43 per cent the year before.³ Boards feel that, and the standard response — shorter cycles, faster iteration — is a sensible response to that term alone.

The second term is the one nobody manages, and it is the subject of this piece. Solution-space expansion is not a mood. It has a source, and the source has a direction: the falling cost of cognition is feeding it. Every actor in your market can now search more possibilities, build more of them, test more of them and enter more of them than they could two years ago. That includes the people you compete with, the people you sell to, and the people who have not started yet.

Cheap thinking does not make the future easier to understand. Cheap thinking creates more futures to understand.

This distinction is not pedantry, because the two terms call for opposite responses. Horizon compression alone says: plan in shorter cycles. Solution-space expansion alone says: hold optionality, build portfolios. The joint condition says something neither of those does — that the value of choosing which question to ask has risen, because the answers have got cheap. Cheap answers increase the value of choosing the question.

Three channels: customers, competitors, constructors

Solution-space expansion sounds abstract until you draw the routes. There are three that cheap cognition drives directly, and they behave differently enough that a firm can be exposed to one and not the others.

Channel one — customers

The first channel is the one incumbents notice last, because it does not look like competition. Your customer’s cost of thinking fell too. They can now do the first pass themselves, or have it done, and arrive at the meeting holding it.

Legal services show this most clearly because the unit of sale is so visible. Research cited by Clio finds that 67 per cent of corporate legal departments and 55 per cent of law firms expect AI to change how hours are billed, and that clients have moved faster than the industry has; 71 per cent already prefer a flat fee for an entire matter.⁴ Note the direction of that gap. The demand side repriced before the supply side decided to.

What the customer channel expands is not the number of competitors. It is the number of ways the customer can meet their own need — and every one of those is a future in which some part of your revenue does not exist.

Channel two — competitors

The second channel is the familiar one, with an unfamiliar mechanism. Your competitors did not get better at your business. They got cheaper at changing their business. When the cost of designing, modelling, testing and specifying a new offer collapses, the rate at which competitors can attempt structural change rises — and each attempt is another plausible future you have to consider.

Public markets have already started pricing this. Sapphire Ventures’ software indices fell 20 per cent, and pure SaaS 23 per cent, through 18 February 2026, with the sector decoupling from the broader market — the IGV index down 32 per cent while the Nasdaq was essentially flat. They attribute it to pricing moving away from seats towards measurable value, and to model improvements raising displacement risk for incumbents. Their summary line is worth reading twice: “risk is up, and terminal value assumptions are down.”⁵ Sapphire hedge it themselves — investors may be selling first and asking questions later. Even so, the market is now discounting a term that most boards do not have on their agenda.

Channel three — constructors

The third channel is the entrants who do not exist yet, and it is the one most often argued badly. The lazy version is “there are more startups now.” The data does not straightforwardly support that. Crunchbase recorded $300 billion invested across 6,000 startups globally in the first quarter of 2026, with AI taking 80 per cent of it — a record by a distance. But seed deal counts fell 30 per cent year over year, to 3,800; the dollar increase came entirely from larger rounds.⁶ Capital is concentrating, not scattering.

So the constructor channel is not primarily about venture-funded startups. It is about the falling cost of construction available to everybody — including a customer who decides to build the thing rather than buy it, and including you. The number of parties who can credibly attempt to build a replacement for part of your value chain has risen, whether or not any of them raised a round.

Productivity applied to a depreciating unit

Here is where the argument stops being interesting and starts being expensive.

The standard enterprise response to all of the above is a productivity programme. Assistants for everyone, a target for time saved, a dashboard. It is fundable, it is measurable, it is career-safe, and in most firms it is genuinely working — people really are faster.

Having all your staff go ten per cent faster or twenty per cent more productive does not get you through the Fog. In a particular and identifiable set of circumstances, it gets you to zero.

The mechanism is not that productivity is bad. It is that productivity applied to a depreciating commercial unit accelerates the economics of its depreciation without giving you ownership of the successor unit. If the thing your invoice is denominated in is the thing cheap cognition compresses, then every efficiency gain shrinks your own revenue base — and does so faster the better your programme runs.

Clio put the same point in a single sentence from the vendor’s side of the table: a tool that turns a four-hour task into a one-hour task “creates a revenue problem under hourly billing and a margin opportunity under fixed fees.”⁴ Same tool. Same efficiency. Opposite economics, decided entirely by the unit of sale.

A board announcing that everyone is now using AI and thinking faster is barely interesting, because everybody will be able to do that. The interesting question is the one nobody has been assigned: do we possess machinery for directing abundant cognition at the questions that determine whether this company still deserves to exist?

The arithmetic of faster stranding

Assertions about anti-strategic productivity are cheap. Here is the arithmetic, worked, on a composite firm.

What this is: an explicitly labelled worked model with stated assumptions. The parameters are illustrative — chosen because they are round, defensible and recognisable to anyone who runs a practice — not measured data from any firm. The purpose is to show the shape of the mechanism and where its hinges are, so a reader can substitute their own numbers.

The firm, before

A forty-consultant advisory firm. Each consultant delivers 1,000 billable hours a year, so annual capacity is 40,000 hours. The standard engagement takes 100 consultant-hours and bills at $300 per hour: $30,000. At 400 engagements a year the firm bills its full capacity and turns over $12.0 million. Fully loaded staff cost is $180,000 per consultant ($7.2 million), overheads are $2.4 million, so total cost is $9.6 million and operating margin is $2.4 million — 20 per cent.

Before AI touches anything, note how much slack is already in the model. Clio’s 2025 Legal Trends Report data puts law-firm utilisation at around 38 per cent, realisation at 88 per cent and collection at 93 per cent — firms collect revenue for only a fraction of the workday even when everything is working.⁷ Our composite is deliberately generous to itself. The mechanism below does not depend on the firm being badly run; it bites hardest when the firm is well run.

The compression

The firm deploys AI properly. Drafting, research, first-pass analysis and document review all get faster. The standard engagement now takes 80 consultant-hours instead of 100 — a 20 per cent compression, which is a modest and widely reported order of magnitude, not a heroic one.

Suppose the firm passes that through, because its unit of sale is hours and it bills what it works. It has used its own AI investment to shrink its revenue unit by twenty per cent. The customer has captured much of the cognition-cost collapse. The firm’s product has not changed. Its pricing has not changed. Its value proposition has not changed.

A successful, well-executed, on-budget AI programme has moved the firm from a twenty per cent margin to break-even, and every internal metric improved while it happened. Hours per engagement: down. Output per consultant: up. Client satisfaction: probably up. The dashboard is green and the firm is worth less.

The four escapes, tested one at a time

The obvious objection is that no firm would just sit there. If a consulting firm becomes twenty per cent faster it does not mechanically go out of business: it could fill the freed capacity, preserve pricing, increase volume, or improve margins. Each of those is a real escape. Each has a condition attached, and the conditions are the whole point.

Escape one — fill the freed capacity. Capacity is now 40,000 ÷ 80 = 500 engagements. At $24,000 each, that is $12.0 million: revenue restored exactly. The condition is that demand exists for 25 per cent more engagements at the same price. If the firm can only sell 440, it delivers 35,200 hours, bills $10.56 million, runs at 88 per cent utilisation and earns $0.96 million — a 60 per cent fall in profit from a 10 per cent shortfall in demand. This escape has a hard dependency on market growth that the firm does not control, and every competitor’s capacity rose at the same moment.

Escape two — hold the price per engagement. Bill $30,000 for 80 hours. Revenue holds at $12.0 million, margin returns to $2.4 million, and the effective rate rises from $300 to $375 an hour. This is the escape that works — and notice what it actually is. The firm has stopped selling hours. It is now selling a defined piece of work at a fixed price, which is a different commercial unit with different risk, different scoping discipline and different competitive exposure. It is a good move. It is not a pricing tweak.

Escape three — raise the rate. Arithmetically identical to escape two: $375 per hour for 80 hours. Behaviourally much harder, because the client can observe that the work took less time and is being asked to pay more per unit for it. The Thomson Reuters Institute’s Law Firm Rates Report 2026 found that regardless of whether firms discount aggressively or hold firm on realisation, they end up collecting roughly the same amount per hour.⁸ Rate rises are being competed away.

Escape four — keep the dividend quietly. Deliver in 80 hours, bill 100. This is the historical pattern, and it deserves to be taken seriously rather than dismissed. When Westlaw and LexisNexis compressed legal research from hours in a library to minutes at a terminal, “the efficiency dividends didn’t flow to clients. Firms captured it, and billing rates continued climbing.”⁹ That is the strongest case against everything in this article, and the honest answer is that it worked for thirty years.

Two things are different this time. The earlier tools optimised around the professional — better retrieval, faster organisation — leaving the core work untouched. Today’s tools do the work itself: drafting, reviewing, analysing, summarising. And the client has the same tools, which means they no longer merely suspect that the work got faster; they can perform a version of it and see. SignalFire’s framing of the window is precise: in the short term AI may expand margins if pricing holds, and early adopters may benefit significantly — but “that window won’t stay open for long as prices adjust across the market over time.”⁹ Escape four is real and temporary, which makes it a harvest, not a strategy.

The compounding case

Now run the channels together, which is what actually happens, and which no single-variable analysis catches.

Assume 30 per cent of the original 100-hour engagement was first-pass analysis — the scoping, the market scan, the baseline model, the competitor summary. Assume the client now does that themselves, because their cost of thinking fell too. The addressable engagement drops to 70 pre-compression hours, and with the firm’s own AI applied, to roughly 56 hours. At $300 an hour, the engagement is worth $16,800 — down 44 per cent from $30,000.

The bottom two rows are the finding. To stand still, this firm must win 79 per cent more engagements — and at that volume it is exactly at capacity, so there is no slack left for the next increment of compression. If it wins no new work at all, revenue falls to $6.7 million against a $9.6 million cost base: a $2.9 million loss, from a firm that did nothing wrong operationally.

And it must win those engagements in a market where every competitor’s effective capacity rose by the same 25 per cent at the same time. Supply rose exactly as willingness to pay per unit fell. That is not a demand problem the sales team can fix with more activity.

One more move, and it is the one that should worry a partnership most. An AI-native entrant is not obliged to stop at “80 hours instead of 100.” It can ask why the customer is buying hours at all.

Boundary cases that earn their place

If forecasting cannot work under these conditions — and it cannot, because the plausible-future set is now too large to enumerate and the drivers move faster than the planning cycle — then what does?

Push one variable to a structural extreme and see which of your assumptions survive. We have written the method up at length: the human supplies the pivot, the machine supplies the proof; strip the vague terms, stretch the variable, stress the structure, stage the decision. This article does not re-teach it. It adds one test that decides whether a boundary case is worth running at all.

If pushing the variable to zero or to infinity collapses or transforms the set of legal moves, you have a pivot. If pushing it to either extreme just makes the situation “more of itself” — more expensive, slower, larger — but leaves the legal moves unchanged, you don’t.

Apply that to the two probes most firms reach for. “What if AI adoption hits 100 per cent?” fails the test: at the extreme, everyone has the tools, and the set of moves available to you is the same set, run faster. Nothing changes shape. “What if competent generic advice is effectively free?” passes: at the extreme, an entire category of promise you currently sell becomes unsellable, a pricing structure becomes indefensible, and a hiring model becomes a liability. The legal moves change shape, not just size.

Two disciplines keep this honest.

The probe is not a prediction. “Competent generic advice is effectively free” is a boundary probe. It is not a claim that advice will cost zero dollars, and it carries no date. You do not need to know whether it arrives in February 2027 or October 2028. If the answer is that a large share of your commercial proposition disappears when competent advice becomes free, you have learned something enormously valuable today. Thought experiments work under Fog because they exchange timing certainty for structural certainty.

You pierce the Fog locally; you never clear it. This is not modesty, it is mechanism. A search apparatus that is good at finding possibilities manufactures Fog as a side-effect of working well — every good search surfaces new candidate futures, and the solution space grows. Add the cheap-thinking paradox and it is worse than that: every other actor’s search is doing the same thing to you. The correct posture is not “run the search until the Fog lifts.” It is to build an operating rhythm that is productive inside permanent Fog.

The Strategic Search Ledger

A boundary case that produces a good conversation and no artefact is entertainment. The artefact is a ledger, and it has six columns.

The fifth column is the one that does the work, and it comes with a rule: an entry does not close until the decision-changed column is non-empty. Without a decision delta, the exercise is scenario theatre — a wall of futures, a satisfied room, and an unchanged budget.

This is a compressed, executive-row descendant of the twelve-field Question Ledger in the Terminal Value Doctrine, which remains the definitional artefact and owns the full schema and the gap audit. What this six-column form adds is the closing rule. The Question Ledger makes the search inspectable; the Strategic Search Ledger makes it consequential.

One completed row

That row took a morning. It rejected one future, changed two present decisions, and produced a question the firm did not have last week. None of it required knowing when anything will happen.

Where the paradox stops

An argument that only runs one way is a slogan. Cheap cognition does not always multiply futures. Sometimes it genuinely reduces uncertainty, and the difference is structural rather than a matter of degree.

Audit is the clean case. KPMG describe AI moving fieldwork beyond sample-based testing to “reviewing full data populations, identifying anomalies, trends, and exceptions” — instead of testing a sample of transactions, analysing the whole population.¹⁰ Independent trend analysis reports the same shift, and adds the consequence: it raises “regulator expectations of what is now technically achievable.”¹¹

Nothing multiplies here. The population of transactions is finite and enumerable, the test is well-defined, and the answer set does not grow when you think harder — it converges. Sampling existed because attention was expensive. Cheap cognition dissolves the constraint and the uncertainty falls.

This is why our own Cognition Scarcity Audit lands on both sides of the line. Its first, second and fourth scarcity signatures — analysis performed only on a sample, events investigated only when they escalate, synthesis that waits for someone to remember to ask — are all bounded. Cheap cognition simply fixes them, and firms should go and fix them. Its third signature — planning limited to a handful of scenarios — is unbounded, and that is precisely where more cognition produces more futures rather than fewer. The same audit contains the boundary of the paradox.

The practical instruction is therefore not “be careful with AI.” It is: sort your questions before you spend cognition on them. Flood the bounded ones — there is enormous, uncontroversial value sitting in census-instead-of-sample work. Discipline the unbounded ones, because that is where undirected thinking manufactures the very Fog you are trying to see through.

What changes on Monday

Four questions organise everything above, and they are the questions a chair can put on the agenda without a consultant in the room.

What becomes cheap? What therefore stops being valuable? What remains scarce? And how do we systematically reason our way toward the business that captures that remaining value before somebody else builds it?

Then three concrete moves.

Name your commercial unit and test it for compressibility. Write down what the invoice is denominated in. Then ask what happens to it when the work behind it takes 40 per cent less time and the client can observe that. If the answer is “revenue falls,” your productivity programme is a depreciation accelerator and should be re-scoped, not celebrated. Keep it — it pays — but stop calling it strategy.

Run one boundary case properly, and use the pivot test to choose it. One variable. Pushed to a structural extreme. Chosen because at the extreme the set of legal moves changes shape. Not five scenarios. One, worked until something breaks.

Close one ledger row with a decision that changed. Six columns. The fifth is not optional. If, at the end of the exercise, the budget, the hiring plan and the pricing sheet are all exactly what they were before, the search did not happen — you had a conversation about the future and filed it.

None of this clears the Fog. It cannot: the same cheap cognition that lets you search is funding everyone else’s search, and a good search apparatus manufactures new possibilities as a by-product of working. What it does is trade timing certainty, which you cannot have, for structural certainty, which you can — and leave behind a record of which futures you rejected and why, so that when one of them starts arriving you are executing rather than debating.

AI makes thinking cheap, but it does not make strategy easy. Cheap thinking destroys old scarcities while multiplying the number of plausible futures. The board’s job therefore shifts from extrapolating yesterday to allocating disciplined cognition against terminal-value questions: identify the load-bearing variable, push it to the boundary, search what survives, and build the successor before the market does.

The firm that goes twenty per cent faster at selling hours has not answered any of that. It has just arrived at the same place sooner.