Why AI adoption is an identity problem before it is a technology problem.

RAND's 2024 study of artificial-intelligence projects landed on a number that surprised almost no one in the room: more than 80% of enterprise AI initiatives fail. The figure is roughly double the failure rate of non-AI IT projects.^[2024] What surprised people was the diagnosis. The researchers traced failure not to model performance, not to data availability, not even to budget. They traced it to misaligned stakeholder understanding — a polite way of saying that the humans involved did not agree on what they were doing, why they were doing it, or who was accountable when it was done.

The picture from the major institutional research programs is the same. McKinsey's 2025 State of AI finds 78% of organizations now use AI in at least one business function, but only 1% of executives describe their generative-AI roll-outs as mature.^[2025] Gartner forecasts at least 30% of generative-AI projects will be abandoned after proof-of-concept by the end of 2025.^[2024]BCG's longitudinal work on AI value capture continues to find roughly 70% of value comes from people, processes, and adoption — not from the algorithm and not from the data.^[2020]

The argument of this article is that these numbers are not describing a technology problem. They are describing an identity problem. Enterprise AI fails between Level 0 and Level 2 of the PolyCognitive maturity model — between "no adoption" and "value with risk" — and what stalls organizations at each of those points is not the absence of tools. It is the unresolved human question of what people are for, once the tools arrive.

Begin with a finding from behavioral economics that has gone quietly load-bearing in the AI literature. In a series of forecasting experiments, Dietvorst, Simmons and Massey showed that people shown an algorithm and a human making the same prediction will consistently abandon the algorithm after it makes a single visible error — even when the algorithm is more accurate overall.^[2015]The effect, which they named "algorithm aversion," is robust across domains and replicates reliably.

What makes the aversion finding important is what came after it. A second wave of research, led by Logg, Minson and Moore, looked at who exhibits aversion and who does not. The answer was sharp: lay people often show the opposite reaction — what the authors called "algorithm appreciation" — weighing algorithmic advice more than identical advice from a human. But domain experts, the people whose judgment is being recommended against, reliably weight the algorithm less.^[2019] Resistance, in other words, is not evenly distributed. It is concentrated at exactly the layer of the organization where the most experienced practitioners sit.

This is why blanket adoption campaigns reliably stall. The CHRO announces an enterprise license, the CEO posts on LinkedIn, the heads of function nod, and the people whose work would actually be changed — the senior practitioners — quietly do not change. They are not lazy and they are not behind. They are doing exactly what the evidence predicts: weighting the algorithm less than the organization assumes they will.

Why does resistance concentrate at the senior practitioner layer? Cao and colleagues, writing in Computers in Human Behavior, ran the most careful empirical test of the question to date. Their answer, drawn from a multi-study programme across knowledge-work settings, was unambiguous: employees resist AI most strongly when it threatens occupational identity, not when it threatens productivity.^[2023]The framing the organization usually adopts — "this will save you time" — is at right angles to the felt threat, which is "this will make me legible as someone whose skill is now optional."

Kellogg, Valentine and Christin's synthesis in Academy of Management Annals extends the picture. Their review of algorithmic management practices identifies six new forms of organizational control that AI introduces — recommendation, restriction, recording, rating, replacing, and rewarding — each of which provokes a predictable form of resistance.^[2020] The pattern they document is consistent: organizations that treat AI rollout as a technology project fail at adoption; organizations that treat it as a labor-relations project succeed.

The implication for the human layer of leadership is not sentimental. It is structural. The leader's first move at Level 0 is not to push harder on the tool. It is to surface what is being protected and to name what gets adopted toward. Until the identity question is on the table, the productivity argument cannot land. Microsoft's 2024 Work Trend Index makes the cost of skipping this step visible: 75% of knowledge workers already use AI at work, but 78% bring their own tools rather than use the sanctioned ones, and only 39% have received any company training.^[2024] The organization is not absent from AI — it is absent from the conversation about AI, while the work continues underneath.

Suppose the identity question is resolved and adoption climbs. What then? The PolyCognitive framework names a second stuck-point — Level 1, Adoption Without Value — that is harder to see than Level 0 precisely because the usage telemetry looks good. Tools are used. Dashboards are green. The metrics that matter to the business have not moved.

The Upwork Research Institute's 2024 study of knowledge workers placed a number on this: 77% of employees report that AI has addedto their workload rather than reduced it, and 47% say they don't know how to achieve the productivity gains their employers expect.^[2024] This is the productivity-theater layer. Activity rises. Drafts get longer. Decks get prettier. Meetings get summarized. The organization's output looks busier and feels more polished, but the cycle times that matter to customers and the error rates that matter to operations stay flat.

The Dell Innovation Index makes the leadership consequence concrete: only 31% of leaders can quantify the impact of their generative-AI initiatives on business outcomes, despite 76% saying the technology will be significant or transformative.^[2024] When leaders cannot point to outcome shifts, accountability drifts toward what can be measured — tokens used, prompts run, time spent in the tool — and the team learns that using AI is the deliverable.

The accountability move is older than AI. Hammer wrote the essential warning in 1990: organizations that "pave the cow paths" by automating existing processes capture marginal gains, while organizations that obliterate and reengineer capture order-of-magnitude improvements.^[1990] The human-layer translation in 2026 is the same instruction in different language: stop rewarding AI activity, start rewarding outcome shifts, and define value in metrics that pre-date AI in the function.

Identity and accountability resolved, the next layer surfaces. Level 2 — Value With Risk — is the level at which the wins arrive and the side effects arrive with them. The most-cited illustration in the literature is the Samsung incident: in May 2023, Bloomberg reported that Samsung Semiconductor engineers had leaked confidential source code and meeting notes into ChatGPT on three occasions within 20 days, resulting in an enterprise-wide ban on generative AI.^[2023]The cases since have multiplied. Cyberhaven Labs's 2024 telemetry showed sensitive corporate data sent to AI tools growing 485% year-over-year, with the share of data classified as sensitive nearly tripling from 10.7% to 27.4% of all AI-bound traffic.^[2024]

What makes the shadow problem a human-layer problem rather than a policy problem is the predictable failure mode of policy at this level. Blanket prohibitions drive usage underground. The work still has to get done; the policy says the sanctioned tool is too slow or unavailable; the employee uses an unsanctioned tool from a personal browser. The organization gets the worst of both worlds — the risk without the visibility.

The human move at Level 2 is to translate risk into rules that survive contact with the work, and to design enforcement that does not punish disclosure. The instruction is unusual because it requires the leader to make the sanctioned path faster than the shadow alternative, not slower. Most governance regimes do the opposite. They make the sanctioned path safer by making it slower, and learn — after a Samsung-class incident — that they have been optimizing for the wrong thing.

The strongest empirical finding on the human side of enterprise AI adoption is also the simplest. Ransbotham and colleagues, in the MIT Sloan Management Review × BCG longitudinal study, surveyed approximately 3,000 managers across 28 industries and found organizational value from AI is roughly five times more likely when the CEO is personally involved in AI strategy and when AI is integrated into employees' daily workflows.^[2023]McKinsey's equivalent finding is that AI leaders versus laggards are roughly three times more likely to have their CEO and senior team directly engaged in setting AI strategy.^[2025]

The pattern is older than AI. Westerman, Bonnet and McAfee's study of 400+ large firms found that "Digital Masters" — organizations with strong digital capability and strong leadership engagement — outperformed peers by 26% on profitability. Organizations with technology capability alone underperformed.^[2014] Fountaine, McCarthy and Saleh, writing in Harvard Business Review, isolate the mechanism: the biggest barriers to AI value capture are organizational and cultural — shifting from siloed to interdisciplinary work, from experimental to evidence-based decisions, from rigid to agile — and only leaders sitting on the relevant authority can move those.^[2019]

Engagement, in this literature, is not visibility. It is not the CEO posting about AI on LinkedIn. It is the CEO making decisions with AI in the room, on the timescales that the rest of the organization watches. Senior engagement, operationalized, is a human-layer move: it tells everyone below what is permissible to adopt, what is rewarded to demand, and what is punished to skip.

The PolyCognitive framework names five capabilities, each unsticking one of the five maturity levels, each running on a human and an AI track. The human track of each capability is the translation of the research above into a leadership move at a specific stuck-point.

1. H.1 · Unsticks Level 0

   Surface and resolve the real reasons people resist. Resistance is identity, not productivity. The leader's first move is to name what is being protected and to make the choice to adapt safer than the choice to wait. Driving AI Adoption →

2. H.2 · Unsticks Level 1

   Hold people accountable for outcomes, not AI activity. Anchor accountability in metrics that pre-date AI in the function. Refuse to substitute tool usage for outcome shifts. Extracting Value from AI →

3. H.3 · Unsticks Level 2

   Make risk concrete and set rules people will actually follow. Translate categories into per-workflow do/don't rules. Make the sanctioned tool faster than the shadow alternative. Managing AI Risk →

4. H.4 · Unsticks Level 3

   Help people release workflows their identity is built around. Distinguish the workflow from the identity it carries. Redesign roles before redesigning the process. Run transitions in cohorts. Redesigning Workflows →

5. H.5 · Unsticks Level 4

   Protect the skills AI can't replace by making people use them. Build deliberate practice into normal work. Protect apprentice pipelines from premature AI handoff. Rehearse recovery. Preserving Human Edge →

The conclusion the literature converges on, across institutional research programmes that rarely agree on anything, is that the human layer of AI adoption is decisive in a way the algorithm layer is not. The framework presented on this site is one articulation of what that means operationally: five stuck-points, five capabilities, two layers. The AI layer matters. But it cannot do the work the human layer must do first.