Artificial Intelligence and Strategic Decision-Making: Evidence from Entrepreneurs and Investors

Confidence 0.78 · last confirmed 2026-06-08

A peer-reviewed paper by Felipe A. Csaszar (Ross School of Business, University of Michigan), Harsh Ketkar (McCombs School of Business, UT Austin), and Hyunjin Kim (INSEAD), published in Strategy Science 9(4):322–345 (December 2024), in the journal’s Special Issue on the Theory-Based View. It is the wiki’s first academic-strategy treatment of how large language models change the strategic decision-making (SDM) process — distinct from the wiki’s many practitioner/consultancy sources on AI adoption, and the first to engage the theory-based-view of strategy (Felin & Zenger) head-on.

The paper’s vantage is strategy-as-discipline: not “should firms adopt AI” but “what happens to the cognitive machinery of strategy — search, representation, aggregation — when LLMs can generate and evaluate strategies.” It supplies the wiki’s first controlled empirical evidence that LLMs perform at a level comparable to entrepreneurs and investors on realistic, forward-looking strategy tasks.

TL;DR

• Three questions: (1) How might AI-augmented SDM look? (2) How effective are current LLMs at strategic decisions? (3) What are the implications of AI in SDM?
• Method 1 — vignettes: AI-augmented reimaginings of classic SDM tools (Porter’s Five Forces, Devil’s Advocate, etc.) demonstrating the feasibility of automating commonly used strategy tools with current LLMs.
• Method 2 — two empirical studies: (a) with a start-up accelerator, an experiment comparing LLM-generated business plans against those of entrepreneurs seeking VC; (b) using a start-up competition, comparing LLM evaluations of business plans against those of VC and angel investors. Finding: LLMs generate and evaluate entrepreneurial strategies at a level comparable to humans in realistic, high-uncertainty contexts.
• Method 3 — cognitive-process framework: SDM rests on three processes — search, representation, aggregation (Csaszar & Steinberger 2022) — and the paper theorises how AI affects each, linking the SDM outcomes it changes (generation + evaluation of strategies) to competitive advantage.
• Competitive-advantage futures: depending on how AI capability progresses, advantage could stay Ricardian (rooted in unique resources), become Schumpeterian (innovation-driven), or erode entirely if everyone can generate comparable strategies cheaply.
• Theory-Based-View tension: AI both supports and challenges the TBV. Challenges — LLMs are next-word predictors, which may limit forward-looking causal theory creation (Zellweger & Zenger 2023); they rely on past data, risking reproduction of conventional strategies and reduced novelty (Felin & Holweg 2024); their generality may miss firm-specific context. Support — AI could expand the reach of theory-based strategising (e.g. virtual strategy simulations).
• New approaches enabled: AI can enhance the speed, quality, and scale of strategic analysis and enable novel methods like virtual strategy simulations — but ultimate firm-performance impact depends on competitive dynamics, not the tool alone (a strategy-theoretic restatement of the wiki’s micro-productivity-trap / paradox-of-access pattern).

What was actually ingested

Full article (pp. 322–345, ~24 pages incl. both empirical studies, the cognitive-process framework, discussion, and references; converted from the INFORMS PDF via pdftotext). Identity verified against the cover/title page — authors, Strategy Science vol. 9 no. 4, DOI 10.1287/stsc.2024.0190, CC-BY-NC-ND. Note on dating: the wiki uses the journal issue date (December 2024) for the slug/date_published; the paper was accepted Sept 2024 and published online in advance Nov 18, 2024.

Key claims, with detail

The financial-trading analogy bounds the claim

Algorithmic trading now accounts for >78% of trading decisions (SEC 2020) because trading reduces to quantitative, rule-based inputs. SDM is different: it depends on open-ended, qualitative textual inputs and outputs (news, user stories, market reports → strategic plans, memos). Computers historically couldn’t handle this — until LLMs. The paper’s wager is that LLMs cross the threshold that kept SDM “largely inaccessible to AI,” for three reasons: LLMs handle strategy’s textual data; they match/surpass humans on reasoning-heavy professional exams (medicine, law); and their training corpora encode strategy-relevant knowledge (consumer preferences, competitor info).

Empirical: LLMs are human-comparable at generating and evaluating strategy

The two studies are the load-bearing contribution — they move the claim from speculation to evidence in realistic strategy environments (actual accelerator applicants; actual VC/angel evaluations), complementing prior LLM-business studies that examined either non-SDM tasks (writing, creativity — Boussioux, Girotra, Noy & Zhang) or stylised SDM (internal-consulting tasks — Dell’Acqua et al.; small-business mentoring — Otis et al.). Csaszar et al. push into forward-looking entrepreneurial strategy under genuine uncertainty.

The search / representation / aggregation framework

The paper’s general contribution is a framework tying AI use in SDM to firm outcomes via the three cognitive processes it changes:

• Search — generating the candidate set of strategies (AI widens or narrows it);
• Representation — how the strategic problem is framed/modelled (AI can increase or flatten representational complexity);
• Aggregation — combining judgments into a decision (AI changes who/what aggregates).

These map onto the quality and heterogeneity of firms’ strategies, the complexity of strategic representations, the role of business experiments, and the importance of theories in strategy — the bridge to the TBV.

Competitive advantage: Ricardian → Schumpeterian → or gone

A sharp, quotable claim for the strategy page: as AI capability progresses, the nature of competitive advantage may change — it “could remain Ricardian (based on unique resources), become Schumpeterian (driven by innovation), or potentially cease to exist altogether.” If AI commoditises strategy generation, advantage migrates to complementary assets (proprietary data, unique execution) — convergent with the paradox of access the wiki holds via Anand-Wu (“because everyone can use it, it becomes harder to capture value with it”).

Dynamic-capabilities reading (Warner & Wäger)

Lightly within the W&W lens: this paper touches digital-sensing/digital-scenario-planning — AI-augmented SDM is precisely “analyzing scouted signals; interpreting digital future scenarios; formulating digital strategies,” and the paper’s virtual strategy simulations are a direct instance of AI-enabled scenario planning. (Roles inherited: cso, cdo, strategy-consultant.) The paper sits mostly outside the transformation-process cells — it is a strategy-theory contribution about AI as a decision tool, not an org-transformation case — so it carries a single, defensible tag rather than being stretched across the vocabulary.

Cross-source positioning

Source	Construct	What Csaszar et al. adds
Dell’Acqua et al.	Jagged frontier of AI in knowledge work (consulting RCT)	Extends the augmentation evidence into strategic decision-making under genuine uncertainty, not stylised consulting tasks.
Boussioux et al.	LLMs generate/evaluate ideas vs crowds	Parallel finding in a strategy setting: LLM-vs-entrepreneur generation and LLM-vs-investor evaluation.
Carroll & Sørensen / theory-based-view	Strategy-as-theory; analogical reasoning	The TBV-special-issue companion: how AI supports and challenges the theory-based view (causal-theory creation vs past-data conventionalism).
automation-vs-augmentation	Augment vs automate	SDM as an augmentation frontier — AI augments the strategist’s search/representation/aggregation rather than replacing the decision.
micro-productivity-trap / Anand-Wu paradox of access	Task gains ≠ firm value; access erodes advantage	”Ultimate impact on firm performance will depend on competitive dynamics” — the strategy-theoretic version of the same caution.

Linked entities and concepts

Existing pages (touched or referenced):

• theory-based-view — heavy: the paper is in the TBV special issue and engages Felin & Zenger directly.
• strategy — heavy: AI in SDM; the Ricardian/Schumpeterian/erased competitive-advantage futures.
• automation-vs-augmentation — moderate: AI-augmented SDM as an augmentation frontier.
• analogical-reasoning — light: search/representation as the cognitive substrate strategy reasoning draws on.
• enterprise-ai-adoption — light: a strategy-discipline complement to the adoption literature.

Dangling (single-source mention, deferred per CLAUDE.md author-entity promotion rule):

• Felipe A. Csaszar — lead author (Professor of Strategy, Michigan Ross). Promote on second-source mention.
• Harsh Ketkar — co-author (Assistant Professor, UT Austin McCombs).
• Hyunjin Kim — co-author (Assistant Professor of Strategy, INSEAD).
• Felin & Zenger, Zellweger, Felin & Holweg — cited theory-based-view scholars (not authors here).

Source-quality flag

• Strengths: peer-reviewed (Strategy Science, INFORMS); two empirical studies in realistic strategy settings (accelerator + competition) rather than toy tasks; a clean cognitive-process framework; the wiki’s first academic-strategy anchor on AI + SDM and first to engage the theory-based view directly. Open-access (CC-BY-NC-ND).
• Caveats: studies use 2023–24-vintage LLMs (capability has since advanced — both raising and complicating the findings); “comparable to entrepreneurs and investors” is a level-of-performance claim, not a firm-performance claim (the authors are explicit that firm impact depends on competitive dynamics); generalisability beyond entrepreneurial/VC settings is untested.
• Per CLAUDE.md §Lifecycle: baseline 0.70 + 0.05 (peer-reviewed) + dual-study empirical strength → confidence 0.78.

Open questions for the wiki

• Re-test with frontier models. The empirical comparisons predate 2025–26 frontier models; a replication would sharpen the “comparable to humans” claim.
• Does the Ricardian→Schumpeterian→erased trichotomy get a concept-page home? It is a candidate addition to strategy / theory-based-view as AI-era competitive-advantage scenarios.
• Second Csaszar / Kim source would promote the authors from dangling to entities and strengthen the wiki’s academic-strategy spine.