ICYMI
Double (AI) Agents. What intelligence services can teach companies about AI agents.
The Centaur Age Is Here. Which of Yours Are Lame? Human-AI teams are proliferating across the enterprise. Most of them are making things worse.
The Plausibility Crisis. As AI-generated output becomes more abundant, convincing and unreliable, senior executives are becoming the last line of defence.
In 2000, David X. Li published a paper introducing the Gaussian copula function for modelling default correlation in credit portfolios. Within five years, the dominant rating agencies controlling the structured finance market (Moody’s, S&P and Fitch) had adopted variants of the same formula. When subprime defaults exceeded the shared model family’s assumptions, their ratings failed in the same way. The Financial Crisis Inquiry Commission concluded that the agencies were “key enablers of the financial meltdown.”
Now, a similarly concentrated enterprise LLM market creates related risks for corporate strategy. Reliance on a few leading AI models could push firms toward convergent strategies, reducing competitive differentiation and potentially creating new systemic risks.
Strategic Monoculture
Strategy itself, in the form of board memos, M&A theses, competitor analyses and similar documents, is now drafted through a small number of foundation models with overlapping training datasets, similar default framings and the same structural reflexes. Anthropic, OpenAI and Google together accounted for 88% of enterprise LLM API usage in December 2025.
The compression is substantive as well as stylistic. Romasanta, Thomas and Levina tested six frontier models across more than 15,000 strategic dilemmas (Harvard Business Review, March 2026): approximately 96% of responses chose differentiation over cost leadership, 93% chose augmentation over automation, and prompt engineering shifted the bias by less than 2%. The models sound alike and they also recommend alike.
The pattern extends to the humans using them. Dell’Acqua et al.’s field experiment with 758 BCG consultants found that AI-assisted work was roughly 40% higher quality, but group-level idea diversity fell by around 41%. Meincke et al. found a similar pattern in product innovation: GPT-4-generated ideas scored higher on average purchase intent but were less novel and more similar to each other. Switching vendors does not help: Wenger and Kenett (2025) found that LLM responses are far more similar to each other than human responses are to each other. Cross-model convergence is structural.
We might term this Strategic Monoculture: the convergence of corporate thinking around the priors of a small number of foundation models. Consulting firms have driven convergence for decades through application of frameworks and best practices (DiMaggio & Powell, 1983; Abrahamson, 1996). But that convergence was slow and observable, whilst LLM-driven convergence moves faster and is harder to track.
Strategy Matters More Than Ever
This would matter less if execution were still a strong moat. But AI is devaluing execution relative to strategy as a source of competitive advantage. Stanford HAI’s 2025 AI Index reports that the cost of querying a GPT-3.5-level model fell from $20.00 to $0.07 per million tokens in 18 months (a 280× collapse). Agentic workflows are automating supply chains. Operational AI is spreading across every sector.
Defences That Don’t Work
Three defences are frequently raised but flawed in practice:
“Better prompts will fix it”. Lee, Kizilcec et al. (COLM 2025) tested 30,000 admissions essays and found that prompting did not alleviate homogenization. The style changes, but the worldview doesn’t.
“A human in the loop will fix it”. A human in the loop will not fix the problem by itself if the human starts from, and is anchored by, the model’s draft.
“Use external advisors”. External advisors can’t address the problem unless their AI stack, data sources, and challenge process are genuinely differentiated. McKinsey’s Lilli runs on Cohere and OpenAI, BCG’s tools run on GPT-4o, and Bain put 13,000 consultants on ChatGPT Enterprise in August 2024. The fresh perspective is running on the same engine.
Rebuilding Differentiated Thinking
Leaders can pursue three routes to help sustain or rebuild original strategic thinking:
Diversify the cognitive supply chain. No firm would sole-source a critical component in its physical supply chain. The same logic applies to the cognitive one. Different AI models teamed with different humans should serve different strategic functions so that cross-functional debate brings different priors to the table. Importantly, diversity should be sought in the AI models, by going beyond the usual suspects and by adding company-specific fine-tuning. Goldman Sachs is moving in this direction, with its GS AI Platform integrating multiple approved models.
Make AI provenance a disclosure requirement. Every material strategy paper should state which model, what data, which prompts, and at what stage AI was introduced. It’s the cognitive equivalent of declaring a conflict of interest. Any firm that would require an external advisor to disclose a conflict should require the same from an AI-assisted strategy paper.
Redesign the decision process. Most firms have bolted AI onto an unchanged workflow. The workflow itself needs rebuilding around three phases: diverge, where AI and humans together explore the broadest option set; challenge, where different models and teams stress-test the emerging consensus; and commit, where human judgement alone owns the narrowing. Walmart’s agentic AI framework uses a co-pilot model and explicitly distinguishes between actions suitable for autonomous execution and areas where human oversight or approval remains essential.
Four Tests for the Board to Request
Boards can apply a few simple tests to ascertain whether the organisations is thinking strategically about strategic thinking.
The convergence test. Run your competitors’ annual reports through the same model your strategy team uses. How different is the output from your own board paper?
The model test. Which of our recent material decisions would have been different if we’d used a different model, or no model at all?
The independence test. Do our layers of strategic challenge (e.g., management, non-executives, external advisors) use different cognitive infrastructure?
The diversity test. Have we designed our AI-assisted workflows to produce more diverse thinking than either humans or models would produce alone, or have we just made the same thinking faster?
The Lesson of the Copula
The lesson of the 2008 rating-agency failures was not that the ratings were wrong. It was that they were all wrong in the same way. The question for firms in 2026 is whether their strategy has the same property.
Footnotes & Sources
• Li, D. X., “On Default Correlation: A Copula Function Approach,” Journal of Fixed Income 9(4): 43–54, March 2000. The Gaussian copula formula for modelling default correlation, subsequently adopted across the structured finance industry. Link
• Financial Crisis Inquiry Commission, The Financial Crisis Inquiry Report, US Government Printing Office, January 2011. “The three credit rating agencies were key enablers of the financial meltdown.” Link
• Menlo Ventures, 2025: The State of Generative AI in the Enterprise, December 2025. Anthropic 40%, OpenAI 27%, Google 21%; combined 88% of enterprise LLM API usage. (Methodology caveat: Menlo is an Anthropic investor.) Link
• Romasanta, A., Thomas, L. D. W. & Levina, N., “Researchers Asked LLMs for Strategic Advice. They Got ‘Trendslop’ in Return,” Harvard Business Review, March 2026. Six frontier models tested across 15,000+ simulated strategic dilemmas; ~96% chose differentiation, ~93% augmentation, prompt engineering shifted bias by <2%. Link
• Dell’Acqua, F., McFowland, E., Mollick, E. et al., “Navigating the Jagged Technological Frontier,” HBS Working Paper 24-013, September 2023; later published in Organization Science, 2026. Field experiment with 758 BCG consultants; AI-assisted work around 40% higher quality, while group-level idea diversity fell by 41%. Link
• Meincke, L., Girotra, K., Nave, G., Terwiesch, C. & Ulrich, K., “Using Large Language Models for Idea Generation in Innovation,” The Wharton School Research Paper, SSRN, September 2024. GPT-4-generated product ideas scored higher on average purchase intent but were less novel and more pairwise-similar than human ideas. Link
• Wenger, E. & Kenett, Y., “We’re Different, We’re the Same: Creative Homogeneity Across LLMs,” arXiv:2501.19361, January 2025. LLM responses more similar to each other than human responses are to each other. Link
• DiMaggio, P. J. & Powell, W. W., “The Iron Cage Revisited: Institutional Isomorphism and Collective Rationality in Organizational Fields,” American Sociological Review 48(2): 147–160, 1983.
• Abrahamson, E., “Management Fashion,” Academy of Management Review 21(1): 254–285, 1996.
• Stanford HAI, 2025 AI Index Report. Cost of querying a GPT-3.5-level model on MMLU fell from $20.00 per million tokens in November 2022 to $0.07 by October 2024. Link
• Lee, J., Kizilcec, R. et al., “Poor Alignment and Steerability of Large Language Models: Evidence from College Admission Essays,” COLM 2025, arXiv:2503.20062. 30,000 admissions essays; prompting did not alleviate homogenization. Link
• VentureBeat, “Consulting giant McKinsey unveils its own generative AI tool for employees: Lilli,” August 2023. McKinsey is “LLM agnostic”; Cohere and OpenAI on Microsoft Azure. Link
• Business Insider interviews with Alicia Pittman and Scott Wilder (BCG), September 2025. 90% of BCG’s 33,000 employees use AI.
• Bain & Company corporate announcements, 21 February 2023 (OpenAI services alliance) and 17 October 2024 (expanded partnership). WSJ, October 2024: 13,000 Bain consultants licensed for ChatGPT Enterprise (from August 2024).
• Goldman Sachs, GS AI Platform. IMD AI Maturity Report (November 2025) and AIX Expert Network (July 2024): the platform integrates multiple LLMs (OpenAI, Google, Anthropic and open-source), enabling employees to select the most suitable model per task. CFO Denis Coleman described OneGS 3.0 as “positioning AI as a foundational capability rather than a standalone tool” (Goldman Global Conference, 2025).
• Walmart, “Inside Walmart’s Strategy for Building an Agentic Future,” corporate blog, 29 May 2025. Describes a “co-pilot model, with humans and AI working as a team,” and says Walmart is evaluating which actions are suited to autonomous agent execution and where human oversight or approval remains essential. Link