In 2024, a Strategy Science experiment had AI evaluate the same business plans that experienced VCs and angel investors had scored. On financial analysis, AI tracked the humans well — a correlation of r=0.57. On innovation assessment, the number was r=0.21. On market opportunity, r=0.28 [1]. The dimensions where AI is weakest are exactly the dimensions where judgment lives.
I have been thinking about that finding for a while, because it does not say what most people seem to want it to say. The argument is not that AI is bad at strategy. AI is, in fact, very good at strategy. The same paper found that LLM-generated business plans were rated 0.14 standard deviations higher than the ones the entrepreneurs wrote themselves. The argument is that AI is good at the parts of strategy that have always scaled well — the parts where pattern recognition and analytical decomposition do the work — and weak at the parts where someone has to take a position that might not hold.
The question used to be “what should we do?” Now it is “what can we actually make happen?”
The people who built things and sold them do not have a credential advantage in answering that. They have a structural one. Grounded judgment — the kind forged by direct exposure to consequences — is the scarce complement to abundant AI-generated analysis. Not legacy ballast. The filter.
The strategy machine runs faster than anyone can steer
The same Csaszar paper noted, almost in passing, that LLMs “risk reproducing conventional and widely accepted strategies, potentially reducing novelty rather than generating genuinely innovative approaches” [4]. McKinsey, looking at how organizations get strategy wrong from a different angle entirely, observed the same shape in human form: “A vague strategic commitment replaced having a real strategy” — companies dressing up financial targets as strategy and calling it done [11]. AI is excellent at producing the same dressed-up shape, faster.
So the strategic machine now hums. The deck looks beautiful. The slides have data. Internal coherence has never been cheaper.
Which leaves the harder question untouched: which of these analytically tidy options actually survives contact with a customer, a payroll cycle, or a sales team that has to make next quarter’s number?
The gap between planning and execution is widening, not closing
PMI’s 2025 research, drawn from more than 5,800 project professionals and 300 senior executives, gives the disconnect a number. Thirty-five percent of executives name “a disconnect between planning and execution” as the single largest barrier to reinvention — ranked above any other factor they were offered. Only half of projects deliver value exceeding the effort they consume. Thirteen percent fail outright [2].
The AI-era numbers are worse. MIT Sloan’s 2025 review of enterprise AI pilots found that 95% produced little to no measurable business impact. BCG’s 2025 global AI survey came at the same finding from the value side: roughly 5% of companies are realizing significant value from AI, and close to 60% report little or none, despite active initiatives [12].
Then there is the confidence gap. Seventy-eight percent of CEOs report confidence in guiding AI adoption. Twenty-eight percent of mid-level managers do [7].
Same companies. Same projects. Fifty-point gap.
The CEO is reading the strategy deck. The manager is staring at the workflow that has to absorb it.
Someone with consequence exposure still has to design the feedback loops in the first place — and the 95% figure tells you how well the design-it-from-the-deck approach is working.
That difference in vantage point matters when the plan meets the work.
What you learn when the consequences are yours
Nicolai Foss and Peter Klein, building on the economist Frank Knight, define entrepreneurial judgment as decision-making under genuine uncertainty where losses are personal and uninsurable [3]. This is not a personality trait. It is a structural condition. The entrepreneur cannot diversify the consequence of a bad strategic call across a portfolio of fiscal quarters and reorgs. The bill arrives on a date they can name.
Nassim Taleb sharpens the point in Skin in the Game: skin in the game “naturally decreases with scale and distance.” The further removed decision-makers are from the consequences of their decisions, the less reliable the judgment becomes. In systems where everyone involved has something to lose, bad ideas get weeded out quickly. In systems where nobody does, they persist [5].
This is the observation I keep returning to: you only fully understand a system when the consequences of your decisions are direct and immediate.
Selling is the cleanest case I know. Pricing decisions are not theoretical — they determine whether a customer signs. Operational choices are not incremental — they determine whether the thing actually ships. A proposition that looks logically sound can still fail because the buyer’s procurement window is closed, the champion just left, or the buyer simply does not believe you.
A chef who has only cooked in test kitchens knows recipes. A chef who has run their own restaurant knows what happens at 7:42 p.m. on a Saturday when the line is six deep and the dishwasher quit at noon. Same knowledge of ingredients. A different feedback loop. The food is different when rent depends on it.
That is the Foss-Klein point in plain form: judgment under personal, uninsurable risk operates on a different feedback loop than judgment buffered by salary and process — not because the analyst is less capable, but because the system corrects the entrepreneur faster.
McDonald’s spent three years learning what franchisees knew on day one
In 2019, McDonald’s acquired the AI startup Apprente and partnered with IBM to deploy automated voice ordering at drive-throughs. By mid-2024, the system was live at more than 100 U.S. locations across 13 states. Corporate strategy approved it. The efficiency math worked. The roadmap was clean.
Customers got nine sweet teas they did not order. They got bacon on their ice cream. Orders bled into adjacent lanes. The system did not handle accents, background noise, or the actual sound of a parking lot at lunchtime.
The franchisees — the people whose monthly P&L moves with every wrong order — said it on day one. One of them, when McDonald’s finally pulled the plug in June 2024, put it cleanly: the technology would “have to be at least 95% accurate and save money over having a person in the drive through, and the way it is designed now, does neither” [V1].
The strategy persisted for three years past the point its consequence-bearers had invalidated it. The mechanism is not mysterious. A reviewer who would normally push back on a thin recommendation tends not to push back on a confidently explained one. When AI provides confident explanations, review teams accept its recommendations 87% of the time — even when those recommendations contain systematic bias [8]. Confidence substitutes for consequence-tested validity. The deck keeps the project alive longer than the drive-through window can defend it.
The gap between the strategy deck and the drive-through window tends to be measured in years, not insights.
What if the salespeople are just anchoring to yesterday?
Three objections deserve serious answers.
The first is survivorship bias. Most entrepreneurs fail. The ones who survive may be telling themselves a story about earned judgment when the truth was timing, or luck, or a kind market. “Trust the people who have sold things” is, I am aware, the kind of advice a consultant would give to sell consulting.
The redirect is mechanical, not heroic. The argument here is not that founders are wise. It is that consequence exposure produces a structurally different feedback loop — one where bad decisions cost the decider directly and quickly. That loop exists whether the entrepreneur ultimately succeeds or fails. The mechanism is the feedback, not the survival.
The second objection: AI is improving fast. The Csaszar correlations are a 2024 snapshot. The same kind of argument was once made about chess and protein folding, and it did not hold.
That risk is real. I am making no claim of permanence here. The structural advantage today is incentive alignment. AI can ingest the customer transcript, the cash flow projection, and the operations report. It cannot lose its house when the strategy fails. Whether that changes for agentic AI with capital allocation authority is a different essay; for now, the gap remains a question of who bears the cost of being wrong.
The third objection is the one I find hardest to dismiss. Experience itself produces distortions. Anchoring on yesterday’s market. Pattern calcification. Confusing familiarity with understanding. There is an inverted U between information and decision quality — initial exposure improves judgment, extended narrow exposure can degrade it — and it applies to consequence-exposed veterans as much as to anyone [6].
So apply it symmetrically. The same inverted U applies to analytical strategists who have never tested a recommendation in market. Their confidence also exceeds their accuracy. The argument is not that execution experience automatically produces good strategy. It is that the absence of any consequence exposure in the strategic decision creates a systematic blind spot — and AI now floods that blind spot with internally coherent options at a pace no review process was designed for.
AI raises the floor of strategic quality. The ceiling still needs grounded judgment. The best human strategies already incorporate execution feasibility, market sensing, and stakeholder dynamics that AI cannot replicate from training data [9]. AI is not the enemy of good strategy. It is the reason the filter now matters more than the funnel.
The Consequence Filter: how to ground strategy before it ships
Five steps connect the strategic choice to the people who must execute it.
Map the consequence distance. Count the layers between the person making the strategic call and the person who will bear its cost — the salesperson with a quota, the operator with a service-level commitment, the franchisee with rent. Distance is the variable that predicts how long a broken strategy will live.
Stress-test against execution constraints before analytical merit. For every option, name the three hardest execution requirements and test each against current capability — can the sales team actually sell this; can operations deliver it; can the organization absorb the change on the timeline you have. Analytical merit means nothing if these answers are no [2].
Install consequence proxies where direct exposure is missing. Pre-mortems led by execution-experienced people with binding kill criteria. Time spent by strategists in customer-facing roles as participants, not observers. Decision rights structured so the consequence-bearer can stop the project, not just file a report. Without binding authority, the proxy becomes a performance — a meeting that happens before the meeting where the real decision is made [5].
Define the question before generating options. Use AI to expand the option space within defined execution constraints, not to define what the constraints are. The HBS research is direct on this: sequencing matters as much as the recommendations themselves. Generative-first produces more variety but lower quality [10]. MIT Sloan adds a useful inversion — when AI is used to argue against a preferred option rather than to generate one, judgment sharpens rather than dulls [14].
Close the loop in 90 days, not 12 months. For each strategic choice, name the three to five assumptions it depends on, and define a measurable execution outcome that would validate or invalidate each within a quarter. If your strategy-to-execution feedback cadence is annual, you are running on a timeline AI has already obsoleted [4].
Bain’s research on founder-led companies makes the structural argument concretely. From 1990 to 2014, founder-led S&P 500 companies outperformed by 3.1x in total shareholder return. Over the most recent decade, founder-led public companies returned 25% per year against the S&P’s 14%. The trait is not founder mystique. Bain calls it the Founder’s Mentality: insurgent mission, front-line obsession, owner’s mindset [V2]. All three are structural — tight feedback loops, proximity to execution reality, personal accountability — and all three are reproducible by organizations that decide to install them.
Three things worth doing this quarter.
Find the three strategic decisions in your organization where the decision-maker sits furthest from the consequence-bearer. Those are your highest-risk strategies — not the ones that look weakest analytically.
Stop evaluating AI-generated options on internal coherence first. Hand the option to someone with direct execution exposure and ask whether they would bet their own resources on it. Listen to the answer before you adjust the slide.
Build one consequence proxy this quarter that has the authority to stop a project, not just comment on one.
The options have never been more abundant. The filter has never mattered more.
References
[1] Csaszar, F. A., Ketkar, H., & Kim, H. (2024). Artificial Intelligence and Strategic Decision-Making: Evidence from Entrepreneurs and Investors. Strategy Science (INFORMS). https://pubsonline.informs.org/doi/10.1287/stsc.2024.0190
[2] Project Management Institute. (2025). New PMI Research Reveals Strategy-Execution Gap Is Undermining Transformation. https://www.pmi.org/about/press-media/2025/new-pmi-research-reveals-strategy-execution-gap-is-undermining-transformation-and-how-to-close-it
[3] Foss, N. J., & Klein, P. G. (2012). Organizing Entrepreneurial Judgment: A New Approach to the Firm. Cambridge University Press.
[4] Lane, J. N., & Jachimowicz, J. (2026). AI Trends for 2026: Building Change Fitness and Balancing Trade-Offs. Harvard Business School Working Knowledge. https://www.library.hbs.edu/working-knowledge/ai-trends-for-2026-building-change-fitness-and-balancing-trade-offs
[5] Taleb, N. N. (2018). Skin in the Game: Hidden Asymmetries in Daily Life. Random House.
[6] Matatihi. (2025). When Analysis Fails: Understanding the Paradox of Expert Decision-Making. https://matatihi.nz/sustainability-insights/2025/1/27/when-analysis-fails-the-paradox-of-expert-decision-making
[7] Turning Data Into Wisdom. (2025). What 2025 Taught Us About Data, AI, and Decision-Making. https://www.turningdataintowisdom.com/what-2025-taught-us-about-data-ai-and-decision-making/
[8] Same as [7], on the transparency paradox in AI-assisted review.
[9] Csaszar et al. (2024) and Lane & Jachimowicz (2026); on AI raising the floor of strategic quality without raising the ceiling.
[10] Lane, J. N., & Jachimowicz, J. (2026), HBS Working Knowledge — on the sequencing and format of AI tools shaping outcomes.
[11] McKinsey & Company. (2024). Getting strategy wrong—and how to do it right instead. https://www.mckinsey.com/capabilities/strategy-and-corporate-finance/our-insights/getting-strategy-wrong-and-how-to-do-it-right-instead
[12] Boston Consulting Group. (2025). Global AI Survey 2025. https://www.bcg.com/publications/2025/global-ai-survey; MIT Sloan / UpStrategyLab synthesis: https://www.upstrategylab.com/blog/ai-exposing-gap-between-strategy-execution
[14] MIT Sloan Management Review. (2025). When Not to Use AI. https://sloanreview.mit.edu/article/when-not-to-use-ai/
[V1] McDonald’s / IBM AI drive-through partnership, 2019–2024. CNBC (June 17, 2024): https://www.cnbc.com/2024/06/17/mcdonalds-to-end-ibm-ai-drive-thru-test.html
[V2] Zook, C. & Allen, J., Bain & Company. Founder-Led Companies Outperform the Rest — Here’s Why. Harvard Business Review (2016) and Bain updates through 2025. https://hbr.org/2016/03/founder-led-companies-outperform-the-rest-heres-why