Psychological safety gets your team to try AI. Job design keeps them using it.
Most AI rollouts stall at the same place: a few people use the tools every day and everyone else drifts back. Two whitepapers published in February 2026, built on one survey of 2,257 employees at a global consulting firm, explain why. Trying AI and sticking with AI are two different decisions, and different things move each one. Here is what the data says, and the four plays we run with AlexDLY customers because of it.
The short version
Paper 1: safety opens the door
A one point rise in felt psychological safety raised the odds of adopting AI by about 30 percent. Once people had adopted, safety had no effect on how often or how long they used it.
Paper 2: the job decides if it stays open
Skill variety and autonomy were the strongest predictors of trying AI. Rising workload went with deeper use. Fear for one's standing went with shallow use.
The play
Spend weeks one and two making it safe to try and finding each role's recurring task. Spend weeks three and four on discretion, the human gate, workload, and the status story.
Where this research comes from
Psychological safety is Amy Edmondson's construct from Harvard Business School: the belief that you can ask a question, try something, or admit a mistake without being punished for it. Her 1999 items are still the standard measure, and the 2026 edition of The Fearless Report, the global benchmark she runs with Connie Hadley and Mark Mortensen, is built around how AI is reshaping that safety at work.
The two papers here use those items. Aaron Reich, Diana Wolfe, Matt Price, Alice Choe, Fergus Kidd, and Hannah Wagner, a team spanning Avanade, Kyndryl, Rotman, and Seattle Pacific University, surveyed every eligible employee of one multinational consulting firm. The survey took about 25 minutes. Of the 2,257 people who answered, 1,256 said they used AI tools in their regular workflow and 1,001 said they did not. Adoption was flat across tenure, region, and most levels, which makes the sample useful: the differences you see below are not explained by who was in the room.
- Safety First: Psychological Safety as the Key to AI Transformation, arXiv 2602.23279, February 26, 2026.
- Work Design and Multidimensional AI Threat as Predictors of Workplace AI Adoption and Depth of Use, arXiv 2602.23278, February 26, 2026.
Paper 1: psychological safety opens the door
The first paper asks a plain question. Does feeling safe predict whether people adopt AI, how often they use it, and how long they have used it? Safety was measured with eleven items across three dimensions: individual safety, team respect, and team learning. Adoption was a single yes or no.
The answer splits cleanly. Safety predicted adoption with an odds ratio of 1.30, meaning each one point rise on the scale went with roughly 30 percent higher odds of using AI at all. That effect held for analysts, consultants, managers, and directors, for people with under a year of experience and over ten, and in Europe, North America, and growth markets. None of those factors changed it.
Among the 1,256 people who had already adopted, safety predicted nothing. Not frequency, not duration. The coefficients were close to zero.
The authors size the effect honestly. It sits in the same range as social influence and facilitating conditions in the technology acceptance literature, below perceived usefulness. Safety is a real lever, and it is a first-week lever.
Paper 2: the job decides whether it stays open
The second paper models the same 2,257 people with two other sets of predictors. Four job characteristics from the Hackman and Oldham model: autonomy, skill variety, task significance, and feedback. And four kinds of AI threat: work is changing under me, I am losing control, my skills are losing value, my status is at risk. All eight entered at once.
For adoption, skill variety was the strongest predictor in the model, odds ratio 1.37, with autonomy next at 1.22. Task significance and feedback did nothing. Among the threats, only loss of control showed a small positive link to adoption, which the authors read as people adopting where use was expected of them, then holding back on depth.
For depth of use, the picture changes again. Perceived changes in work, meaning more tasks and shifting expectations, went with both more frequent and longer use. The authors flag the cross-sectional design and say it plainly: this could be AI causing the workload or the workload driving people to AI. Either way, the heaviest users were the ones who felt work piling up.
Threat to status and position leaned negative on every depth outcome without reaching significance. Task significance predicted shorter use: people whose work carried the highest stakes used AI for shorter spells. Skill variety kept helping. And the models explain little variance overall, under five percent, which is the authors' own caution against treating any of this as a switch.
What it means for a 5 to 50 person business
Read together, the two papers describe a staged problem. Whether a person tries AI depends on feeling safe and on having a role with enough variety and discretion to find a use. Whether it sticks depends on whether the tool lands on a task that repeats, whether they keep control of how it is used, and whether it threatens what makes them valuable.
- Training alone underperforms. The authors say it directly: if adoption tracks skill variety and autonomy, then interventions that only target beliefs will miss when the job gives no discretionary space or no recurring use case.
- Mandates backfire on depth. Loss of control went with adoption and then with shallower use. People comply, then use the tool narrowly.
- Usage dashboards measure the wrong gate. A scoreboard in week one raises the interpersonal risk of a visible miss, which is the exact thing safety exists to lower.
- Watch the workload. If your deepest users are your most overloaded people, adoption is borrowing from them, and the study cannot tell you which way that runs.
- Say what stays human. Status threat dampens depth quietly. Nobody announces they are afraid for their job. They just use the tool less.
The AlexDLY adoption plays
We turned the two gates into a 30-day sprint. Each week is one play, each play is an Agenda task with a checklist, and each has one measure that matches the gate it is opening.
Week 1, make it safe to try. The leader names three tasks that are safe to try AI on with no customer exposure, and shares one AI miss of their own in writing before anyone else does. Ten minute show and tell, misses welcome. No usage scoreboard. Measure: how many people tried once.
Week 2, find the recurring task. Every person lists five weekly tasks. Pick the one that repeats and has steps. Delegate it as an Agenda task with a checklist and a deadline. The assignee runs the agent steps and keeps the human ones. Measure: one delegated task per person, run once.
Week 3, discretion and the human gate. Staff choose how AI fits their own task. Every customer-facing step is marked human-only and every outbound draft waits in the approval queue. Nothing is mandated. Delegate the same task again, because adoption is the first run and depth is the second. Measure: repeat use.
Week 4, workload and status. Write two lists. What will not be added to anyone's plate because AI exists. Which human skills get more valuable here, by name. Put both into OrgMemory so Alex answers from them and the team can see them. Measure: hours the four tasks gave back, totalled from the Agenda, not a usage percentage.
Run it inside AlexDLY
The step-by-step version lives in the Learn Center inside the product, under Use cases for the Agenda. It names the exact screens and buttons for each week, ends with a first move for today, and links the product tour. Open /learn once you are signed in, or start with the free War Room and it is the first use case Alex will point you to.
The product was built around the same two gates before we read these papers, which is why the plays fit. Delegated tasks arrive with a checklist and a prompt, so the recurring use case is concrete. Alex ticks only the steps marked for an agent and the server refuses to tick a human step, so discretion and the human gate are enforced rather than promised. Approvals sit on the calendar. OrgMemory holds the norms where Alex reads them. Read more on the approval model in how to hand follow-up to agents without losing trust.
Read the caveats before you quote the numbers
- One firm, one industry, one survey. Consultants are early adopters with direct exposure to AI's implications for their roles. Your team may differ.
- Cross-sectional and self-reported. The papers are careful to call every finding an association, not a cause.
- Small effects. The adoption models explain about three percent of variance and the depth models about four. Safety and job design matter; they are not the whole story.
- Safety was measured at the individual level, not as a team climate. The papers note that even within a team, people differ in how safe they feel.
Does psychological safety make people use AI more?
It makes them more likely to start. In the 2,257 person study, each one point rise in felt safety raised the odds of adopting AI by about 30 percent, but among people who had adopted, safety did not predict how often or how long they used it.
What predicts whether AI use sticks?
In the companion paper, skill variety kept helping, rising workload went with deeper use, high stakes work went with shorter use, and fear for one's status leaned toward shallower use. Autonomy and a recurring task matter more than another training session.
Should I publish AI usage numbers to push adoption?
Not in the first weeks. A scoreboard raises the interpersonal risk of a visible miss, which is the thing safety exists to lower. Measure who tried once, then repeat use, then hours given back.
Run the 30-day sprint with Alex
AlexDLY turns each play into a delegated task with a checklist, keeps the human steps human, and totals the hours given back. Start with a free War Room and the Learn Center walks you through week one.
Build my free War Room