Does AI improve strategic judgement?

Prediction, option generation and the gap between rated plans and realised performance

This retrospective review considers source versions available by 1 September 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: strategic judgement; prediction; decision quality; entrepreneurship; AI assistance

1. The outcome that strategy research must explain

A strategy document can be coherent, well rated and produced quickly while leaving the quality of the underlying commitment unresolved. Strategic choice includes deciding what objective to pursue, which alternatives deserve consideration and what resources to commit under uncertainty. AI can contribute to each operation. Evidence that it performs one of them well cannot establish that the complete sequence improves organisational outcomes.

The central question is whether AI-supported analysis improves strategic judgement and how that claim should be tested. This review distinguishes prediction, option generation, evaluation and commitment. It focuses on research relevant to marketing strategy without treating a marketing plan as a generic writing task. A plan's factual adequacy, internal consistency and persuasiveness are assessable properties; competitive advantage and realised performance require additional observations.

The distinction is visible in the emerging strategy literature. Csaszar, Ketkar and Kim compare AI-generated and evaluated strategies with human work in entrepreneurial settings, reporting performance comparable to entrepreneurs and investors on the assessed tasks [1]. Those comparisons establish capability within a specified assessment. They leave open how strategic choices are implemented and how competitors respond after adoption.

2. Method and source boundaries

Targeted searches conducted on 21 October 2026 combined strategic decision-making, generative AI, entrepreneurial performance, prediction, exploration and cognitive bias. 18 primary research or theoretical sources were selected from publisher pages, author manuscripts and working-paper records. They include experiments on strategic evaluation, professional tasks and business advice, alongside established conceptual distinctions. This is a narrative synthesis, not a systematic review or a meta-analysis.

The comparison records whether a study evaluates a document, a decision process or an economic outcome. It also distinguishes random assignment of tool access from the amount or manner of subsequent use. Users who choose to interact intensively with a tool may differ from those who do not. A reported association between engagement and performance cannot automatically be treated as the causal effect of a prescribed usage dose.

Several sources were available only as abstracts or publication records. Their discussion is limited accordingly. The source-access register identifies these cases. No replication, fresh participant study, quantitative bias adjustment or common effect-size conversion is reported. The figures organise the distinctions needed to interpret the existing evidence and contain no invented performance estimates.

3. Prediction, explanation and choice

3.1 Different analytical targets

Shmueli distinguishes predictive and explanatory modelling and the different decisions involved in building and evaluating them [5]. A model can forecast an observed outcome without identifying the causal effect of a proposed intervention. For strategic decisions this matters because a business usually wishes to change an outcome, not only predict which outcome will occur under existing behaviour. A correlation-based forecast can be useful while leaving the intervention question unanswered.

Kleinberg and colleagues describe policy problems in which better prediction is itself a useful input to decision-making [6]. Their distinction helps identify where an AI prediction can add value, but the prediction still operates within a decision problem containing costs and objectives. An accurate forecast of likely response does not specify how much the organisation should spend to obtain it or whose interests should determine the objective.

3.2 Objectives and constraints

An optimisation task begins with an objective function and a feasible set. Strategic judgement often includes deciding which objective matters and whether the apparent constraints can be changed. If evaluation supplies a fixed objective to the model, it tests performance conditional on that objective. It does not assess whether choosing that objective was itself a good strategic decision.

A marketing example makes the measurement distinction concrete. Predicting which existing customers will purchase differs from estimating which customers will purchase because of an intervention. Neither result alone settles whether the organisation should prioritise short-term sales, future demand or another outcome. Those are definitions of the decision problem. They require evidence and authority beyond the score produced by a predictive model.

3.3 Bounded search

Simon's behavioural model of rational choice provides historical context for decisions made under limited information and computational capacity [8]. This reference is used conceptually, without extracting a numerical prediction about modern AI. Expanding analytical capacity changes what can be considered, but it does not make every possible alternative observable or every future consequence calculable.

The number of generated options is therefore a process count. Its interpretation depends on whether the options differ substantively, remain feasible and receive enough scrutiny. Ten variations of one assumption do not necessarily represent a broader strategic search than two genuinely different explanations. An evaluation needs a definition of search coverage rather than assuming that output volume supplies one.

4. Direct evidence on strategic tasks

4.1 Generation and evaluation of plans

Csaszar and colleagues use entrepreneurial and investor settings to examine AI's role in generation and evaluation, and discuss search, representation and aggregation [1]. Their work makes strategy a set of cognitive operations that can be studied separately. Comparability on assessed plans does not establish equivalence across all strategy work, especially when the outcome is a judgement about a proposal rather than the realised return from implementing it.

In the plan-generation comparison, the authors recruited 250 US-based evaluators with investment experience and compared business plans under a preregistered procedure [1]. The evaluators' experience makes their judgements relevant to proposal assessment, while still leaving implementation unobserved.

A plan-rating experiment can isolate whether evaluators prefer one set of texts under a common assessment. Its interpretation still depends on the information supplied to each producer. If one condition receives a clearer problem statement, a richer evidence packet or more revision opportunities, the treatment combines those inputs with the generation method. Fair comparison requires specifying the whole assistance arrangement, including human editing.

4.2 Bias and engagement

Cristofaro, Giardino and Augier report a laboratory market-entry study involving 273 business professionals. Their abstract describes a curved relationship between use and bias, with intermediate engagement associated with stronger processing and lower biases, and differences related to critical engagement [2]. Full-text access was restricted. The abstract does not establish that prescribing a universal middle level of usage will improve strategic judgement in every setting.

The distinction between assignment and engagement is particularly important for this result. Tool availability can be randomised, while the user's chosen behaviour within that condition remains a response. To infer the causal benefit of a specific engagement pattern, a study must isolate that pattern or justify its identification strategy. Describing an observed pattern is useful without converting it into an unsupported operating rule.

4.3 Innovation proposals and multiple quality dimensions

Boussioux and colleagues compare human-crowd and human-guided AI solutions to a circular-economy challenge. Their results distinguish novelty from strategic viability, value and overall quality, with outcomes depending on the search arrangement [11]. This is evidence about evaluated early-stage proposals. It does not show that a high-rated solution will survive implementation, competitor imitation or customer adoption.

The search manipulation also differs from ordinary revision. In differentiated search, the model receives instructions to depart from previous solutions; independent search starts separate instances from the initial prompt [11]. The comparison concerns how a candidate set is generated, rather than an assumption that every additional conversational turn improves a proposal.

The separation of quality dimensions also affects aggregation. A composite score can conceal trade-offs if novelty rises while feasibility falls, or the reverse. Weighting those dimensions is a substantive choice. An assessor should be able to inspect the components and understand whether a higher overall score reflects the attribute the decision-maker actually values. Otherwise, a comparison can reward a different objective from the one under discussion.

5. Professional-task evidence and its transfer

Dell'Acqua and colleagues studied 758 consultants and found that AI assistance improved speed, task completion and quality on selected tasks within the assessed capability frontier, while correctness worsened on a selected task outside it [9]. The same broad occupational setting therefore contained different responses to assistance. A label such as strategic work is too coarse to predict the result of a particular assignment.

Noy and Zhang's published experiment reports productivity and quality improvements in professional writing [10]. It supports the proposition that assistance can improve an evaluated deliverable under specified conditions. Its outcome is not a longitudinal estimate of strategic success. A persuasive, polished plan may be valuable communication, but the commercial consequences of its recommendation require a different evaluation.

These adjacent experiments also clarify why time saved is not the same outcome as decision quality. A tool can reduce drafting time while leaving evidence collection unchanged. Another can increase time spent comparing options while improving the final decision. If the evaluation compresses both into one productivity score, it needs an explicit rule connecting time, quality and the business objective. Faster work alone does not establish that rule.

6. Economic outcomes and heterogeneity

Otis and colleagues randomised access to a GPT-4 business assistant among Kenyan entrepreneurs. Their published abstract reports no statistically established average effect on revenues and profits and materially different effects by baseline performance [12]. This is closer to an economic endpoint than a plan-rating task. It also shows why an average treatment effect can be insufficient for understanding who benefits from advice.

Subgroup findings require careful reading. Baseline performance is measured before treatment, whereas later implementation behaviour can be affected by treatment. Those variables play different roles in interpretation. A result suggesting that users implemented advice differently does not imply that identical advice would have identical consequences if imposed on everyone. The organisation and the user's capacity to act are part of the intervention setting.

The field result should not be used to reverse every positive laboratory finding. It concerns different participants, problems and outcomes. A short task can establish a capability that contributes to value without measuring the full process that produces that value. A field intervention can fail to produce an average economic gain despite useful components. The studies become comparable only when the claim is narrowed to the outcome each actually observes.

An economic effect also needs a time horizon. Implementation costs can occur before benefits, and a strategic decision can change the distribution of risk as well as average profit. The reviewed evidence does not provide a general discount rate or a universal period over which AI-supported strategies should pay back. Those quantities cannot be supplied by a writing-productivity experiment or an expert rating of a proposal.

7. Exploration, evaluation and competitive interaction

March's exploration–exploitation framework distinguishes search for new possibilities from the use of established knowledge [4]. It supplies a conceptual vocabulary for separating expanded option generation from refinement of an existing route. The framework predates generative AI. It does not establish that a particular model architecture or prompting technique achieves an optimal balance for a business.

Gans's working paper formalises a strategist who may be human or AI within an adapted decision framework [7]. The paper is theoretical, so its conclusions depend on model assumptions and are not treated here as observed firm-performance effects. Formal theory can clarify the conditions under which a decision mechanism matters while leaving the empirical prevalence of those conditions open.

Competitive interaction introduces another level of analysis. If several firms obtain similar analytical improvements, each firm's absolute decision quality can improve without every firm gaining a relative advantage. Conversely, a better decision can create value in a category even when competitors also improve. Evaluations should identify whether the outcome is operational efficiency, absolute profitability or performance relative to rivals.

This is also why a strategy assessment cannot be reduced to the plausibility of a story about the future. A document may describe a coherent sequence that depends on unobserved customer responses or on rivals remaining inactive. Those assumptions can be made explicit and tested where possible. Fluency of explanation is a property of the representation, not evidence that its contingent events will occur.

8. Evaluating AI-supported strategic judgement

8.1 Compare decisions under common information

A useful comparison preserves the evidence available at the time of choice. Supplying later outcomes to one condition creates hindsight rather than a prospective test. Evidence packets, tool access and time constraints should be recorded. If one group can commission additional research, the intervention includes that expanded information-gathering process and should be described accordingly.

The reference decision also needs to be specified. Expert agreement can evaluate professional acceptability, but experts may share assumptions or disagree legitimately. A realised outcome can assess a prediction, yet outcomes contain chance and contextual change. Neither measure alone is a complete standard for strategic quality. A strong design can examine reasoning, information use and later performance while preserving the distinction among them.

8.2 Keep rejection visible

Strategy involves rejecting options as well as selecting them. If evaluation considers only surviving proposals, it can miss whether a system screened out useful alternatives. A test of option generation needs coverage of plausible alternatives; a test of selection needs the candidate set on which selection operated. Better scores among survivors may reflect a narrower set rather than better judgement.

Camuffo and colleagues' trial trained entrepreneurs in a scientific approach to testing business ideas and compared their subsequent decisions with a control condition [3]. It is not an AI intervention. Its relevance is the direct study of decision method and revision rather than treating a polished proposal as the final outcome. Testing assumptions can improve the information available for a choice even when it leads to abandonment or change.

8.3 Distinguish confidence from calibration

A confident recommendation can be correct or incorrect. Calibration concerns whether stated confidence corresponds to accuracy across a set of cases. It requires repeated observations with outcomes that can be assessed; the tone of one answer cannot establish it. Strategic uncertainty also includes factors for which no stable reference frequency is available. Reporting those limits is part of describing the decision, not a substitute for making it.

9. Application to marketing strategy

Marketing strategy combines market evidence, organisational capability and choices about customers and commitments. The reviewed studies support testing AI on specified components such as synthesising supplied evidence or comparing alternatives against explicit criteria. They do not establish that a model can independently infer every relevant capability, customer constraint or competitive reaction from a broad request for a strategy.

A factual claim about a market, a causal claim about a customer response and a proposed strategic commitment should remain distinguishable in the output. Each has a different route to verification. Market facts can be traced to sources; causal claims need appropriate comparisons; commitments need feasibility and authority. A single citation attached to an entire recommendation may support only one premise in that chain.

When comparing AI-supported work with a previous process, the unit of comparison should include review and correction. A first draft that appears faster can transfer work to downstream reviewers. A slower analytical phase can reduce later rework. Measuring the complete decision process avoids attributing an isolated production gain to the entire strategy function. The relevant boundary depends on the business question being tested.

The human–AI meta-analysis separates augmentation from synergy, a useful distinction when a strategic plan is rated more highly with assistance. [13] The automation taxonomy distinguishes information acquisition, analysis, decision selection and implementation. [14] Lee and See relate trust to appropriate reliance rather than to unqualified acceptance of automated advice. [15]

Cognitive-forcing research examines designs that reduce overreliance in a defined decision task. [16] Bansal and colleagues test whether explanations improve complementary team performance, rather than assuming that more explanation necessarily helps. [17] The Productivity J-Curve treats complementary investment and the timing of measured returns as part of technology adoption. Better-rated strategic advice and later firm performance consequently remain distinct outcomes. [18]

10. Limitations and conclusion

The selected literature spans formal theory, laboratory judgement, professional tasks and a field intervention. Their outcomes are heterogeneous and several sources were available only in summary form. No evidence establishes a universal treatment effect of AI on marketing-strategy success. Models and interfaces also change, so a result belongs to its tested intervention rather than an abstract category of intelligence.

Existing research demonstrates useful strategic capabilities and uneven consequences. AI can generate and evaluate proposals, improve some professional tasks and influence real business outcomes. Whether those contributions improve judgement depends on the task, evidence, engagement and implementation. The defensible evaluation keeps prediction, explanation, option quality and realised performance separate, then tests the connections relevant to the decision. A stronger document is one possible result; a better strategic commitment remains an empirical question.

Source access and appraisal register

The additional comparative sources in references 13–18 are used for the specific distinctions cited in the discussion. Access is limited to previously inspected primary-source abstracts, publication records or manuscripts. Their inclusion does not imply a fresh full-text risk-of-bias appraisal or direct validation in the application discussed here.

Sources were consulted on 21 October 2026. Access categories describe the material inspected for this narrative review. An abstract supports only its reported summary; it does not establish a check of the complete analysis or supplementary material. No independent replication or formal risk-of-bias scoring was undertaken.

References