Strategy

Confidence 0.92 · 21 sources · last confirmed 2026-10-25

The discipline of making integrative choices that position an organisation to win on a chosen playing field by creating value for customers, employees, and the firm. Distinct from strategic planning (which is a list of activities the firm intends to do) and from operations (the execution of those activities). The two foundational sources in this wiki frame strategy as theory-of-winning (Martin 2022) and as value-creation (Oberholzer-Gee 2022); the two definitions are complementary, not competing.

Working definitions (two complementary lenses)

Lens 1: Strategy as theory of winning (Martin)

Per Martin 2022:

“A strategy is an integrative set of choices that positions you on a playing field of your choice in a way that you win. Strategy has a theory: here’s why we should be on this playing field, not this other one, and here’s how, on that playing field, we’re going to be better than anybody else at serving the customers.”

The theory must be coherent, doable, and translatable into actions. Planning is the comfort-zone alternative: a list of cost-side activities the firm controls, without internal coherence or a theory of customer outcomes.

Lens 2: Strategy as a plan to create value (Oberholzer-Gee)

Per Oberholzer-Gee 2022:

“Strategy’s simple. It’s a plan to create value.”

Value is the gap between willingness to pay (WTP) at the top and willingness to sell (WTS) at the bottom of the value stick. Strategy is the choice of which value-stick to operate on (which market) and which moves to make to lengthen the stick (raise WTP, lower WTS, or both).

The two lenses are complementary: Martin tells you where to play and how to win; Oberholzer-Gee tells you what you’re optimising for once you’ve chosen.

Key claims

Planning ≠ strategy (Martin 2022)

The load-bearing distinction. Planning lives on the cost side (square feet leased, people hired, products launched) — you control it. Strategy specifies a competitive outcome (customers wanting your product enough that they buy enough of it to make you profitable) — you don’t control it; customers decide.

“It’s much easier to say, ‘I’ll build a factory, I will hire more people, et cetera’ than ‘I will have customers end up liking our offering more than those of competitors.‘”

The competitor risk that compounds the comfort-trap: “while you’re planning, chances are at least one competitor is figuring out how to win.”

The value stick (Oberholzer-Gee 2022)

A vertical bar from WTP (top) to WTS (bottom). Three wedges:

Wedge	Formula	Beneficiary
Customer delight	WTP − price	Customers
Firm margin	price − cost-of-supply	The company
Employee / supplier value	compensation − WTS	Employees / suppliers

Three ways to raise WTP: quality (better product/service); complements (razor/razorblade; printer/cartridge; espresso machine/capsule); network effects (the more users, the higher each user’s WTP).

Two ways to lower WTS — and the critical distinction between them:

• Pay more money = redistributes value from the firm to employees. “Value is just redistributed […]. There’s no value created.”
• Make the job a better job (training, generous promotion rules, work-from-home, more interesting work) = creates value. WTS drops, the stick lengthens.

The load-bearing implication: invest in job quality, not just compensation.

Two worked examples — Southwest Airlines and Best Buy

The two foundational sources each carry one canonical case study:

Southwest Airlines (Martin 2022)

US major air carriers in the era were “playing to play” — “playing to participate, maybe buy more planes, get more gates, maybe grow some, not having a theory of here’s how we could be better than our competitors.” Southwest had a theory of winning with Greyhound as the substitute target:

Choice	Why it fit the theory
Point-to-point routes	”You only make money when you’re in the air” — no idle ground time
737-only fleet	One type → simplified gates, systems, training, simulations
No meals	Short flights — saves cost and time
No travel agents; online booking	”Less expensive for everybody and more convenient”

Outcome: substantially lower cost → substantially lower prices → growth → “flies the most passenger seat miles in America.”

Best Buy turnaround (Oberholzer-Gee 2022)

≈$1B quarterly loss → 20%+ ROIC, via two value-stick moves under a new CEO:

Move	Value-stick effect
Stores as warehouses (ship from store down the road instead of central DCs)	Raises WTP (faster shipping)
Store-in-a-store (Microsoft/Samsung/Lenovo/Sony branded space inside Best Buy stores)	Lowers WTS for vendors (cheaper retail presence than standalone Apple-style flagships); lowers WTS for employees (specialise on one vendor’s catalogue → easier work → engagement up)

Methodological punchline: “We started with ideas about how to create value before we thought about how to capture a fraction of the value that we created.” Value creation precedes value capture.

Three practices for escaping the planning trap (Martin 2022)

1. Accept the angst. Strategy can’t be proven in advance. “That is not being a bad manager. That is being a great leader because you’re giving your organization the chance to do something great.”
2. Lay out the logic explicitly — “what would have to be true about ourselves, about the industry, about competition, about customers for this strategy to work?” Then watch the world unfold and tweak. “Strategy is a journey — a mechanism for tweaking it, honing it, and refining it so it gets better and better as you go along.”
3. Keep it to one page: where you choose to play, how you choose to win, capabilities you need, management systems, and the aspirational goal. “Then you lay out the logic, what must be true for that all to work out the way we hope.”

Strategy in the industrial-AI era — “winners determined by adoption speed”

Carrier 2026 (Senior Lecturer in System Dynamics, MIT Sloan) supplies the 2026 industrial-AI restatement of the strategy vs planning distinction:

“This technology is available to everyone. So winners will be determined not by who has access to the technology, but whose organization adopts it faster in a way that actually helps its system.”

The mechanism is direct restatement of Martin’s planning-trap at the technology-adoption scale. Carrier’s worked example: the Heineken Mexico brewery where his students built a relatively simple AI agent that grabbed machine + cloud + maintenance data on demand — shrinking a 6-hour changeover (with only 15 minutes of actual information content) to 15 minutes, yielding a million extra cases of beer per month. The methodological claim Carrier draws from the case: “This is a relatively simple agent, but it actually changed the way the work is done, not simply improve the workflow” — the Oberholzer-Gee 2022 value-stick lens reads this as a WTS-reducing move (lower per-case labour cost) that also raises WTP (more product per period at the same quality).

Carrier’s “pick the right agent level” heuristic — “there’s no reason to jump to a level five agent when a simple rule-based agent will do” — is also operationally important: in strategy terms, value capture from AI requires matching agent complexity to the system’s missing-feedback-loop, not to fashion.

The general industrial-AI strategy claim from Carrier: “Our ability to adopt and absorb the technology are going to be the limit” over the next 3–5 years — the micro-productivity-trap thesis at the industrial-AI scale, with the binding constraint explicitly named as adoption capacity, not technology.

Strategy as participation, not document — Erginbilgiç at Rolls-Royce

Erginbilgiç 2026 (Rolls-Royce CEO since Jan 2023; McKinsey-named “case study in the art of corporate transformation”) supplies a substantive strategy-formation doctrine that complements rather than competes with the Martin / Oberholzer-Gee lenses. Where Martin asks what is the theory of winning and Oberholzer-Gee asks where does value come from, Erginbilgiç answers an orthogonal question: how does an organisation arrive at — and align around — a strategy?

The doctrine: strategy is produced via participation, not delivered as a document.

• Personal CEO attendance at 25–30 strategy workshops (~12:50): not a single off-site retreat, but a sustained series across the organisation.
• 500+ employees through the strategy-formation process (~12:55–13:08): “every like 500-plus people joined. When you are done, whole organisation is aligned because they were in the room when you were making the decision.”
• Two protocol rules per workshop (~13:14–13:32):
  1. “No hierarchy in the room” — so junior voices can challenge senior assumptions during formation.
  2. “This is going to be chaotic” — explicitly licensed divergence: “many strategic conversations don’t open up. It closes. It becomes like plan conversations. You need to open it up so that all the out-of-blue ideas come into the room, even if you don’t do them, they are in the room.”
• Alignment is the output of participation, not the output of communication. The conventional model is strategy formed in the C-suite → cascaded down → alignment via communication discipline. Erginbilgiç inverts this: alignment is generated during formation because the people who will deliver the strategy are in the room when the strategic choices are made. “Whole organisation is aligned because they were in the room when you were making the decision.”
• Explicit contrast with consultant-led strategy formation (~12:43): “I don’t do strategy in the dark room with consultants.” Consultants are not banned, but they are not the substrate of strategy formation — the organisation itself is.
• Granularity as the validation criterion (~11:57–12:25, “second pillar” of the four-pillars framework): a strategy is granular when “everybody knows in the company their role in delivering strategy. If you don’t have granular strategy, only ten people know it and then ten people are surprisingly surprised that that strategy is not delivered.” Granularity is the antidote to the strategy-as-aspirational-deck failure mode.

This doctrine adds a process layer beneath Martin’s theory of winning and Oberholzer-Gee’s value creation: both lenses are silent on who arrives at the theory and how. Erginbilgiç’s answer — everyone affected, through chaotic non-hierarchical workshops, with the CEO personally present at scale — is a non-trivial commitment that explains why most strategy-deliverables-from-consultants don’t produce alignment even when the analysis is correct. Strategy without participation is a plan; strategy with participation is a contract.

It also supplies a non-AI control case for the wiki: the doctrine is articulated in an industrial-aerospace context with no AI substrate, suggesting strategy-as-participation is a transformation primitive rather than an AI-era pattern. The convergence with Allen 2026’s demanding-leader-as-cadence-primitive observation is consistent: both anchor transformation mechanics that pre-date and survive the AI substrate.

Strategy is a journey, not a one-shot plan

Both foundational sources converge here from different vocabularies:

• Martin: “strategy is a journey — a mechanism for tweaking it, honing it, and refining it.”
• Oberholzer-Gee: “the strategy is about looking forward, seeing the future, planning for the future.”

This convergence is consistent with Sterman 2026’s feedback-loop framing in systems-thinking — strategy is a continuously-updated theory under feedback, not an open-loop issue → data → optimal-choice → implement sequence. Different literatures, the same operational stance.

AI and strategic decision-making (Csaszar, Ketkar & Kim 2024)

The wiki’s first academic-strategy treatment of how AI changes the strategic decision-making (SDM) process — Csaszar, Ketkar & Kim (2024), Strategy Science — supplies two things the practitioner sources above don’t: controlled empirical evidence and a cognitive-process model.

• Empirical: across a start-up accelerator (LLM-generated business plans vs entrepreneurs seeking VC) and a start-up competition (LLM evaluations vs VC/angel investors), LLMs generate and evaluate forward-looking entrepreneurial strategies at a level comparable to humans in realistic, high-uncertainty settings.
• The cognitive-process model: SDM rests on search, representation, and aggregation; AI affects each, changing the generation and evaluation of strategies and, through them, competitive advantage. AI-augmented SDM can enhance the speed, quality, and scale of strategic analysis and enable new methods (virtual strategy simulations).
• The competitive-advantage trichotomy (a sharp addition to this page): as AI capability advances, the nature of advantage may stay Ricardian (rooted in unique resources), become Schumpeterian (innovation-driven), or erode entirely if everyone can cheaply generate comparable strategies. The third branch is the strategy-theoretic statement of the paradox of access and the micro-productivity-trap: “the ultimate impact on firm performance will depend on competitive dynamics,” not on the tool. If strategy generation commoditises, advantage migrates to complementary assets (proprietary data, unique execution) — and, per the theory-based-view, to the firm-specific, falsifiable theory of value an LLM trained on consensus cannot easily produce.

This is the AI-era extension of the page’s strategy-as-a-theory-under-feedback stance: the lenses above (Martin’s theory of winning, Oberholzer-Gee’s value stick, the TBV) describe what a strategy is; Csaszar et al. describe what happens to the cognitive work of producing it when LLMs can do parts of the search/representation/aggregation loop.

The durable-advantage principle: time-to-do vs time-to-happen ( AWS Sydney 2026)

The sharpest one-line statement of where advantage survives the commoditisation branch above: “AI compresses the time it takes to do things. It does not compress the time it takes for things to happen.” Brovich’s “Only Moats That Matter” slide sorts advantage by which clock governs it: the “hard to do” moats (workflow embeddedness, software scale, integration lock-in, engineering complexity) erode because AI compresses doing-time; the “hard to get” moats (compounding proprietary data — years of operations; network effects — years of adoption; regulatory permission — years of process; capital at scale — decades of trust; physical infrastructure — years of building) appreciate because their constraint is elapsed time, trust, and physics, which no capability gain shortens. This is the durable-advantage refinement of the trichotomy: when strategy generation commoditises, advantage migrates to whatever is bottlenecked by time rather than effort — the strategy-theoretic complement to “rent the model, own the harness” and the firm-specific theory of value.

Earning the right to innovate: “Proven, Better, New” ( Lenny’s Podcast, June 2026)

Mark Pincus (Zynga founder) supplies a founder-vantage product-strategy framework distinct from this page’s HBR-theory lenses: Proven (start from something already validated in the market — “earn the right to innovate” before differentiating), Better (not marginally improved, but decisively better — enough that “10 out of 10 people” say yes), New (only after proven + better is earned should a genuinely novel element be added). The framing reframes disciplined imitation-then-improvement as a legitimate strategic starting point, against a cultural bias that over-rewards perceived originality.

A companion claim sharpens where ambition should sit: narrower scope, executed decisively, outperforms broad, diffuse ambition — Pincus argues the most ambitious long-term outcomes come from less ambitious initial scoping, not more. This is a founder-practitioner complement to Martin’s planning ≠ strategy discipline above: both warn against strategy-as-sprawling-plan, from different vantages (execution focus vs. cognitive-work discipline). Per Lifecycle rules this single-source, book-promotional interview does not lift the page’s confidence; its value is a founder/product vantage the page’s consulting- and academic-altitude sources don’t otherwise carry.

Seven sources, a deficit asymmetry, and three decision axes ( McKinsey, August 2026)

The wiki’s most operational treatment of competitive advantage, and the first that tries to make the construct measurable from outside the firm.

The definition, with three qualifiers that each do work. Competitive advantage is “the distinct set of hard-to-replicate assets and operating models that a company creates that earn superior returns over time.” Banholzer tests each: if everyone can do it, “that’s the beauty of capitalism… it’s just par”; if the market doesn’t pay for it, “nobody really cares — unique is not the same as hard-to-replicate superior returns”; and durability is “measured not in months or even a couple of years, but in decade-plus.”

Seven sources, detectable outside-in. Screening the top 5,000 global companies, LaBerge required three properties for a marker to be usable — variation within an industry, specificity enough to act on, and accuracy — and tested them on metals and mining precisely because it is the hard case (“commodity, B2B, not an industry where the majority of players are household names”). Across 180+ companies the markers spread widely rather than clustering. Seven themes held across industries: innovation and IP; customer access and channel; brand and reputation; unique assets and resources; operational excellence; scale; financial strength. Their relative weight shifts by sector — brand and IP carry a far higher premium in medical technology than in automotive components, where “you have to be really, really good to truly make that a differentiated point that your customer is going to pay for.”

The deficit asymmetry — the most actionable finding on this page. Advantages cluster and compound: strength across four of the seven “really outperformed quite a bit.” But the downside is not symmetric. Companies with no deficits did “all right… they didn’t drastically outperform, but they did okay,” whereas “if they have a deficit in one of the seven areas, economic profit goes from slightly positive to massively negative. And if they have two more, it almost doubles again.” The rule Banholzer draws is a direct constraint on strategy: “you have to have an advantage and you have to be good enough on everything else.” He is explicit that this inverts the personal-development intuition: “at a personal level we talk about strength-based growth… we’re finding at an enterprise level, you do need to address your weaknesses because it will hamper your strengths.” His worked case is a chemical company that believed it led on a polymer speciality and was over-delivering there while carrying a route-to-market deficit — misallocating on both sides at once.

Three decision axes. Efficiency — allocate against what customers actually pay for, with failure modes on both sides (over-engineering past diminishing returns; neglecting a capability until “it actually starts to destroy the value you’ve created by building out some other strength”). Durability — how long the edge lasts, whether trends are about to topple it, whether you have “taken your foot off the gas.” Extensibility — can the advantage travel, with the constraint stated plainly (“a cement company can’t usually win successfully moving into software”) and the pay-off being adjacencies where the advantage is valued rather than merely applicable.

Cadence. “The cadence for the companies that do this really well and grow very well is much faster than annually. It’s usually quarterly or monthly” — continuous tracking rather than periodic clean-sheeting, “so that as things start to spike, they see them with enough lead time that they can do something about it.”

Five closing questions, which read as the page’s most portable artifact: do you know what matters to your customers (they often either disagree you’re good at it, or don’t care)? Do you know how good is good enough? Are you gaining or losing ground relative to peers in the plays you’re already in? What trends should you monitor, and continuously or twice a year? And when you make a move, are you using competitive advantage to inform it?

Per Lifecycle rules confidence lifts only 0.88 → 0.90: this is first-party consulting research supporting a service line, the underlying article is not ingested, and the outside-in detection claim — the methodological premise for the whole product — is validated in the episode only by “for the companies that we did know or had experts on… they were defensible.”

Own vs. rent as a strategy decision, taken per capability ( Sequoia, August 2026)

The most reusable framework in a talk that is otherwise unevidenced. Huang’s first step in her sovereign-AI roadmap is explicitly a strategy step, and its central move is refusing the binary: “Sovereign AI isn’t binary. You’re not 0% or 100% sovereign. An important part of the strategy is to draw the lines for which intelligence you want to own and which you’re comfortable outsourcing.”

Four decision factors, applied per capability rather than per company:

Factor	The question
Cost	how large a COGS line item is this?
Speed / latency	”is it a P0 or not?”
Performance	”it used to be that you would choose open weights when you didn’t care about performance” — now potentially inverted in-domain
Proprietary data	is the data you would feed the model genuinely specific to your business?

Her worked example is the crisp one, because the two answers land on opposite sides within the same product: in coding, agents are still mostly rented (“you want strong out-of-the-box performance and latency isn’t a P0”) while tab-autocomplete is mostly owned (“you really really care about speed and these API calls are so frequent that the costs really rack up”). Security companies own for speed plus post-training control; bio companies own for proprietary-data value.

Two further claims bear on this page:

• The battleground moves down a layer. “That battleground is increasingly becoming not just the race for the application layer but the race for the intelligence layer… It’s actually about who can own the intelligence itself and shape better intelligence in the product.” If right, this is a claim about where defensibility sits — and it implies that a firm whose value capture rests on UI, workflow and go-to-market over a rented model has built on a substrate it does not control.
• Legibility as a commercial variable, not a marketing one. “Every single buyer right now is choosing their AI champion. They’re getting the same pitch over and over again, they’re trying to discern which vendor is sophisticated enough to take me to the promised land. Increasingly that means putting out your own research.” The prescription — technical marketing, a separately branded research group, “publishing research with high taste” — treats published research as a purchasing signal in a market where buyers cannot evaluate the product directly.

Confidence unchanged. This is investor content addressed to its own portfolio, with no measurement of any kind behind any of the four factors; per the vendor-source rule it adds a framework without lifting the page’s confidence. Contrast with Banholzer & LaBerge above, whose own-vs-build claims at least gesture at an underlying dataset.

Debates and supersession

• Martin’s Southwest beats hub-and-spoke vs Sinek’s infinite-game. The wiki’s first multi-source productive tension on the strategy page. Resolution filed in strategy-finite-vs-infinite-game synthesis: Sinek operates one layer above the strategy lenses — he asks which game you are in; Martin asks how to win the round you are in. The two are answering different questions about the same case (Southwest), not competing definitions of strategy. No supersession declared.

• Value capture vs value creation. Oberholzer-Gee insists value creation precedes value capture; the implication is that strategy starts with WTP/WTS analysis, not with profitability targeting. This contradicts a long line of strategic-planning practice that starts from financial targets and works backward. No supersession declared — the question of which framing to start from is operational, not theoretical.

• Two “10%” figures from the same research programme, measuring different things. Banholzer & LaBerge report that as few as 10% of organisations have full alignment on what their competitive advantage is. West et al. report that about 10% have data tracking the drivers of the advantage they claim (against ~57–60% who say they are confident). Both are verbal characterisations of McKinsey research the wiki has not ingested. They are compatible, since a firm can be aligned without data or have data without alignment, but they should not be cited interchangeably. (Added 2026-09-21.)

• Where the strategy failure sits: design or passage. Martin and Oberholzer-Gee locate the failure in the strategy (a plan mistaken for a theory of winning; value capture pursued before value creation). West et al. locate the larger differentiator after design: the climbers on McKinsey’s power curve were better at all three stages but “most differentiated by their ability to mobilize”. The two readings meet in Zimmerman’s claim that “it is difficult to mobilize vague strategy”, which makes a good design a precondition for mobilization rather than a rival explanation. The measurement that separates mobilization from execution in McKinsey’s comparison is not described, so this is not settled. (Added 2026-09-21.)

• No supersession events. Two sources in the same year (2022); they’re complementary rather than competing. Future ingest of Playing to Win (Lafley & Martin 2013) and Better, Simpler Strategy (Oberholzer-Gee 2021) would deepen but not retire either framing.

• A practitioner’s planning loop beside a measured regression. Demaree (Onebrief, YC Root Access) describes AI-generated military plans played out in a physics-based war-game simulation, with after-action reviews fed back “again and again until you’ve achieved true superhuman decisions”, and estimates the value of superhuman command and control as “closer to a 10x” increase in force structure, his own estimate. Allen & McDonald measured mid-to-late-2025 frontier models on a business strategy simulation and found them below earlier models and MBA students. The two do not test the same thing: the paper scores single runs of models given structured information; Onebrief iterates plans through a simulator, and the talk reports no results. Whether such a loop corrects an exploitation bias of the kind the paper found is open. (Added 2026-10-03.)

Related concepts

• strategic-foresight — strategy and foresight together: where to play / how to win / how the future might reshape both. Martin’s what-would-have-to-be-true logic-laying is the foresight discipline at the firm scale.
• dynamic-capabilities — Teece’s sense / seize / transform framework is the upstream capability that produces a viable strategy; the value-stick is the discriminator for which seize-and-transform moves are value-creating vs value-redistributing.
• systems-thinking — Sterman’s strategy is a journey, world is feedback converges on Martin’s tweak as you go. Different literatures, same operational stance.
• infinite-game — Sinek/Carse’s frame recasting which game you are in before strategy lenses ask how to win it. Cross-walk filed in strategy-finite-vs-infinite-game synthesis.
• theory-based-view — Felin & Zenger’s TBV is the falsifiability discipline (novel/simple/elegant/falsifiable/generative) applied to Martin’s theory of winning and Sinek’s Just Cause.
• strategic-centering — McGrath’s organizing-principle lens: a third modern reframing alongside Martin (theory of winning) and Oberholzer-Gee (value stick). Centering sits upstream — it chooses the dimension (mission / customer / technology / national ecosystem / friction erasure) along which the firm pursues coherent opportunity sets as industries dematerialize and the classical anchors (Porter / RBV / Blue Ocean) lose force.
• micro-productivity-trap — task-level AI gains failing to translate to firm-level results is a strategy failure: the firm has a plan (deploy AI tools) but not a strategy (a theory of how the deployment creates competitive value).
• dynamic-capabilities — Erginbilgiç’s strategy-as-participation doctrine is the strategy-formation correlate of Teece’s sensing and seizing capabilities: the workshops are the firm’s distributed sensing engine, and the chaotic-divergence rule licenses the seizing-of-options that hierarchical strategy-formation suppresses.

Open questions

• Single primary sources for each lens; deeper primary-source ingest of Martin’s Playing to Win (Lafley & Martin 2013) and Oberholzer-Gee’s Better, Simpler Strategy (2021) would strengthen the concept significantly.
• The value-stick lens is silent on dynamic WTP/WTS — what happens when AI agents (acting as buyer proxies, per Ognibeni 2026 / Price 2026) change whose WTP is being measured? Open question for a future ingest.
• The Southwest case is 50 years old. A canonical 2020s-era strategy case anchoring both lenses simultaneously would strengthen the concept’s currency.

Can an LLM do strategy? A peer-reviewed measurement, and a regression (added 2026-09-09)

[[2026-03-11-allen-mcdonald-how-well-can-ai-do-strategy-simulation-benchmark|Allen & McDonald (Strategy Science, March 2026)]] is the wiki’s first peer-reviewed, multi-model measurement of LLM performance on a task built to have the defining properties of strategy, and it belongs on this page as much for its operational definition of strategy as for its results.

The definition it uses, and why it matters here. Strategic decisions are resource commitments shaping long-term performance, distinguished by five properties: complexity/interdependence, high stakes, multiperiod partially-irreversible commitments, delayed and noisy feedback, and deep uncertainty. That list is this page’s most useful operational test for whether a decision is strategic — and the paper’s argument is that no existing AI benchmark captures any complete subset of it, which is why apparently strategy-adjacent evidence (idea generation, plan evaluation, chess) has told us so little.

The instrument. The Back Bay Battery simulation — eight years allocating a constrained R&D budget between a declining core technology and an emerging one, across three customer segments, while setting prices and forecasting sales. It is the innovator’s-dilemma structure made scorable: “spreading investments too thin leads to mediocrity, ignoring disruption leads to decline, and over-committing too early leads to collapse.” Crucially, participants are never told what to maximise — profit, revenue and future positioning must be weighed, exactly as in real strategy.

The results, and the one that should change how this page is read.

• Early models (GPT-3.5, GPT-4 era) score well below MBA students.
• Late-2024/early-2025 models — o3-mini highest, then o4-mini, Claude Sonnet 4, Gemini 2.0 Flash — exceed the historical MBA average, by timing investment and pricing to achieve profitability and growth.
• Mid-to-late-2025 frontier models (GPT-5, o3, Gemini 2.5 Pro) regress below both the earlier models and the students.

The mechanism is a systematic exploitation bias: high cumulative profit, very low emerging-technology revenue — defend the core, starve the future. Legible in the models’ own reasoning (“We will not invest any R&D into QSC this year, preserving our limited capital for the core business” — Gemini 2.5 Pro), and it survives removing the simulation’s firing rule.

Why this is a strategy finding and not only an AI finding. The failure mode the frontier models fall into is the single most-studied failure mode in this literature — over-exploitation of a profitable declining core — and they fall into it while optimising a defensible local objective. That is the innovator’s dilemma restated as a property of a reasoner rather than of an organisation: the bias does not require misaligned incentives, entrenched middle management or resource-allocation politics to appear. It emerges from a system with none of those things.

And the finding with the widest reach: benchmark transfer breaks. The correlation between simulation performance and GPQA Diamond is positive from GPT-3.5 through o4-mini, then reverses; against LM Arena it plateaus or reverses. The authors’ warning is directed at exactly the inference a strategist would otherwise make: “scholars and developers risk mistaking progress in adjacent domains for progress in strategic decision making.” Anyone selecting a model for judgment work on the strength of headline benchmarks is, on this evidence, reading the wrong instrument.

Read with its limits. The MBA comparison is contextual rather than controlled (students were primed by coursework and worked unsupervised); the composite score is a within-dataset visualisation; and the study measures strategic reasoning over supplied information, not strategic agency — information was fed to the models by a program. See the source page for the full set, including the open question of whether the regression reflects a strategic deficit or a goal-inference one.

Mobilization: the step between strategy and execution (added 2026-09-21)

McKinsey, September 2026 take the question on from where Banholzer & LaBerge left it. Banholzer and LaBerge ask how a firm can see its advantage. This episode asks why a strategy that sees it clearly still fails to change what the organisation does.

The claim. “You can’t simply go from strategy to execution… you need to mobilize.” Mobilization is separate work that needs top-team attention: putting each strategic initiative in an executive’s hands, doing enough planning beneath each choice, reallocating talent, capital and “agentic budget” away from what does not matter, and making sure those commitments reach budgets and plans. In a five-year comparison, companies that rose to the top 20% of economic profit and stayed there were better than companies that fell at design, mobilization and execution alike. They were most differentiated by mobilization. Zimmerman quotes Andrew Marshall: “Your strategy can only influence your competitors if it survives your system first.”

A standard for strategy, needed because “it is difficult to mobilize vague strategy”:

1. A value-creation thesis clear enough that someone who hears only the thesis can make choices that fit it.
2. True choices, each becoming one or a few owned initiatives. “Not just ‘we’re going to digitize’… strategy is not done till it becomes one.”
3. The choices form a portfolio across horizons of value and levels of risk, so the team can judge whether it is leaning into uncertainty enough.
4. Trade-offs named early. “If you’re not naming what’s going to stop… you’re probably not going to stop things.”

Items 2 and 4 are close to Martin’s integrative choices and to Erginbilgiç’s granular strategy, where everyone knows their role. What is new is the unit: a choice is not finished until an executive owns it as an initiative. Zimmerman’s test for the thesis (clear enough to derive one’s own choices from) is close to McGrath’s “permissionless action”, in which a shared strategic center lets people act without waiting for approval.

Two biases, with a fix for each. Anchoring to last year’s budget: use year one of the strategic financial plan, not last year’s budget, as the first draft of next year’s budget. Evaluating bold moves one at a time: loss aversion makes each move look riskier alone, so stack-rank bold moves against everything else being funded. Both are ways of putting into practice the finding that power-curve climbers invest about twice what their average competitor does in their chosen priorities. Incremental budgeting, which moves allocations “by a few points” a year, cannot produce that.

Strategy quality, measured by self-report. The share of executives saying their strategy passes at least four of McKinsey’s ten tests fell from 35% (2010) to 21% (2024). Lacroix blames context (the speed of change, the volume of information) rather than capability. Because the figure is self-assessed, a fall could also reflect more honest self-assessment. The episode does not address that.

AI as a strategy question. Lacroix’s three vectors are productivity “to stay in the game”, innovation to defend and extend profit pools, and deliberate choices to shape new market structures. The first matches the micro-productivity-trap’s claim that productivity gains alone are competed away. The third is a strategy choice in the full sense used on this page.

Confidence unchanged at 0.92. This is first-party consulting research promoting a practice and a book, and none of the underlying studies is ingested.

AI already makes strategic decisions; where do you want to make them? (added 2026-10-11)

Mollick (HBR IdeaCast, Oct 2026) argues that any long-running agent work “involves tons of strategic decisions and judgment calls”, so the question is no longer whether AI makes strategic decisions but “where do you want to make decisions and where not”. A leader can ask for five options with defences, or for a built and tested prototype. Because capability is jagged, which judgments to keep is found by experiment. His two strategic mistakes: anchoring on today’s capability, and trusting vendors or consultants who “if they have successful cases, it’s for GPT-4”. Borrowing Tushman’s ambidexterity, he asks four questions: what have you stopped doing because of AI, what impossible thing are you building, what are you building that does not work yet, and do you have leadership, lab and crowd in place.

A head start, not a moat (added 2026-10-25)

Egeriis (Danske Bank) on Bain’s Winning with AI (Oct 2026): Danske Bank’s CEO does not believe AI currently gives a moat: “with the pace of change… [it] is not something that’s there… you have a few months or six months or 12 months of head start,” in which the bank can win customer meetings, better cost to serve and “that extra bit of market share.” Infrastructure and data products are harder to copy, and “the biggest but also most intangible… competitive advantage is the leadership and culture.” Co-host Andrew Ng adds that an “ephemeral advantage” can buy “quite durable” market share. The opposite view is in the corpus: Bornet expects early adopters to build “compounding intelligence advantages” and “insurmountable learning gaps.” The two are recorded as a contradicts edge between their source pages.

What incumbents hold against AI natives (added 2026-10-25)

Diasio (EY) on WorkLab (Dec 2025): his clients fear most “the competitors and the disruptors that they do not see… It’s the AI natives,” companies of “40 or 50 people” with “multibillion dollar valuations.” Asked what an incumbent brings, he names four things:

• scale, and the ability to deal with “the messiness, which is a corporate reality” that AI-native start-ups dislike;
• experience selling technology into complex enterprises;
• a data flywheel, because “organizational nuance and context is what drives differentiation from just using a general-purpose model”;
• customer reach and contracts.

His example is an unnamed technology conglomerate whose most valuable AI asset turned out to be its game consoles and their community of players. On strategy itself, he asks for “a point of view that you would be comfortable taking to your investors”, written as two to three pages of plain text, before any portfolio of use cases.