There is a strange omission at the center of the AI-adoption conversation. Vendors discuss capability. Consultants discuss strategy. Leadership discusses ROI. Almost nobody discusses trust — and when they do, it appears as a soft, hand-wavy concern, the kind of thing serious people gesture at on the way to talking about something more measurable.
This is wrong, and it is wrong in an unusual way. Trust in technology is one of the most thoroughly studied variables in the entire field of organizational and consumer research. It has been measured, modeled, and empirically validated across thirty years and dozens of major studies. The instruments exist. The findings are consistent. The construct is not soft.
What is happening, instead, is that the people building and selling AI tools have a commercial interest in not measuring trust, because measuring trust would reveal the actual bottleneck they are running into — and the bottleneck is not capability.
Acceptance research has been telling a consistent story since the late 1980s. Whether a person adopts a new technology is determined by a small number of factors that combine in well-understood ways: the expected performance gain, the expected effort, the influence of their social environment, the available infrastructure, the hedonic experience of using the thing, and — mediating most of these — trust. Trust itself decomposes into more specific risks: performance risk (will it actually do what it claims), safety risk (will it harm me or those I am responsible for), privacy risk, financial risk, social risk.
When someone runs the numbers, two findings keep showing up. The first is that trust is not a feeling. It is a structured judgment made of specific risks, each of which can be measured and each of which can be addressed. The second is that the strongest predictors of trust are usually the most operational: performance risk and safety risk dominate. People do not refuse new technology because they are anti-progress. They refuse new technology because they have correctly assessed that the promised benefit is not yet reliable enough to bet their work on.
The current AI conversation systematically ignores both findings.
The first ignorance shows up as the assumption that adoption is a marketing problem. If we just communicated the benefits better, people would use it. This is wrong, and acceptance research has been pointing out that it is wrong since the original Technology Acceptance Model in 1989. Communication does not move trust. Performance moves trust. Safety moves trust. Repeated, observed, reliable behavior of the system moves trust. Slide decks do not.
The second ignorance shows up as the assumption that trust is a downstream consequence of capability. Once the model is good enough, trust will follow. This is also wrong, and the data here is even older than the AI conversation. Trust lags capability — sometimes by years, sometimes by decades — because trust is a function of experienced reliability, not announced reliability. Self-driving cars passed the technical-capability bar for many use cases years before they passed the trust bar, and the gap between the two is the entire reason the rollout has been slower than the technologists predicted. Exactly the same gap is forming, right now, around AI in organizations.
The interesting consequence is that the lever almost no one is pulling is the most powerful one available. If you want adoption inside your organization, you do not need a better model, a better dashboard, or a better slide deck. You need a structured, repeatable demonstration that the thing works in your context, on your tasks, with the failure modes named and bounded. Trust is built by removing specific risks, not by adding general enthusiasm.
This sounds obvious. It is also, in practice, almost never done.
There is a sharper version of this for organizations specifically. Trust in AI is not held by the organization. It is held by individual people, one at a time, each running their own private acceptance calculation. An organization does not “trust” Claude or Copilot or any other system — its employees do, or do not, in numbers that aggregate into something that looks like organizational adoption from a distance. Which means the question is never does our company trust AI but what is the distribution of trust across our team, and what specific risks are dominant for which people.
That is a measurable question. It is not being measured. The vendors who could measure it have no interest in doing so. The consultants who could measure it have no instruments. The leadership who could commission it does not know the literature exists.
The license is paid. The trust is not. That is the gap.
Apparatus: The acceptance-research lineage referenced is Davis, Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology (MIS Quarterly, 1989, the original TAM); Venkatesh, Morris, Davis & Davis, User Acceptance of Information Technology: Toward a Unified View (MIS Quarterly, 2003, UTAUT); and Venkatesh, Thong & Xu, Consumer Acceptance and Use of Information Technology: Extending UTAUT (MIS Quarterly, 2012, UTAUT2). Trust-risk decomposition draws on Featherman & Pavlou, Predicting E-Services Adoption: A Perceived Risk Facets Perspective (International Journal of Human-Computer Studies, 2003). Composite operator and leadership conversations from the DACH organic sector, Q1–Q2 2026.