The people in the middle of GenAI adoption

The numbers that open most AI adoption reports are not measuring the same thing. The Federal Reserve's research team demonstrated that simply reweighting by employment type and changing the time window can shift adoption figures by 5 to 40 percentage points — across studies examining the same underlying phenomenon.[5] When McKinsey reports 88% of companies using AI in at least one function, and a rigorous controlled study finds real adoption rates between 2% and 16% depending on how you define "use," the gap isn't noise. It's a measurement problem, and most organisations have built their adoption strategies on top of it without noticing.

The distinction that matters is between what Burton-Jones and Straub called "lean" and "rich" adoption[6] — the difference between logging into a tool and actually restructuring work around it. McKinsey's own 2025 State of AI report, based on nearly 2,000 respondents across 101 countries, found that 80% of organisations have superimposed AI onto existing processes without changing anything beneath them. They've named this "AI theater."[7] Meanwhile, MIT's NANDA group, after analysing over 300 public deployments, found that approximately 95% of generative AI pilots produced no measurable P&L impact.[8]

The question worth asking — before designing any adoption programme — is not "how do we get people to use AI?" but "what is actually blocking them, and is it what we think it is?"

It's Not the Technology

The most actionable finding from the 2025–2026 literature is also the most counterintuitive: the primary predictor of non-adoption among hesitant populations is not anxiety about AI specifically, but general comfort with technology. Shata's study of 294 educators — one of the few GenAI-specific studies to examine non-users directly — found that concerns about AI per se were not statistically significant in predicting intention to use (p=0.056). What was significant, with a beta of 0.576, was general technological comfort.[3]

This reframes the problem considerably. Programmes built around AI literacy — explaining what large language models do, addressing fears about job displacement, demonstrating safety features — are targeting the wrong variable. The friction is upstream: it's the interface, the unfamiliarity with any new digital tool, the absence of slack time to experiment. The response should be UI simplicity and protected exploration time, not better explanations of how transformers work.

Job displacement, notably, was the least important concern among both users and non-users in Shata's study. The actual psychological threat is elsewhere — in what happens to professional identity and the perceived value of expertise when a tool can produce a serviceable first draft in 90 seconds.

The Peer Effect Is the Mechanism

The most rigorous study of the period is Baym, Jaffe, and Dillon's work through Microsoft Research, published in HBR in March 2026. With 557 participants, it is the first to compare two adoption levers directly within the same programme, on a clearly identified GenAI object. The result: peer-to-peer influence produced +8.9 percentage points in regular adoption; formal training produced +6.1. For advanced AI agent use, peer influence produced +10.4 points. The difference isn't marginal — it's structural.[1]

The symmetry matters too. Gartner's December 2025 survey (N=2,986) found that 37% of employees don't use AI simply because their colleagues don't.[11] Peer influence is not just a positive driver; its absence is an active barrier. Social proof works in both directions, and in the absence of visible adoption around them, most people default to inaction.

A qualitative study of Microsoft Copilot users (arXiv, February 2025) confirmed the pattern: 7 out of 10 had bypassed institutional training materials entirely in favour of peer learning and direct experimentation. The takeaway is not that training is useless — BCG's data suggests that a minimum of five hours of structured exposure, ideally in person, correlates strongly with sustained regular use. The takeaway is that budget and energy invested in cascade e-learning modules is structurally less efficient than budget invested in facilitating informal exchange between people doing the same work.

Professional Identity as the Unspoken Barrier

Brynjolfsson, Li, and Raymond's randomised controlled trial — 5,172 participants, published in the Quarterly Journal of Economics — found that novices benefit most from AI assistance, with an average 15% productivity gain.[2] In a knowledge-intensive context, this creates a problem the literature has not yet directly addressed: if the tool most helps those with the least expertise, how do senior practitioners with strong professional identities interpret that signal?

Shata documents this explicitly. Practitioners who opted out of GenAI articulated their resistance not as fear of replacement but as a threat to the integrity of their own work.

Hartyándi's research identifies what they call the "Janus face" of AI perception — the coexistence of genuine interest and genuine threat in the same individual, where affective distrust can block adoption even when cognitive openness is present.[13] In contexts where expertise is the primary product — professional services, advisory, research — this dynamic is amplified and poorly understood. The threat isn't to employment; it's to the proposition of value itself. Designing programmes that treat this as an irrational resistance to manage, rather than a coherent professional concern to address, produces compliance without adoption.

The Verification Paradox

A less-discussed mechanism, absent from most adoption frameworks, is what might be called the verification paradox. In high-accountability work — consulting, legal, audit, research — the productivity promise of generative AI is structurally complicated. The faster a tool generates content, the more labour is required to verify it. The QJE data showing 15% productivity gains comes from a contact centre context. In contexts where a factual error in a deliverable destroys professional reputation and client trust, the calculus changes.

Prasad et al. (2025, N=470) make the point sharply: a practitioner who doubts the reliability of AI outputs will not cross from intention to actual use, regardless of their initial motivation. The tool must be trusted before it will be used in high-stakes situations. And trust, once lost, is expensive to rebuild.[4]

The Trust Window

Deloitte's TrustID data documents a 31% erosion in employee trust toward employer AI initiatives over just three months following poorly calibrated rollouts.[9] S&P Global (2025) found that 42% of companies had abandoned the majority of their AI initiatives before reaching production — a 2.5x increase in a single year.[12] Gartner places generative AI firmly in the Trough of Disillusionment, with under 30% of CEOs satisfied with ROI despite an average investment of nearly $2M per company.[11]

MIT NANDA identified a "learning gap" at the heart of generative AI systems: LLMs do not retain feedback between sessions, do not adapt to context over time, and do not improve automatically with use. The implication is that the benefits of AI in a professional context accrue to the individual who has learned to use the tool well, not to the tool itself. The expertise is in the human-AI interaction, not the model — which reframes onboarding from a one-time event into an ongoing capability-building practice, and makes the quality of peer networks the primary infrastructure for sustaining adoption.[8]

The strategic implication is about sequencing more than speed. Each failed or over-promised initiative consumes a non-renewable stock of trust. Getting there slowly and building a foundation of genuine usefulness is not caution; it's strategy.

What Actually Works: Three Entry Points, Not One

The research converges on a model that is hybrid and differentiated, not monolithic. Westover's comparison of 127 organisations found that a hybrid structure — central AI governance combined with distributed implementation teams embedded in specific business contexts — achieved a 74% success rate for AI initiatives, against 12% for organisations without both dimensions. Contextualised training produced 36% more retention than generic training.[10]

Different populations require different entry points, and conflating them into a single programme is one of the most consistent failure modes in the literature.

For practitioners under delivery pressure who primarily need efficiency — the entry point is a quick demonstration of a specific irritant being resolved, delivered by someone doing the same work, measurable in minutes saved. Not a training module; a live example from a peer.

For those with lower baseline digital comfort — the entry point is protected experimentation time, a simple interface, and the explicit removal of performance pressure during the learning phase. The target is general technological comfort before AI-specific adoption. The HBR framework AWARE identifies three psychological needs at stake: competence (feeling effective), autonomy (retaining control), belonging (maintaining meaningful relationships). A programme perceived as prescriptive or surveillance-adjacent frustrates all three simultaneously.

For senior practitioners with strong professional identities — the entry point is co-construction, not demonstration. The AI is positioned as an amplifier of their existing expertise, not a replacement for it. The framing that works is not "here is what AI can do" but "here is how AI makes what you already know go further." The cases that produce the highest engagement in this group are the ones they have defined themselves.

What the Literature Doesn't Cover

Almost every study in this corpus was conducted in an industrial, banking, or public sector organisation — not one examines a context where expertise is the primary product sold. The dynamics of knowledge-intensive professional services — where billing structures are built on time spent rather than output delivered, where junior work is compressed not expanded by AI, where the expertise itself is the value proposition to the client — constitute a research gap currently filled only by inference and anecdote.

The compression of junior tasks from hours to minutes by generative AI is, in a time-and-materials billing model, a direct revenue reduction. The democratisation of expertise documented by Brynjolfsson et al. is, for the senior practitioner whose knowledge premium is their primary market differentiation, a competitive threat. These tensions are not irrational, and they are not addressed by the literature. They require organisations in that position to build their own evidence base — and to be honest, internally, about the fact that the adoption challenge they face is not the same one the research describes.

The organisations that navigate this well in the next twelve months will not necessarily have spent the most, or moved the fastest, or chosen the most capable models. They will have been the most precise about what they were actually asking people to change, and the most honest about what the change was genuinely worth.

---

Research synthesis — Corpus 2024–2026 — April 2026 Sources: 80+ publications across peer-reviewed journals, working papers, and proprietary surveys. GenAI-specific sources prioritised; mixed ML/AI sources retained with explicit notation. Vendor-produced materials downgraded.