The Evidence You're Waiting For Doesn't Exist Yet. The Evidence You Need Has Existed for Fifty Years.
You're betting on AI either way. The only question is what you anchor the bet to.
AI is new. The human being handed it is not. That sentence is the whole argument of this entry, and the whole reason this publication exists. The long-term studies of AI adoption you would love to base your decisions on cannot exist yet, and everyone deciding anything about AI right now is doing it without them. What follows is the honest version of what to do about that: what the missing evidence cannot tell you, what fifty years of behavioral science already can, and the question that settles the argument with anyone who says it is too early to act.
The studies you want cannot exist yet
Let us say the quiet part first: the evidence everyone would prefer does not exist. Nobody has followed a thousand companies for ten years and come back to tell us which AI strategies actually paid. Not because researchers are lazy, and not because the field is unserious. The reason is simpler and more forgiving than that: the technology is too young for the question. The kind of AI now sitting in workplaces has only existed for a few years, and research that watches organizations over the long term needs, unavoidably, the long term: years of observation, and then the analysis, the review, and the publication stacked on top. You cannot argue with a clock, and there is nobody here to blame.
But there is a second, crueler layer, one that even people who read research for a living tend to miss. The object of study mutates faster than the instrument that measures it. Suppose the perfect five-year study of today's models launched this morning. By the day it publishes, the models it studied will have been replaced several times over, the interfaces will have changed shape, and the price of intelligence will have moved. The study would arrive as history, not guidance. This is the evidence lag, and AI at work sits deep inside it: the technology forces decisions on a schedule the science structurally cannot match.
Now put that gap next to the field data on what happens meanwhile. The opening entry of this publication laid out the numbers: 88% of companies now use AI, and about 6% see profit impact from it. MIT's NANDA project put it even harder: roughly 95% of enterprise generative-AI pilots deliver no measurable P&L impact. Read those together with the lag and the situation gets clear. The losing battle is not a forecast. It is happening now, during the wait, to the people waiting.
There is no spectator seat
The instinctive response to missing evidence is to wait for it, and in most of life that instinct serves you well. Here is why it fails for AI: waiting is not a neutral position. The field evidence on shadow adoption says your people are already using AI, at home confidently and at work quietly, whether or not leadership ever made a decision. So the organization that waits is not outside the game. It is in the game with no plan, absorbing the risks of unguided AI use while collecting none of the returns, and falling behind on the one asset that compounds: competence, which is built by practice the waiting organization is not doing.
Which dissolves the comfortable framing entirely. The choice was never evidence versus no evidence, because nobody gets that choice; the longitudinal evidence does not exist for anyone, including every vendor, consultant, and competitor currently projecting confidence. Everyone deciding anything about AI right now is betting without it. The only variable actually under your control is what you anchor the bet to.
The stable layer
So anchor to the thing that is not moving. Look at where the failures concentrate, in the numbers above and in every system this publication has examined in production: not in the model's benchmark scores, which keep improving, but at the junction where the technology meets people. Whether they hide it or share it. Whether they trust it too much, too little, or exactly as far as it has earned. Whether one person's quiet competence becomes an operation's capability or stays personal leverage. That reading of the numbers is ours, and we mark it as ours. But notice what it implies if it is even half right: the layer where AI succeeds or fails is the one layer that has not changed, and it is the best-studied layer in the whole system.
Because the psychology of a human meeting a new tool did not ship last quarter. Whether people feel safe enough to try things and admit errors: Amy Edmondson operationalized psychological safety in 1999, on the back of a research tradition older still. Whether people believe they can learn the tool, the belief that carries adoption into return: Albert Bandura spent four decades building the self-efficacy literature. How training must be designed so skill survives contact with real work: Robert Bjork named desirable difficulties in 1994. Why exploration requires a secure base to return to: Bowlby and Ainsworth built that evidence over half a century, and Feeney and Thrush showed it operating between adults in 2010. And the newest wing, how people over-trust and under-trust automated advice, is accumulating fast in the decision sciences right now. AI is new. The human being handed it is not. When this publication makes a claim, this is the ground it stands on: not predictions about a mutating technology, but decades-deep findings about the constant in the equation.
Four disciplines that stand in for time
Anchoring is the first discipline, not the whole method. Acting honestly inside the evidence lag takes four habits, and you can audit every one of them against this publication's own record. First, mark the seams. Established science and proposed synthesis never wear the same label here: what is verified gets a citation, what is ours gets claimed as ours, and the methodology page keeps the full ledger of which is which. Second, bet in public. A proposed theory is only honest if it can lose: The Secure Base in AI went out with a falsifiable prediction attached and the declared consequence that if the data comes back flat, the theory dies in public.
Third, verify before citing, every time. Every statistic in this publication traces to a primary source read before use, because in a field this noisy, secondhand numbers are how honest people end up lying. And fourth, the one that pays your operation directly: build local evidence streams. The instruments in these entries, keeping score of the tool by task type, running the same job twice and watching the path, sending the system's monthly numbers to the team that runs it, are all the same move: bookkeeping for judgment. An operation that keeps those books is running a small longitudinal study of the only organization whose results it needs: its own. The field's evidence is years out. Yours can start Monday.
The question that settles it: what is the alternative?
Everything above can sound like special pleading until you put it next to the other options actually on the table. There are four, and it is worth watching what each one anchors its bet to.
Waiting for the science anchors the bet to a date that keeps receding, and pays the meanwhile-price we already walked through: shadow adoption without guidance, risk without return, a competence gap that compounds while everyone stands still. It is the most respectable-looking option on the list, and the only one guaranteed to convert the evidence lag into pure loss. Trusting the vendor's evidence anchors the bet to testimony from the party selling the outcome, and the one rigorous test this publication keeps citing should end that habit alone: the most intuitive trust feature in the industry, showing users the AI's explanations, failed to reduce how often people accepted wrong answers in controlled trials, and sometimes made it worse. Vendor intuition is not just unproven. It has already been caught pointing backwards.
Copying what everyone else does anchors the bet to the average adopter, and the average adopter is the losing side of the statistic: when 6% profit, imitation is a 94% bet against yourself, dressed as prudence. And running the standard IT playbook, licenses, logins, a training slide, a go-live date, anchors the bet to the assumption that this is an installation problem, which is exactly the assumption the failure numbers keep punishing: the technology installs fine, and the returns die at the human layer the playbook never touches.
Four alternatives, four anchors: a receding date, conflicted testimony, a failing average, a wrong assumption. That is the field. Against it stands the option this publication was built on: anchor to half a century of tested science about the humans doing the adopting, mark what is proposed, bet in public, and let your own operation start generating the evidence the journals cannot hand you yet. We are all about research. That is exactly why we refuse to idle waiting for it: respecting evidence means using the mature evidence that exists, not treating its absence in one field as permission to guess in all of them.
What sound looks like on a Monday
None of this requires a research department. Locate your processes on the map before any tool talk. Size the human loop to the task, and spend the deliberate pause where a wrong answer is expensive. Give people a safe place to build the skill before the skill is needed in front of a client. Start the tally that turns trust into a number about a task. And write down, honestly, what you are assuming versus what you know, because that habit, more than any tool decision, is what the evidence lag actually demands of you.
The studies will come, and this publication will read them the day they land, and will report what they say about our own predictions, whichever way they cut. Until then the question stands, for us and for anyone selling you certainty in either direction. You are betting on AI either way. Anchored to what?
If you want the anchored version of the very first decision, the Impact vs. Risk Matrix is the free one-page tool: where the impact is high and the risk is low, start there.
Frequently asked questions
Should companies wait for more research before adopting AI?
Waiting feels prudent, but it is not a neutral position: it is a bet that losing slowly is safe. Three things happen during the wait. Your employees adopt anyway, privately and without guidance, so the organization absorbs the risks of AI use without the returns. The competence gap compounds, because AI skill is built through practice that the waiting organization is not doing. And the evidence being waited for keeps receding, because the technology changes faster than long-term studies can be run on it. The alternative to waiting is not recklessness; it is anchoring decisions to the mature behavioral science that already exists about how people adopt tools, build skill, and misplace trust.
Is there scientific evidence for AI adoption strategies?
Not the kind most people mean by the question: there are no long-term, organization-level studies of AI adoption outcomes, because the deployed technology is only a few years old and keeps changing. But the evidence that exists is stronger than the field's reputation suggests, if you look one layer down. The failures happen where AI meets people, and that layer is one of the best-studied territories in behavioral science: psychological safety (Edmondson, 1999), self-efficacy (Bandura), training design and desirable difficulties (Bjork, 1994), secure-base exploration (Bowlby; Feeney and Thrush, 2010), and a growing decision-science literature on overreliance and trust calibration (Buçinca et al., 2021; Schoeffer et al., 2025). The technology is new. The science of the humans using it is not.
Why are there no long-term studies of AI in the workplace yet?
Two honest reasons. First, time: workplace AI at current capability has only existed for a few years, and longitudinal organizational research needs multi-year windows plus peer review and publication on top. Second, and less appreciated, the moving-target problem: by the time a multi-year study of today's models is published, the models it studied will have been replaced several times over. The object of study mutates faster than the instrument that measures it. This is nobody's failure; it is a structural property of studying a fast-moving technology. It also means the rational response is not to idle until the studies arrive, but to anchor to research whose object does not mutate: human behavior.
What is the evidence lag in AI?
The gap between when a technology forces decisions and when settled science about it can exist. AI at work is deep inside that gap: organizations must make adoption decisions now, while the longitudinal evidence that would guide those decisions is years away and aimed at a moving target. The evidence lag does not excuse guessing. It defines what disciplined practice looks like: anchor claims to adjacent mature science, mark clearly what is established versus proposed, put falsifiable predictions on record before the results come in, and instrument your own operation so it generates local evidence while the field catches up.
How can a business generate its own evidence about AI?
By instrumenting the work it is already doing. Keep score of AI outputs by task type, right ones and caught ones, so trust is calibrated to your own record instead of vendor benchmarks. Put anything in your operation that gets called an automation through the same job twice: if it works the same way both times, it is a fixed system you can audit and improve; if an AI is choosing the route and the route changes, its behavior cannot be studied the same way and it deserves closer watching. Send the system's monthly numbers to the team that operates it, so the trend is visible to the people who create it. None of this is a research program; it is bookkeeping for judgment. An operation that does it accumulates something the published literature cannot give anyone yet: longitudinal evidence about AI in the one organization that matters to it, its own.
Sources
- Stanford Institute for Human-Centered AI (2026). The AI Index Report.
- McKinsey & Company (2025). The State of AI. Global survey.
- MIT NANDA (2025). The GenAI Divide: State of AI in Business.
- Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383.
- Bandura, A. (1997). Self-efficacy: The exercise of control. W. H. Freeman.
- Bowlby, J. (1988). A secure base: Parent-child attachment and healthy human development. Basic Books.
- Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.
- Feeney, B. C., & Thrush, R. L. (2010). Relationship influences on exploration in adulthood: The characteristics and function of a secure base. Journal of Personality and Social Psychology, 98(1), 57–76. https://doi.org/10.1037/a0016961
- Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188. https://doi.org/10.1145/3449287
- Schoeffer, J., Jakubik, J., Vössing, M., Kühl, N., & Satzger, G. (2025). AI reliance and decision quality: Fundamentals, interdependence, and the effects of interventions. Journal of Artificial Intelligence Research, 82. https://doi.org/10.1613/jair.1.15873
- Reich, A., Wolfe, D., Price, M., Choe, A., Kidd, F., & Wagner, H. (2026). Safety first: Psychological safety as the key to AI transformation. arXiv:2602.23279
Published: July 13, 2026