Psychology of adoption

Editorial note. This chapter preserves the earlier concise English edition. A full translation of the expanded Spanish edition published on July 21, 2026 is in progress.

Rogers describes the macro curve —innovators, early majority, laggards— and The four links already place it in the fourth link. But between the curve and the person on Tuesday at 5 p.m. there is a space the method cannot leave implicit: how it feels to use a system that suggests, classifies or decides.

At Distribuidora Norte, the salesperson who reads the “customer at risk” alert as their own tool adopts; the one who reads it as a performance ranking before the manager avoids it. The difference is not in the model but in the psychological design of the flow —and in whether diagnosis mapped first the incentive that blocked sharing information.

An AI can be technically correct and psychologically untouchable. This chapter names the phenomena that explain why a well-built AI still fails in those who should use it —and translates them into design decisions.

Key phenomena

Phenomenon	What it is	Design response
Automation bias	Accepting what the AI says without questioning, even when the human knows more	Moments of mandatory human judgment where the AI stays silent and asks for judgment
Control paradox	Too much control overwhelms; too little alienates	Override levels calibrated by role and by case risk
Identity threat	”The AI says my experience doesn’t count”	Design that shows human contribution; AI as assistant, not judge
Calibrated trust	Blind distrust vs. blind trust	Transparency of limits, explainability proportional to risk

These phenomena are not excuses to avoid automating: they are requirements so automation returns power instead of taking it away.

Trust, in numbers

The fourth phenomenon —calibrated trust— is no longer a design intuition: it has been measured at scale. The global study by the University of Melbourne and KPMG (Gillespie, Lockey et al., 2025; more than 48,000 people across 47 countries) finds a scissors that orders the whole chapter: 66% already use AI regularly, but only 46% are willing to trust it. Heavy use, low trust —exactly the terrain where psychological design decides whether adoption holds or decays. And the floor-level detail sharpens the reading: 58% use it at work, but 57% hide that they use it; 66% trust what the model returns without verifying it; only 47% received any training. Much use, little verification, hidden use, and almost no one trained. This is not distrust of AI: it is miscalibrated trust —blind where it should doubt, hidden where it should be a conversation—, which is precisely what a well-designed flow can correct.

These are global averages across 47 countries: they illustrate the shape of the problem, not a number to promise a client. For local scale, the anchor remains regional evidence.

The erosion of the scaffold

There is a cost of adoption that shows up in no usage metric: what happens to a person’s skill when AI takes off their plate, precisely, the hard tasks through which they learned the craft. Matt Beane studies this in operating rooms and warehouses (The Skill Code, 2024): the path by which a novice becomes an expert rests on three conditions —the three C’s: challenge (facing tasks at the limit of one’s ability), complexity (seeing the whole problem, not a slice) and connection (the bond with an expert who corrects in context)—. When an intelligent system absorbs the entry-level tasks —the “easy” ones that were really the practice ground—, all three weaken at once: the novice is no longer challenged, no longer sees the whole problem, and stops needing the expert. Productivity rises today; the training of tomorrow’s expert goes dark with no one measuring it.

I call this the erosion of the scaffold: the silent decay of the path by which an organization forms its own experts. It is not automation unemployment —that is seen and debated—; it is a loss of future capacity that the productivity metric does not register, because that metric improves while the scaffold falls. It is the second, quieter floor of identity threat: there the expert feels the AI discredits what they know; here, the system stops producing the experts who would come next. The answer is not technological: it is deciding which learning tasks are reserved for the person even though the AI could do them, because their value lies not in today’s output but in tomorrow’s skill.

Technostress of the gray zone

Generative AI brings a technostress different from earlier waves. The analysis by Högemann et al. (2025) of young professionals —those supposed to be most comfortable with the tool— finds that their stressors are not about usability but about ambiguity: not knowing whether using it is allowed, fear of exposing data or infringing rights, a perceived dependence that unsettles (“do I still know how to do this without the AI?”). It is not produced by a hard interface; it is produced by the absence of clear rules. It speaks directly to the 57% who hide their use, and to the erosion of the scaffold: the one who hides fears the gray zone; the one who fears dependence feels the question about their future skill. The bridge’s answer is not a motivational memo: it is removing the gray zone —explicit use policy, clear permission, an honest framing of what is delegated and what is not—, because much of that stress is sociotechnical debt felt in the worker’s body.

The team AI reshapes

The psychology of adoption does not end with the individual: AI changes the shape of the team. The field experiment by Dell’Acqua et al. (2025), with 776 professionals at Procter & Gamble, leaves two findings. First: individuals with AI matched the performance of teams without AI —one well-assisted person reached what previously required a group—. Second, more sociotechnical: AI broke functional silos. Without AI, R&D people proposed technical solutions and Commercial people commercial ones, each from their discipline; with AI, both produced balanced solutions, as if the tool lent them the other side’s vocabulary.

This is not a promise of replacement, it is a hint about redesign: if AI makes a profile cross the border of its silo, diagnosis must ask what collaboration becomes possible that was not before —and what is lost if the team dissolves into isolated individuals, because Beane’s connection is precisely what a team provides and a lone individual does not—. The study also observed more positive emotional responses to the conversational interface: AI covered part of the social role of the teammate. Powerful and risky at once —a teammate who motivates but does not train is, once again, scaffold that erodes—. And the same caveat as always: it is a resource-rich multinational with bounded tasks; it illustrates a mechanism, it does not promise a number.

ADKAR and change fatigue

The ADKAR model —developed by Prosci for change management— names five thresholds each person must cross: Awareness (why something is changing), Desire (willingness to participate), Knowledge (how to do it), Ability (real capacity to do it day to day) and Reinforcement (support so the change sticks). When adoption stalls, asking which of the five is missing is more useful than talking about generic “resistance”. If desire is missing, more training does not help: incentives and fears must be addressed. If reinforcement is missing, that explains why three months later everyone went back to the spreadsheet. And the KPMG scissors puts a number on a concrete threshold: with only 47% trained, much of the fabric is stuck in knowledge and ability —not in ideological distrust—.

Change fatigue appears when teams live sprints without real autonomy, or when each quarter announces a new “transformation” without closing the previous one. The result is floor-level cynicism: “this too shall pass”. The answer is not more communication or more hype: it is consolidating one step before opening the next front.

Connection with HCAI

Shneiderman proposes Human-Centered AI (HCAI): high automation and high human control are not opposites. In the agentic layer —when software decides and executes on its own— that translates into judgment in the design: auditable rules, points where the system must escalate to a person, revocation mechanisms. User psychology defines where to place those control points so they are actually used. An “override the AI suggestion” button that takes twelve clicks is not used; one that appears when risk warrants it —legitimizing the operator’s experience— is. Sociotechnical design is also design of social permission to contradict the model when needed.

Where it enters the method

Link 1: interviews detect identity fears and scars from failed pilots.
Links 2–4: UX, copy and flows incorporate confirmations, visible override, feedback to the expert when the model is wrong.
Sociotechnical hygiene: workarounds from poorly managed distrust are a signal of psychological failure even when usage metrics look acceptable.

Adoption is not a marketing appendix or persuasion manual: if resistance brings an operational reason, it is a requirement —not an obstacle to overcome.