65% of marketers and 64% of journalists were using AI at work. Their managers said they saw almost none. Both were right.
July 2026 · Case 6 of 6
As you read — hold this question
Employees are using AI at work but hiding it. Organisations say they see little AI use. Both statements are true simultaneously. If that's the problem — who has the adoption failure, the employees or the organisation?
65%
of marketers, 64% of journalists, 30% of lawyers — using AI at work in a Danish study. Yet their managers reported seeing almost no AI use and few productivity gains.
Professor Ethan Mollick named them "secret cyborgs": employees who are using AI to substantially improve their work but are deliberately not telling their employers about it. The individual productivity data is clear — consultants completed tasks 25% faster using GPT-4 (BCG/McKinsey study); GitHub Copilot produced 26% productivity improvement in coding (using the now-obsolete GPT-3.5); a Danish study found AI halved working time for 41% of the tasks employees do. Yet organisations report seeing little of this. The reason is not that the gains are imaginary. It is that individual AI productivity gains do not automatically translate into organisational AI productivity — and the gap between them is created by rational, self-protective behaviour by employees who have good reasons to hide what they're doing.
Why employees hide AI use — Mollick (2024)
Quiz: Trust & Safety
Six reasons rational employees keep their AI use invisible
Ethan Mollick's research identifies a systematic pattern: employees know they are using AI effectively, recognise the value, and still choose not to share. Each reason below is rational given the incentive structures of most organisations.
⚠
Fear of punishment from vague policies
They received a warning about "improper AI use" but were never told clearly what that meant. Rather than ask and risk being monitored, they hide their use entirely.
🏆
Fear of losing "hero" status
They are being praised for their rapid, high-quality output. If they reveal it's AI-assisted, they suspect colleagues and managers will respect them less.
📉
Fear of triggering layoffs
They know that when organisations discover technology creates productivity gains, the next move is often headcount reduction. Revealing AI capability is perceived as contributing to their own job insecurity.
🎁
No reward for disclosure
They have invested time in discovering effective AI workflows. Without a mechanism to be recognised or compensated for that knowledge, they see no reason to give it away.
📈
Gains will become new baseline expectations
If they reveal they can do 40% more work using AI, 40% more work will be expected. Same salary, higher load. The rational move is to pocket the personal time saving, not surface it.
📡
Want to share — but have no mechanism to do so
This is the most structural reason. Some employees genuinely want to share what they have discovered. But the organisation has no channel for this — no internal forum, no way to contribute learnings, no reward for knowledge sharing. So the knowledge stays with the individual.
The visibility problem
Why individual gains don't become organisational gains
The core problem Mollick identifies is a visibility failure. AI use that improves individual performance is invisible to the organisation. This creates two compounding effects.
Level
The effect of AI invisibility
The consequence
Individual
Each employee must discover AI use cases for themselves, from scratch — no learning from colleagues who have already done this
Mistakes are repeated across the organisation; effective workflows never scale; most employees remain at zero while a few carry all the AI capability
Organisational
Leaders lack information about how AI is actually being used, which tools are generating value, and which business processes are already being affected
Cannot make informed strategic decisions about which AI to adopt; cannot allocate resources to areas where impact is already occurring; strategy is built on absence of data
The historical pattern of technology adoption is clear: a very small number of early users — the "tinkerers" — figure out the use cases, and then everyone else copies them. The AI era is forcing every individual employee to be their own tinkerer. This is structurally inefficient and historically unprecedented — and it produces the pilot-hell pattern where organisations are stuck between fragmented individual experimentation and the pressure to achieve organisational-scale impact.
Booyse & Scheepers (2024) — seven barriers to AI in decision-making
What qualitative research found when asking senior managers directly
Booyse and Scheepers (2024) conducted qualitative interviews with 13 senior managers from South African organisations that had experienced AI adoption for automated decision-making. They applied Adaptive Structuration Theory (AST) — a framework for understanding the dynamic interaction between technological structures and social systems — to identify and classify barriers. Seven emerged.
Barrier 1
Human Social Dynamics
Resistance to change; interpersonal power relationships; group norms that discourage AI adoption even when individuals might be willing
Barrier 2
Restrictive Regulations
Legal and compliance constraints that limit what automated decision-making can do, even when technically capable
Barrier 3
Creative Work Environments
Knowledge workers who define their value through judgment and craft resist AI that automates decisions they consider core to their professional identity
Barrier 4
Lack of Trust and Transparency
Employees and managers who cannot see how the AI reached a decision cannot trust it enough to act on it — the black-box problem applied to organisational decisions
Barrier 5
Dynamic Business Environments
AI models trained on historical data struggle in rapidly changing environments — the model's training data is already out of date by the time it is deployed
Barrier 6
Loss of Power and Control
Managers who previously held decision-making authority resist AI that automates those decisions — the threat is not to their job but to their status and influence
Barrier 7
Ethical Considerations
Concerns about bias, fairness, accountability, and the appropriateness of delegating decisions — especially consequential ones — to automated systems
The solution — crowd + lab
How Mollick proposes to close the visibility gap
Mollick's prescription is not a single intervention but a dual-track R&D approach: the Crowd (bottom-up, decentralised) combined with the Lab (top-down, focused). Neither works without the other.
The Crowd — user innovation at scale
Drawing on Eric von Hippel's user innovation theory: the people best positioned to discover AI use cases are the people doing the actual work. The crowd approach means: (1) reducing fear — clear, permissive policies with a bias toward allowing AI use; (2) aligning rewards — recognising and compensating disclosure of effective AI workflows; (3) modelling positive use — executives visibly using AI so that employee adoption is normalised; (4) providing access — giving employees frontier models and platforms to experiment with; (5) building community — creating mechanisms for sharing what is being discovered.
The Lab — focused internal R&D
A dedicated team of subject-matter experts and technologists whose job is to build, not analyse. The lab takes the use cases the crowd discovers and turns them into deployable products. It also: builds organisation-specific AI benchmarks (because generic benchmarks don't measure performance on your specific tasks); builds prompts and tools that work for your context and measures them; builds "provocations" — demonstrations of what AI can do that create visceral understanding for those who haven't engaged with it yet; and continuously tests new models against your specific business-critical tasks.
Psychological Safety — the prerequisite
Mollick identifies psychological safety — the belief that disclosing AI use will not lead to punishment or penalty — as the key prerequisite for the crowd approach to work. Organisations with low psychological safety on AI will not see the crowd innovate, regardless of incentives, because the risk calculation for disclosure is negative. Establishing psychological safety is not a policy change; it is a culture change that requires visible, consistent leadership behaviour over time.
Take this away
The adoption problem is not that employees won't use AI. It is that the incentive structures of most organisations make it rational to hide effective AI use rather than share it. The gap between individual AI productivity and organisational AI productivity is not a technical gap — it is a governance, culture, and incentive design problem. Solving it requires the crowd approach (reduce fear, align rewards, model positive use) combined with the lab approach (dedicated internal R&D that builds from discovered use cases).
What is the "visibility problem" in AI adoption — and how does it operate at both the individual and organisational level?
The visibility problem is that individual AI productivity gains are invisible to the organisation. At the individual level: each employee must discover AI use cases from scratch, with no learning from colleagues. Mistakes are repeated; effective workflows don't spread; the few who have figured out AI use carry all the capability while the majority remain at zero. At the organisational level: leaders lack data about how AI is actually being used, which tools generate value, and which processes are already affected. This means AI strategy is built on the absence of information — leaders cannot make informed decisions about investment, platform selection, or capability development because they cannot see what is already working in their own organisation.
Question 2 of 3
Name four of the six reasons Mollick identifies for why employees hide their AI use — and explain why each is described as rational rather than irrational.
Any four of: (1) Fear of punishment from vague policies — rational because the policy says "improper use may be punished" without defining improper; the safe move is to hide use rather than risk being an example. (2) Fear of losing hero status — rational because credit for high-quality output may be reduced if it is attributed to AI rather than personal skill. (3) Fear of triggering layoffs — rational because the historical pattern is that productivity-enhancing technology is used to justify headcount reduction. (4) No reward for disclosure — rational because the employee has invested in discovering AI workflows and receives nothing for sharing that knowledge. (5) Gains become new baseline — rational because disclosed productivity gains become expected performance rather than rewarded performance; same pay, more output. (6) No mechanism to share — structural rather than behavioural; the employee may want to share but the organisation has created no channel for it.
Question 3 of 3
What is Adaptive Structuration Theory (AST), and how was it applied by Booyse and Scheepers (2024) to identify barriers in AI decision-making adoption?
Adaptive Structuration Theory (AST) is a framework for understanding the dynamic interaction between technological structures and social systems — how technology shapes human behaviour and how social norms and structures in turn shape the way technology is adopted and used. It goes beyond purely technical or purely social explanations by treating technology adoption as a two-way interaction. Booyse and Scheepers (2024) applied AST in qualitative interviews with 13 senior managers from South African organisations that had experienced AI adoption for automated decision-making. The AST lens revealed that the barriers are not simply technical (system limitations) or simply social (people reluctance) but emerge from the dynamic between the two. The seven barriers identified — including human social dynamics, lack of trust, loss of power and control, and ethical considerations — all reflect this dynamic interaction between the technology's capabilities and the social and organisational context it operates in.
Module 4 Video
Module 4 · Video Walkthrough
Sources
Mollick
Mollick, E. (2024). AI in organisations. One Useful Thing [blog]. Summarised in: The AI Daily Brief, "Making AI Work: Leadership, Lab, and Crowd." Module 4 reading, Flinders University BUSN9049, 2026.
Booyse & Scheepers
Booyse, D. & Scheepers, C.B. (2024). Barriers to adopting automated organisational decision-making through the use of artificial intelligence. Management Research Review, 47(1), 64–85. https://doi.org/10.1108/MRR-09-2021-0701
von Hippel
von Hippel, E. (2005). Democratizing Innovation. MIT Press.
Module transcript
Module 4 The Challenge of Enterprise AI Adoption. AI Daily Brief / Mollick summary. Flinders University, 2026.