Back to Insights

INSIGHTS / COGNITIVE AI GOVERNANCE

The Executive Judgement Gap

Why AI adoption is advancing faster than leadership capability — and why that matters for the future of executive decision-making

By Merlin Lockey GSM, Founder and Principal Educator, ZenithWellPublished November 202610 min read

Executive summary

Organisations are adopting artificial intelligence quickly. Eurostat reports that 20% of EU enterprises used AI in 2025, up from 13.5% a year earlier, and 88% of organisations in McKinsey’s 2025 global survey use it in at least one function. But being able to deploy AI is not the same as being able to judge what it produces. This article introduces the Executive Judgement Gap, a ZenithWell conceptual framework. It describes the distance between an organisation’s AI adoption capability and its leaders’ capability to interrogate, verify, challenge and govern AI-generated intelligence. Research on automation bias, on the link between confidence in AI and reduced critical thinking, and a field experiment with 758 BCG consultants all suggest the same thing: fluent, confident output can be persuasive even where it is wrong. Using an illustrative banking scenario, the article argues that human oversight without reasoning competence is a formality. It also sets out five questions every board should ask. It then presents ZenithWell’s four disciplines, DECONSTRUCT, VERIFY, STRESS-TEST and GOVERN, as a connected method. It concludes that executive reasoning capability must be assessed and developed rather than assumed.

Every week, somewhere in Europe, an executive committee approves a recommendation that was drafted, summarised or substantially shaped by artificial intelligence. The paper is well structured. The argument flows. The references look credible. The conclusion is confident. And around the table, a group of capable, experienced people nod it through.

The question this article asks is not whether that recommendation was right. It is whether anyone in the room was equipped to know.

Part 1 — The AI adoption acceleration

The pace of organisational AI adoption is no longer in serious dispute. According to Eurostat [1], 20.0% of EU enterprises with ten or more employees used AI technologies in 2025, up from 13.5% a year earlier, and Eurostat’s detailed statistics show that the share among large enterprises had reached 55% [2]. Among larger organisations surveyed globally, the picture is more advanced still: McKinsey’s 2025 State of AI survey of 1,993 respondents found that 88% reported their organisations regularly using AI in at least one business function [3].

Adoption, however, is not the same as mastery. The same McKinsey survey found that only around one-third of organisations had begun to scale AI across the enterprise [3]. And the motivation behind much of this investment deserves attention. In IBM’s 2025 CEO Study of 2,000 chief executives, 64% said the risk of falling behind drives investment in some technologies before they have a clear understanding of the value those technologies bring, and 37% agreed it is better to be “fast and wrong” than “right and slow” [4].

Those figures describe organisations acting under competitive pressure, often rationally. But they also describe a pattern in which the capability to deploy is running ahead of the capability to understand. As AI-generated analysis moves from drafting emails into market assessments, risk summaries, legal research, investment memoranda and board papers, it begins to shape the information on which consequential decisions are taken.

That raises a question most AI readiness assessments do not ask: is the reasoning capability of the people accountable for those decisions keeping pace?

Part 2 — Introducing the Executive Judgement Gap

ZenithWell proposes a simple analytical lens for that question. We call it the Executive Judgement Gap. It is a conceptual framework, not an established academic theory or a validated measurement model, and it is offered as a way of examining organisational readiness rather than as a finding.

The framework distinguishes between two capabilities that are often treated as one.

The first is AI adoption capability: an organisation’s ability to access and deploy AI technologies, integrate them into operational processes, generate and use AI-assisted information, improve productivity, and introduce AI-supported analysis into decision-making.

The second is executive reasoning capability: the ability of an organisation’s leaders to interrogate AI-generated intelligence; to distinguish factual claims from assumptions and inference; to identify unsupported assertions and evidential weaknesses; to examine the reliability and relevance of sources; to challenge conclusions and consider alternative explanations; to recognise uncertainty, bias and limitation; and to establish human oversight, accountability and defensible decision protocols.

The Executive Judgement Gap is the distance between the two. Where adoption advances faster than reasoning, the organisation becomes increasingly dependent on intelligence its decision-makers are not adequately equipped to scrutinise.

There are good reasons to expect the two capabilities to develop at different rates. Adoption capability can be bought: licences, platforms, integrators and consultants are readily available, and their progress is easy to measure in use cases, users and hours saved. Reasoning capability cannot be bought; it must be developed, and it is rarely measured at all. Adoption also generates its own demand for reasoning. Every additional AI-assisted report adds to the volume of material that someone, somewhere, must evaluate before acting on it. The more successful the deployment, the heavier the evaluative load on leadership.

There is some evidence that boards recognise the issue, at least in relation to AI knowledge. In Deloitte’s 2025 survey of around 700 board members and executives across 56 countries, 66% said their boards had “limited to no knowledge or experience” with AI — an improvement on 79% in the previous edition [5]. That measures familiarity with AI, not reasoning about its outputs, and it should not be stretched further. But it suggests that oversight capability cannot simply be assumed to exist.

THE EXECUTIVE JUDGEMENT GAP

When the ability to deploy AI outpaces the ability to interrogate, evaluate and govern its outputs

The Executive Judgement GapAI adoption rises rapidly while collective executive reasoning develops more gradually. The widening distance represents unexamined decision exposure. Conceptual illustration, not empirical data.Time →THE EXECUTIVEJUDGEMENT GAPUnexamined Decision Exposure

The Executive Judgement Gap — Unexamined Decision Exposure

AI Adoption Capability

Deployable tools, automated drafts, interaction volume

Collective Executive Reasoning Capability

Interrogation, verification, evidentiary governance

Figure 1. The Executive Judgement Gap — a ZenithWell conceptual illustration, not empirical data.
AI adoption is not, in itself, evidence of organisational readiness. The ability to use AI does not confer the ability to judge it.

Part 3 — When plausibility is mistaken for reliability

Generative AI produces fluent language. Fluency is a property of expression, not of evidence. The central risk for executives is not that AI-generated analysis is obviously wrong; it is that it is persuasively incomplete.

Human factors research has studied a version of this problem for decades. A systematic review by Goddard, Roudsari and Wyatt, published in the Journal of the American Medical Informatics Association in 2012 and drawing on 74 studies from healthcare, aviation, military and other settings, describes automation bias as the tendency to over-rely on automation, accepting its output “as a heuristic replacement of vigilant information seeking and processing” [6]. It produces two kinds of error: commission errors, in which people follow incorrect advice, and omission errors, in which they fail to act because the system did not prompt them. That research predates generative AI, but the mechanism it describes is directly relevant to any executive reading a confident machine-generated recommendation.

More recent studies point in a similar direction, with important caveats. In a 2025 study presented at the ACM CHI conference, Lee and colleagues at Microsoft Research surveyed 319 knowledge workers about 936 real examples of generative AI use. Higher confidence in the AI’s ability to perform a task was associated with less self-reported critical thinking, while higher confidence in one’s own ability was associated with more. The authors are clear that these are self-reported associations, not causal findings [7]. A 2025 study by Gerlich in the journal Societies, surveying 666 participants in the UK, reported a negative correlation between frequent AI tool use and critical thinking scores, partly mediated by cognitive offloading. It too is cross-sectional and self-reported, and its reported coefficients vary between tables; it is best read as a signal worth investigating, not as proof [8].

The most instructive evidence for executives comes from a field experiment. Dell’Acqua and colleagues, in research published in Organization Science in 2026, studied 758 Boston Consulting Group consultants. On tasks within AI’s capability, consultants using GPT-4 completed 12.2% more tasks, 25.1% more quickly, at significantly higher quality. On a task designed to fall outside that capability, consultants using AI were 19 percentage points less likely to reach a correct answer than those working without it [9].

The authors call this a “jagged technological frontier”. For leaders, its most important feature is that it is invisible. The output looks equally polished on either side of the line. Nothing in the prose tells you whether you are reading the work of a system operating within its competence or beyond it. Only the reader’s reasoning can do that.

Plausible is not proof. A well-written answer and a well-evidenced answer can look identical on the page.

Part 4 — The boardroom accountability problem

Consider the following illustrative scenario. It is a constructed example, not a description of any real organisation or documented incident.

A mid-sized European bank’s executive committee receives a forty-page analysis recommending a significant shift in its lending strategy: reducing exposure to commercial real estate and expanding into a fast-growing segment of small-business lending. The analysis has been assembled by a strategy team using an AI research assistant. It is professionally formatted, identifies clear market trends, cites what appear to be reputable sources, and ends with a confident recommendation and projected returns.

The committee reads it. The argument is coherent and consistent with what members have heard elsewhere. Questions focus on implementation timing. The recommendation is approved.

What nobody examined was whether the evidence supported the conclusion. Two of the cited market reports were summaries of summaries; one statistic was drawn from a different jurisdiction; the growth trend rested on a period of unusually low default rates; the alternative explanation — that the segment’s growth reflected loosened underwriting across the market — was never considered; and no one had been asked to argue the other side.

None of the committee members was negligent in the conventional sense. They read the paper, discussed it, and decided. But they approved a line of reasoning they had not tested.

European regulation has already named this risk. For high-risk AI systems, Article 14 of the EU AI Act requires that the people overseeing them be enabled to “remain aware of the possible tendency of automatically relying or over-relying on the output” — automation bias — particularly where systems provide information or recommendations for decisions taken by people [10]. Article 26 requires deployers to assign oversight to people “who have the necessary competence, training and authority” [11]. A strategy analysis of this kind would not necessarily be a high-risk system under the Act, and the Annex III high-risk obligations now apply from 2 December 2027 [12]. But the principle stands on its own: oversight without competence is a formality. Accountability for the decision stays with the people who approved it, whatever tool helped draft it. The same expectation runs through the OECD AI Principles, the first intergovernmental standard on AI, which since their 2024 update have listed accountability among their five values-based principles [13].

The risk is not hypothetical. In McKinsey’s 2025 survey, nearly one-third of respondents reported negative consequences arising from AI inaccuracy [3].

Which leaves the question the scenario was designed to pose: can an organisation genuinely claim to exercise meaningful human oversight if its decision-makers lack the capability to critically interrogate the information they approve?

Part 5 — Five questions every executive board should be asking about AI-generated intelligence

These questions are not a compliance checklist. They are a test of whether oversight is substantive.

  1. Can our senior leaders tell an AI-generated assertion from an independently supported conclusion?

    If a claim appears only because the system produced it, it is an assertion. A conclusion requires evidence that would survive without the system.

  2. Can they recognise when credible-looking evidence is not enough to justify a recommendation?

    Citations, charts and confident prose are signals of effort, not of sufficiency. The question is whether the evidence carries the weight the recommendation places on it.

  3. Are they equipped to challenge an AI-generated recommendation rather than simply approve it?

    Challenge requires more than permission. It requires the method and the confidence to ask what would have to be true for this to be wrong.

  4. Do we have clear human accountability for decisions informed by AI?

    Every consequential AI-informed decision should have a named owner who can explain why it was taken.

  5. Could we reconstruct and defend the reasoning behind our consequential AI-informed decisions?

    If a regulator, auditor, court or shareholder asked how a decision was reached, could the organisation show not only what the AI produced, but how leaders tested it?

A board that cannot answer yes to all five with confidence has located its Executive Judgement Gap.

Part 6 — From AI adoption to executive reasoning capability

If the gap is a reasoning problem, the response must be a reasoning discipline. ZenithWell’s approach is built on four connected disciplines. They are designed to work as a single sequence — from taking an AI-generated argument apart to deciding how far it may influence a decision — rather than as four separate topics.

FROM PLAUSIBLE OUTPUT TO DEFENSIBLE JUDGEMENT

The four ZenithWell disciplines

  1. 01

    DECONSTRUCT

    Separate facts, inference & rhetoric

  2. 02

    VERIFY

    Test source reliability & sufficiency

  3. 03

    STRESS-TEST

    Challenge assumptions & edge cases

  4. 04

    GOVERN

    Establish human authorship & auditability

AI-generated intelligenceDefensible executive decision

ZenithWell methodology. Presented as a reasoning framework; not independently validated. zenithwell.net

Figure 2. The four ZenithWell disciplines — one connected reasoning and governance sequence.

DECONSTRUCT. Separate the factual claims, assumptions, inferences and persuasive language within AI-generated information. Most weak analysis is not false throughout; it blends sound facts with untested assumptions in prose that makes them indistinguishable. Deconstruction makes the structure of the argument visible.

VERIFY. Examine the reliability, relevance and sufficiency of the supporting evidence and sources. Is the source real, current and applicable to this jurisdiction and context? Is the evidence sufficient for the strength of the conclusion?

STRESS-TEST. Challenge the conclusion, identify potential bias, examine assumptions and consider alternative explanations. This is where the jagged frontier is most often discovered: by asking what the analysis has not considered.

GOVERN. Establish appropriate human oversight, accountability and defensible decision protocols: who owns the decision, what standard of evidence it requires, how challenge is recorded, and how the reasoning can be reconstructed later.

Each discipline answers part of the gap. Deconstruction and verification address the difference between fluency and evidence. Stress-testing addresses automation bias and the invisible frontier. Governance addresses accountability and traceability. Together they turn oversight from a signature into a practice.

A final point matters as much as the method. Executive reasoning capability should be assessed and developed, not assumed. Most organisations can report how many employees use AI tools; very few can say how well their decision-makers evaluate what those tools produce. Even the EU’s own AI literacy provision, Article 4 of the AI Act as amended by the 2026 Digital Omnibus [14], [15], now requires organisations to “take measures to support” AI literacy while expressly not requiring them to guarantee any specific level for any individual [14]. Literacy for operating AI and the reasoning required to govern its conclusions are different things. Defining and measuring the latter is a leadership responsibility, not a regulatory default.

That is why ZenithWell’s work begins with assessment — establishing how well leaders currently interrogate AI-generated intelligence — and returns to it, so that the development of that capability can be evidenced over time rather than asserted.

Part 7 — The leadership imperative

When the World Economic Forum’s Future of Jobs Report 2025 surveyed employers, analytical thinking remained the most sought-after core skill, with seven in ten companies considering it essential [16]. That finding is about the workforce in general. Its implication for leadership is sharper: in an environment saturated with machine-generated analysis, the scarce asset is not information, but the judgement to know what it is worth.

Most definitions of organisational AI readiness still measure infrastructure, data, use cases and adoption rates. These matter. But they describe only one side of the Executive Judgement Gap. An organisation can be technically ready and judgementally exposed at the same time — and the more successfully it adopts AI, the more exposed it may become.

The leaders who will govern AI well are unlikely to be those with the most tools or the fastest deployment. They will be those who insist that every consequential conclusion can be taken apart, checked, challenged and defended — and who build that expectation into how their organisations decide.

If your organisation is investing in artificial intelligence, perhaps the next question is whether it is investing equally in the human reasoning capability required to interrogate, evaluate and govern it.

The future of responsible AI adoption will depend not only on the intelligence organisations can access, but on the quality of judgement exercised by the people responsible for using it.

I would welcome the views of board members, executives and partners: how does your organisation currently test the AI-generated analysis that reaches the decision table?

About the author

Merlin Lockey GSM is the Founder and Principal Educator of ZenithWell. His professional work spans more than 25 years across higher education, corporate communication, quality management and executive consulting in international and cross-cultural settings. His current work focuses on executive reasoning capability, decision integrity and the governance of AI-generated intelligence.

About ZenithWell

ZenithWell is a corporate and institutional partner for developing executive reasoning capability in the age of AI. It works with leaders and boards to assess and strengthen their ability to interrogate AI-generated intelligence, identify evidential weaknesses, challenge conclusions and establish defensible decision-making practices. To explore ZenithWell’s work in executive reasoning capability and cognitive AI governance, visit zenithwell.net.

A private executive conversation

Test your organisation's reasoning exposure

Try the Free AI Reasoning Audit

References

Numbers in square brackets in the text refer to the entries below, listed in order of first citation. All links accessed 12–14 November 2026.

  1. [1] Eurostat. 20% of EU enterprises use AI technologies. News release, 11 December 2025. https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20251211-2
  2. [2] Eurostat. Use of artificial intelligence in enterprises. Statistics Explained, 2025 data. https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises
  3. [3] McKinsey & Company (QuantumBlack). The state of AI in 2025: Agents, innovation, and transformation. Global survey, 25 June–29 July 2025, n=1,993. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-2025
  4. [4] IBM Institute for Business Value with Oxford Economics. IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles (2025 CEO Study). 6 May 2025, n=2,000 CEOs. https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles
  5. [5] Deloitte Global. Governance of AI: A critical imperative for today’s boards, 2nd edition. Survey January–February 2025, approx. 700 board members and executives, 56 countries. https://www.deloitte.com/za/en/issues/trust/progress-on-ai-in-the-boardroom-but-room-to-accelerate.html
  6. [6] Goddard, K., Roudsari, A. and Wyatt, J. C. (2012). Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association, 19(1), 121–127. https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/
  7. [7] Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM. https://dl.acm.org/doi/10.1145/3706598.3713778
  8. [8] Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. Published 3 January 2025; corrected 10 September 2025. https://doi.org/10.3390/soc15010006
  9. [9] Dell’Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, published online 11 March 2026. https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf
  10. [10] European Commission, AI Act Service Desk. Regulation (EU) 2024/1689 (AI Act), Article 14: Human oversight. Consolidated text as at 27 July 2026. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14
  11. [11] European Commission, AI Act Service Desk. AI Act, Article 26: Obligations of deployers of high-risk AI systems. Consolidated text as at 27 July 2026. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-26
  12. [12] European Commission. AI Act: Regulatory framework for AI (application dates following the AI Omnibus). Accessed 14 November 2026. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
  13. [13] OECD. OECD AI Principles. Adopted 2019, updated 2024. https://www.oecd.org/en/topics/ai-principles.html
  14. [14] European Commission, AI Act Service Desk. AI Act, Article 4: AI literacy (as amended by the Digital Omnibus on AI). Consolidated text as at 27 July 2026. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4
  15. [15] Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 (Digital Omnibus on AI). Official Journal of the European Union. https://www.eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=OJ:L_202601744
  16. [16] World Economic Forum. The Future of Jobs Report 2025. January 2025. https://www.weforum.org/publications/the-future-of-jobs-report-2025