Strategic leadership at high altitude: Investigating how AI affects the required skills of top managers

Bevilacqua, Ferraris, Matzler & Kuděj — how AI changes the skills top managers need

TL;DR

A qualitative study in the Journal of Business Research (vol. 205, available online 2 December 2025, open access). The authors interviewed 23 senior executives at European firms with at least 250 employees in five sectors: automotive, consumer goods, engineering, logistics and telecommunications. Interviewees were CEOs, CTOs, CDOs, CMOs, CSOs, a chief risk officer and a CHRO, among others. The research question is how does AI affect the required strategic leadership skills of top managers? Grounded in upper echelons theory (UET), which holds that a firm’s outcomes reflect its top executives, and coded with the Gioia method, the interviews yield four interdependent skills:

1. AI open mindset (personal level): openness to experimentation and continuous learning, fed from three directions: top managers’ own willingness, ideas coming up from middle managers and the front line, and external pressure.
2. AI strategic co-thinker (organizational level): treating AI as a collaborator in strategy, and filtering its outputs by framing the problem, asking the right questions and checking the answer against context.
3. Multi-level connector (relational level): engaging middle managers inside the firm and building AI relationships with universities, vendors, suppliers and customers outside it.
4. Ethics risk management (organizational level): a responsible-AI culture plus formal AI ethics committees.

The framework’s claim is that these work as one system. The open mindset is the gateway; co-thinking and ethics are a “dual capability” that keep each other in check; the connector spreads all three through the organisation. The paper’s distinctive move for UET is middle managers: they emerge as active co-authors of AI strategy, not implementers, so the authors argue upper echelons research must look below the upper echelon.

Key claims

AI is not one technology, and the paper tries to show it. Prior studies treat AI as a generic piece of digital transformation. Here, generative AI, machine learning, robotics and augmented reality are tracked separately (Table 6). Generative AI raises misinformation and content-integrity issues and supports co-creating strategy and communication. Machine learning drives forecasting and needs technical experts and executives to integrate their insights. Robotics raises safety, liability and workforce well-being. AR supports immersive simulation of decisions and raises privacy and psychological concerns. The distinction is drawn more in the tables than in the interview evidence, which mostly speaks of “AI” in general.

The open mindset has three sources.

• Top-down: “We started from the top! If top managers don’t get it, how can the rest follow?” (automotive). Executives take personal ownership of learning AI tools rather than delegating.
• Bottom-up: “The most impactful AI innovation began with a proposal from middle managers who just tried a new tool and realized its value” (engineering), and “The initiative to use AI for customer support came from the field, where we were struggling with overload and delay” (consumer goods).
• External: “After COVID-19, we understood that the only way forward was to radically change: AI innovation had to become central” (logistics); “if our competitors are already using AI, we can’t afford to stay behind” (consumer goods); and customers who now expect AI-powered experiences.

Co-thinking means filtering. “The quality of the AI response depends on how well you define the problem” (engineering). “We always ask: does this make sense in our context? Could it be misleading? What’s missing?” (automotive). “AI can show you ten options, but only experience helps you pick the one that works in this environment” (logistics). The authors stress that delegating analysis to AI “redefines, rather than replaces, the executive’s ultimate responsibility”: accountability stays with the human decision-maker. One telecoms executive goes further: “Entrusting routine decisions to artificial intelligence allows us to automate processes, reduce costs, and minimize human cognitive biases.”

Middle managers get new roles and more autonomy. Firms run targeted AI training for middle managers with personalised learning paths (“Our middle managers choose their courses based on the technical skills they want to acquire”), sandbox environments and curated datasets. New mid-level roles appear: Data & AI Product Manager, AI Trainer and Data Modeler, Ethical AI Specialist. AI also “simplifies processes that required multiple managerial layers” (telecoms), which the paper does not connect to its claim that middle managers become more central.

Ethics committees as dynamic governance. Committees combine ethicists, lawyers, AI engineers and business strategists, review AI systems before deployment and revise guidelines periodically rather than once (“It’s a continuous process of ethical adaptation”). Not all have veto power, and firms whose committees do have it report more internal trust, regulatory preparedness and stakeholder confidence. That claim rests on interviewees’ accounts, with no comparison across firms. The authors also observe that ethical concerns change with AI maturity: early adopters focus on compliance and avoiding harm, mature ones on trust and a “social license to operate”. One firm’s description of its governance function: “Hand on the wheel and foot on the accelerator.”

A development sequence. The managerial recommendations put the four skills in order: build the open mindset first, then co-thinking, then the connector capabilities, and institutionalise ethics risk management last, as a “mature governance layer”. The suggested formats are AI immersion bootcamps, executive shadowing of data-science teams, reverse mentoring by AI-literate juniors, executive roundtables and cross-functional AI leadership labs. The sequence follows from the framework’s logic; it was not tested.

What it adds to leadership theory. Table 7 maps each new skill to the prior skills it extends. The authors say they go beyond Mumford et al.’s (2007) Strataplex of cognitive, interpersonal, business and strategic skills in three ways. Cognitive and strategic skills must now work with opaque, evolving systems. Interpersonal skills extend beyond the team to whole ecosystems. Ethics moves from a peripheral value to the core of strategic decision-making.

Dynamic-capabilities reading

The paper does not use the W&W vocabulary, but its second-order themes map onto it closely.

• digital-mindset-crafting: the AI open mindset is this cell at executive level. It covers enabling an entrepreneurial mindset and promoting a digital one, with executives “legitimizing failures, sponsoring AI pilot projects, and signaling curiosity as a strategic norm”.
• improving-digital-maturity: AI training programmes for middle managers with personalised paths, experimentation tools, and leveraging AI knowledge inside the firm.
• redesigning-internal-structures: new hybrid roles (AI product manager, AI trainer, ethical AI specialist), cross-functional AI labs, and AI ethics committees as formal structures.
• navigating-innovation-ecosystems: the “AI stakeholders’ relationship” theme: joint labs with universities, co-development with AI software providers, AI-enabled supplier integration and co-creation with customers.
• strategic-renewal/organizational-culture: a test-and-learn culture that treats mistakes as discovery, and a responsible-AI culture in which ethics is “a fundamental condition of innovation”.
• external-triggers: the “proactivity towards business environment stimuli” theme: COVID-19, competitors already using AI, and customers expecting AI-powered service.
• internal-enablers: executive support, in the form of top managers’ AI willingness and visible personal engagement with AI tools.

The paper is also a micro-level account of what Teece calls managerial cognition: the skills of the people who sense and seize. UET and the dynamic-capabilities literature meet here, though the paper cites neither Teece nor Warner & Wäger.

Neighbour sources

The strongest quantitative neighbour is Nguyen et al. (2026). In their survey of 325 managers, top leadership support is the strongest driver of AI adoption level. Its items are managers’ willingness to risk adoption and commitment to acquiring skills, which is what this paper calls top managers’ AI willingness. One study measures the path; this one describes what is inside it.

On the open mindset:

• Linda Hill (Bloomberg Leaders, 2026) gives the same stance in her own terms: wayfinding, where “everything is a working hypothesis.”
• Argenti (HBR, 2026) makes mindset prior to skillset for every role, not just executives.

On co-thinking, Goel (BCG, 2026) names the same two human responsibilities: framing the problem and stress-testing the output. He adds a warning this paper lacks: in BCG’s study of 70 C-suite leaders, those are the skills most at risk of eroding through everyday AI use. The paper treats filtering as a skill executives develop; Goel treats it as one they may be losing.

On middle managers and bottom-up innovation:

• Sadun (Microsoft WorkLab, 2025) calls middle managers the hinge of reskilling and uses incentives to make coaching part of their job.
• DBS (MIT SMR, 2026) is the most developed case in the corpus of innovation organised from below.

A tension worth holding: Lakhani (Oct 2026) puts middle managers among the people most at risk from AI. That claim is about jobs, this paper’s is about roles in adoption, so the two can both hold. The paper’s own telecoms quote, AI removing “multiple managerial layers”, points the same way as Lakhani.

On measurement, the authors’ second future-research line asks for instruments to measure these skills. Weidmann, Xu & Deming (2025) have a validated method for measuring general leadership skill with AI teams. It does not measure these four skills, but it shows such instruments can be built cheaply.

Linked entities and concepts

• Concepts: dynamic-capabilities, enterprise-ai-adoption, responsible-ai, ai-deskilling
• Synthesis: organizational-frameworks-for-ai-adoption, where the framework fills a leadership-skills layer.
• Dangling (single-source mention, deferred per the second-source promotion rule): Simone Bevilacqua (University of Turin and University of Innsbruck; corresponding author), Alberto Ferraris (University of Turin), Kurt Matzler (University of Innsbruck), Michal Kuděj (Pan-European University and Prague University of Economics and Business). Bevilacqua is also first author of a 2025 systematic review of AI and top managers’ leadership (Bevilacqua et al., 2025a), cited throughout and a candidate for acquisition.

Scope and reliability

What was read. The complete article, all 18 pages, from the user-supplied ScienceDirect PDF (originally 1-s2.0-S0148296325007015-main.pdf, renamed to its content slug at acquire time). The tables were read from flattened text; Figures 1 and 2 are images and were read from the rendered PDF.

Sample and method. 23 interviews (6 face-to-face, 17 on Zoom, 30–60 minutes, average 46) with one executive per firm. The authors report saturation after 18 interviews. The interviews were supplemented with company documents and websites. The interview guide has five open questions, two of which ask directly which hard and soft skills executives need, so the skill list partly reflects what executives say they need rather than what they were observed doing. All firms are large, European and in five sectors; finance and healthcare are absent. The firms are anonymised.

What the paper does not show. It does not link the skills to any outcome: firm performance, AI project success or adoption speed. The authors list this as the first gap for future work. Several claims are generalisations from interviewees, notably the trust benefits of ethics committees with veto power. They have no comparison group. The authors acknowledge that the study misses the political side of AI adoption (resource fights, power shifts, job displacement), which bears directly on the multi-level connector. They also call it a temporal snapshot: they expect the co-thinker skill to change as AI becomes more autonomous, and the connector to grow into an “ecosystem architect” role.