When AI Takes On Tasks: What a CEO Needs to Know About Workforce Transformation

Published: 22 April 2026

The debate about AI and jobs is louder than the currently robust evidence, once you look closely. Even so, deferral is not an option. Workforce decisions are, first and foremost, decisions for the executive team. They take effect over the long term, which is why poor decisions often only become visible years later, once corrections have become costly.

The gap between the headline and the evidence

In autumn 2025, a preprint on the Remote Labor Index was published (Center for AI Safety and Scale AI). High-performing AI agents were tested on 240 real end-to-end projects from the freelance market. The setting mattered: deliberately realistic but strictly framed, typically without deep internal system integration and without the kind of process and data access that companies can build in integrated environments.

The result is sobering. The best-performing agent tested reached an end-to-end automation rate of 2.5 per cent. That does not mean AI delivers nothing. The setup likely contributed to the low automation rate, since no deeper system integration took place. It does, however, mean that autonomous, full replacement of real project work, as of 2025, is not broadly supported by empirical evidence, at least not in a setting of this kind. Within integrated corporate processes, AI as augmentation can achieve considerably more.

For the executive team, this means: headcount reductions, hiring, or retraining must not be based on hype. They must be grounded in performance data from your own context. Anyone cutting jobs today on the assumption that everything will be automated within twelve to eighteen months is making a bet that is currently not well supported by evidence.

The quiet transformation already under way

While full automation remains an open question, a second development is already visible today: AI is changing what employees can still do unaided once they use these tools daily.

Michael Gerlich (SBS Swiss Business School) reported in 2025, in a cross-sectional study of 666 participants, a negative correlation between frequent use of AI tools and the capacity for critical thinking, with cognitive offloading, the outsourcing of cognitive work to external systems, modelled as a statistical mediator. The design does not permit a causal claim. It is conceivable that people with certain cognitive profiles simply use AI more often than others. Even so, the finding remains relevant for executive teams and boards. The central question is not only whether AI replaces employees. It is also: which judgements, analyses, and professional assessments will still be possible independently in five years, and which can no longer be made responsibly without a tool?

That is not an IT question. It is a question of organisational resilience.

What I observe from my own leadership experience

Across transformations of recent decades, technological and organisational alike, I see three patterns that companies regularly underestimate with AI.

Pace is overestimated, side effects are underestimated. A demo convinces instantly. In operations, however, it is data access, process integration, role clarity, control points, and acceptance that decide the outcome, not the model. I too have scaled projects too quickly: technically successful, but only economically viable once the organisation had accepted them.

Resistance is not noise. It is a risk indicator. Anyone who involves employees too late, or who cannot explain and train the reasons for the transformation in an understandable way, ends up with shadow processes, use without standards and controls, or withdrawal. Both outcomes cost time, quality, and trust. During the rollout of the GDPR, for example, I saw resistance drop markedly as soon as employees could see the connection to their own data. From that point on, workable solutions emerge that can actually be implemented in operations.

Workload shifts and requires new accountability. AI can save drafting time, but it increases the need for review, judgement, and context. Many plan for efficiency without pricing in review load, liability profile, and training needs; that is precisely where disappointed expectations later arise.

My practical conclusion: AI adoption is less a one-off fork in the road than an ongoing balancing act between speed and the organisation’s capacity to absorb change.

Why the hype persists

Market pressure does not come only from the media and consultants’ slide decks. It comes from capital allocation. Global investment in AI models, data centres, and chips has reached a scale that generates expectations of its own, among vendors, platforms, and in capital markets. As model providers such as OpenAI and Anthropic scale, and NVIDIA and comparable manufacturers carry the infrastructure wave at the same time, a strong incentive system emerges out of growth expectations, market share, and narrative. With every hype cycle, the essential question is who ultimately benefits, and in what form. That puts a great deal into perspective.

How far this narrative can run ahead of operational reality inside companies is shown by the MIT NANDA report 2025 (preliminary findings by Challapally et al., based on 300 AI initiatives, 52 structured interviews, and 153 survey responses). Against 30 to 40 billion US dollars in enterprise investment in generative AI, 95 per cent of organisations achieve no measurable return. Precision matters here too: such statements depend heavily on how return is defined and over what time horizon it is measured. If the assessment at enterprise level takes place after around six months, that is methodically a very early measurement point, particularly since rollouts from pilot to operational scale take considerably longer in many organisations. In mid-sized environments, I have observed implementation times of three to four years for complex end-to-end solutions. The absolute 95 per cent figure should therefore be read with caution; the underlying point nonetheless remains relevant: a measurable return from end-to-end automation is harder to achieve than many articles suggest.

As CEO, one question is therefore often more useful than the next model demo: who has which incentive in the AI-first hype, and who bears responsibility if the organisation cannot absorb it?

What leadership means at this stage

Technology effects are often visible early. Organisational and personnel side effects become visible late, and often differently than the business case suggested. The reverse also holds: inaction has costs too, only delayed. The point is not faster or slower. The point is coupling pace to readiness.

Four decisions that cannot be delegated

Separating replacement from augmentation. Many roles are not being eliminated. They are being rebuilt. Anyone who fails to draw this distinction properly plans past actual need and loses employees who will be strategically needed later.

Protecting core competencies. If cognitive offloading can be a real effect, the executive team must define which capabilities must remain in-house and independent. At a regulated asset manager, this might be investment decisions. At an industrial SME, it might be the company’s own technical problem-solving capability. That is strategy, not IT.

Actively leading workforce planning. Reddy Yanamala (2024) describes a useful three-level framework: analysing skills gaps, allocating resources dynamically, and planning succession proactively. The value lies less in the AI model than in the discipline of managing workforce as a steering object rather than administering it reactively.

Training as a leadership priority. AI changes processes, roles, and control points, and brings new risks with it, including data protection, cybersecurity, misuse, and false confidence. Training is therefore not a peripheral HR task. It is leadership work.

A special risk: AI in HR processes

Anyone deploying AI in recruitment, performance evaluation, or workforce planning encounters a stricter regime. Under the EU AI Act, such systems are generally classified as high-risk (Regulation 2024/1689, Article 6 in conjunction with Annex III, point 4). For Swiss companies, this is practically relevant depending on role (provider or deployer), platform exposure, and EU nexus. Executive search and interim providers with a DACH/EU footprint, for instance, face this challenge directly.

In parallel, Swiss law applies here too, including Art. 328b OR on data processing in the employment relationship and Art. 21 nFADP on automated individual decisions. The consequences are operational: bias controls, traceability, a final human decision, and documented governance must not be project decoration. They must be structurally embedded. Embedding this is part of the board’s oversight duty over the executive team under Art. 716a para. 1 no. 5 OR, and therefore falls directly within the board’s responsibility.

What remains

The empirical picture currently argues against a rapid, broad, autonomous displacement of complex knowledge work. At the same time, it is plausible that significant change will occur over time across nearly every field, including robotics and manual trades. The decisive factor is the time horizon: such transitions are in reality often more capital-intensive and operationally more complex than forecasts suggest. This will limit the pace of implementation in many organisations. What is all the clearer, though, is that requirement profiles, competencies, and governance are shifting, and that shift is already under way.

Under Art. 754 OR, the executive team is liable for the careful performance of its duties, and under Art. 716a para. 1 no. 5 OR, the board holds oversight over that performance. Swiss company law is technology-neutral, so an explicit reference to AI is not required. Accountability and duties of care do not shrink because of AI, quite the opposite, since new risks are added. Careful workforce leadership in the age of generative AI does not mean joining the hype. It means seeing evidence, cultural side effects, and the legal framework together, and deriving robust, verifiable decisions from that view.

Sources

Challapally, A., Pease, C., Raskar, R., & Chari, P. (2025). State of AI in Business 2025: The GenAI Divide. MIT NANDA, July 2025. Preliminary Findings.
Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. DOI: 10.3390/soc15010006.
Mazeika, M. et al. (2025). Remote Labor Index: Measuring AI Automation of Remote Work. Center for AI Safety, Scale AI. arXiv:2510.26787 (preprint, not peer-reviewed).
Reddy Yanamala, K. K. (2024). Strategic Implications of AI Integration in Workforce Planning and Talent Forecasting. Journal of Advanced Computing Systems, 4(1), 1–9.
Regulation (EU) 2024/1689 (EU AI Act), Art. 6 in conjunction with Annex III, point 4.
Swiss Code of Obligations (OR): Art. 328b, 716a para. 1 no. 5, 754.
Swiss Federal Act on Data Protection (nFADP): Art. 21.