Last updated October 8, 2026
In May 2026, 47% of US employees told Gallup that their organization had integrated AI tools. In the same survey, 3 in 10 employees used AI at work a few times a week or more, and about 1 in 7 used it daily, according to Gallup's workplace AI indicator.
Successful AI adoption means employees use AI repeatedly, in the work it was bought for, with enough judgment to catch its mistakes. Behavioral science gets organizations there by treating adoption as specific behaviors to define, diagnose, change and test.
At The Decision Lab, this is the method behind our AI adoption work. In our AI adoption program for a 20,000-person Fortune 500 workforce, confidence in exploring AI tools rose 41% among pilot participants over one month and fell 26% in a control group over the same period. We diagnose adoption barriers with our SPROUT framework, validated with data from more than 20,000 employees. This guide shows how to diagnose those barriers, choose interventions for them, test the interventions and measure durable behavior change.
The four stages of AI adoption
AI adoption is the point at which using AI becomes part of how a person does specific tasks, without reminders or a manager asking. We describe the path there as the Four Stages of AI Adoption: access, activation, usage and sustained behavior change. Organizations often report the earlier stages as adoption, but each stage needs a different intervention.
| Stage | What it looks like | What usually moves it |
|---|---|---|
| Access | Employees have a licence, a login or a tool built into software they already use | Procurement, IT rollout, policy |
| Activation | An employee completes a first useful task with AI | Clear permission, a concrete first task, low setup effort |
| Usage | An employee returns to AI often, but for scattered tasks | Prompts at the moment of need, feedback on results |
| Sustained behavior change | AI is the normal way certain tasks get done, and use holds when attention moves elsewhere | Workflow redesign, manager expectations, peer norms, repetition |
Most rollout metrics capture the first three stages, while the value an organization paid for arrives at the fourth.
The gap between the second and fourth rows takes longer to close than most rollout plans allow. In a study of 96 volunteers by Phillippa Lally and colleagues, published in the European Journal of Social Psychology, a new daily behavior took a median of 66 days to become close to automatic among participants whose progress could be modelled, with individual estimates from 18 to 254 days. The behaviors were simple health routines, so we treat the figure as a lower bound for changing how a professional task gets done.
Gallup's data show the same drop-off between stages. Between May 2025 and May 2026, the share of US employees who used AI at work at least a few times a year rose from 40% to 52%, while the share using it daily rose from 8% to 15%. Gallup also found that frequent users apply AI across more types of tasks, and that employees who use it in more ways report larger productivity gains. When we scope an adoption program, we define success at the fourth stage and work back.
Why AI programs fail to reach everyday use
AI programs usually stall between activation and habit, because using AI carries costs for the individual employee at the moment of use. Across our AI adoption work, including the Fortune 500 diagnostic described in the case studies below, six causes recur. For each one, we set out what the research shows and then what we recommend doing about it.
1. Trust that breaks after one visible error
People hold AI to a stricter standard than they hold themselves or colleagues. In experiments by Berkeley Dietvorst, Joseph Simmons and Cade Massey, published in the Journal of Experimental Psychology: General, participants who watched a forecasting algorithm make mistakes became less likely to rely on it than on a human forecaster, even when the algorithm made smaller errors overall. The experiments used forecasting tasks, so for workplace AI we treat the finding as a practical risk: one visible error in an early draft may be enough for an employee to stop using AI for that task.
What to do: start with tasks where errors are easy to spot, and give employees a quick way to correct and report them.
2. Perceived risk with no clear rules
When employees cannot tell what is allowed, the safest choice is to avoid AI or to use it out of sight. In Gallup's May 2026 survey, 1 in 4 US employees said their organization had communicated a clear plan for integrating AI.
What to do: publish a short guideline for each role that lists the tasks where AI is encouraged and the uses that are off limits.
3. Low confidence in one's own skill
Perceived complexity stops people before they have enough experience to judge whether AI helps them. In a survey of 18,000 Danish workers published by Anders Humlum and Emilie Vestergaard in PNAS, workers who saw productivity potential in ChatGPT were often held back by employer restrictions and a perceived need for training. In our Fortune 500 diagnostic, a skeptic segment stalled at first contact while an enthusiast segment kept experimenting but hit friction, so the two groups needed different programs.
What to do: give each employee group a first task drawn from its own work, small enough to finish in one sitting.
4. Workflow friction
AI that sits outside the tools and steps people already use adds effort before it saves any. In a six-month randomized trial of Microsoft 365 Copilot with about 6,000 workers at 56 firms, Eleanor Wiske Dillon, Sonia Jaffe, Nicole Immorlica and Christopher Stanton found that workers changed tasks they controlled alone, cutting time on email, but did not change meetings or other work that required colleagues to change too. Weekly use peaked at 6 in 10 workers when licences arrived and settled at about 4 in 10.
What to do: put AI into the tool and step where the target task already happens, starting with tasks one person controls.
5. Unclear or negative incentives
Using AI openly can cost employees status. Across four preregistered experiments with 4,439 participants, Jessica Reif, Richard Larrick and Jack Soll of Duke University's Fuqua School of Business documented a social evaluation penalty for using AI: people who received help from AI were rated as lazier and less competent than people who received the same help from other tools. The experiments measured judgments of described employees, so the workplace implication is ours: if performance reviews reward visible effort, employees have a reason to keep AI use quiet or skip it.
What to do: review AI-assisted work on its quality, and have managers use AI where their teams can see it.
6. Social norms set by the team
Whether colleagues and managers use AI shapes what feels normal. In the same Copilot trial, the firm a worker belonged to predicted usage far better than their industry or prior work habits, with average weekly use ranging from about 1 in 5 workers at one firm to 3 in 4 at another. The authors point to firm practices such as training and management as the likely cause.
What to do: compare adoption across teams, then use teams where AI use is already common as examples for similar roles.
A behavioral framework for AI adoption
The framework we use has five steps: define the target behavior, diagnose the barriers for each group, choose interventions matched to those barriers, test them against a comparison group, and iterate before scaling. It follows the same logic we have applied to technology adoption outside AI, from clean cookstoves for the World Bank to digital health platforms, adapted for the specific barriers AI creates.
- Define the target behavior. Name who should use AI, for which task, at what point in their workflow and how often. "Use AI more" cannot be measured or designed for. "Account managers draft first-pass client summaries with AI after every quarterly review" can be.
- Diagnose barriers by employee group. Map how each group moves from first exposure to routine use and where they drop off. For the Fortune 500 client, we combined a review of 48 sources on adoption with workforce interviews, surveys and expert consultation. We now measure the conditions behind adoption with our SPROUT framework, which covers social readiness, psychological readiness, role fit, organizational enablement, and understanding and technical capability.
- Match interventions to barriers. Each intervention should target one barrier for one group. For the same client, we designed 20 candidate programs and used structured, evidence-based criteria to select eight for piloting.
- Test against a comparison group. Run the shortlisted interventions with a small group and compare them with a similar group that did not receive them. We piloted all eight programs for one month with more than 100 employees against a control group, measuring confidence and ease of use before and after. A randomized controlled trial is the strongest test, but it is not always practical, and a well-matched comparison group or a staggered rollout can still separate an intervention's effect from general enthusiasm for AI.
- Iterate and hand over. Keep what moved the target behavior, drop what did not, and give internal teams the protocols to run the programs themselves. Behavior change continues long after a pilot ends, so measurement should continue too.
The order matters more than any single step. Much of our AI adoption work starts with organizations that began at step 3: they trained employees and ran internal communications, then found that usage stayed low because no one had defined the target behavior or checked what was blocking it.
Interventions that increase AI adoption
The interventions that increase AI adoption change the moment of use: what is ready by default, what prompts the behavior, who is seen doing it and what happens after a mistake. Each one below targets a different barrier, and we choose among them based on the diagnosis rather than running all seven at once.
Defaults
People tend to stay with whatever option is already set. In a study by Brigitte Madrian and Dennis Shea of a large US company's retirement plan, published in the Quarterly Journal of Economics, 86% of employees hired under automatic enrollment participated, compared with 37% of similar employees hired before the change. For AI, a default can mean an AI-drafted first version that appears inside the tool an employee already opens, which they can edit or discard. The employee still chooses whether to use it, and skipping it should take one click.
Timely prompts
A prompt works best at the moment the target task begins, rather than in a newsletter sent days earlier. Katherine Milkman, Angela Duckworth and colleagues tested 54 four-week programs with 61,293 gym members in a megastudy published in Nature. Just under half of the programs raised weekly gym visits during the program, by 9% to 27% over a control group, and about 1 in 12 produced a significant effect that lasted after the four weeks ended. We read that result as a planning rule: prompts help people start, and something else has to keep the behavior going once they stop.
Peer modeling
Employees learn what is acceptable by watching colleagues in similar roles. In the Duke study on the social evaluation penalty, the penalty was smaller among evaluators who used AI often themselves. We design for this by making AI use visible within teams, so colleagues in similar roles can see each other's results on real work.
Manager behaviors
Managers decide whether AI use is treated as legitimate work. In Gallup's May 2026 data, employees in AI-adopting organizations who strongly agreed that their manager actively supported their team's AI use were 1.7 times as likely to use AI a few times a week or more as other employees. The data are correlational, so we pair them with specific manager actions we can test: naming the tasks where AI use is expected, using it visibly, setting aside team time to share what worked, and reviewing AI-assisted work on its quality rather than on how much effort it appeared to take.
Feedback loops
People continue a behavior when they can see what it produced. Useful feedback shows an employee the time saved or the quality of the output on their own tasks, and it gives them a quick route to report errors so the guidelines improve. We measure confidence and ease of use before and after each program we pilot, and use the results to revise the next version.
Training design
Training that lets people make and fix errors transfers better than training that shows them the correct steps. In a meta-analysis of 24 studies with 2,183 participants, Nina Keith and Michael Frese found that error management training outperformed error-avoidant training, with the largest advantage on tasks that differed from the ones practiced. For AI, that means practice on real work, including outputs with errors that trainees learn to catch.
Human oversight
People use imperfect tools more when they keep some control over the result. In follow-up experiments published in Management Science, Dietvorst, Simmons and Massey found that participants were considerably more willing to use an algorithm's forecasts when they could adjust them, even by a small amount. In practice, employees should be able to correct AI output and flag its errors. In our AI strategy work with BDC, Canada's bank for entrepreneurs, we specified ten AI capabilities that each keep account managers and other people as the point of trust for the entrepreneurs BDC serves.
The AI Adoption Barrier-to-Intervention Matrix
The AI Adoption Barrier-to-Intervention Matrix links each of the six barriers to the intervention we usually start with and the measure that shows whether it worked. We treat it as a starting point and adjust it once the diagnosis shows which barriers matter for each employee group.
| Barrier | What it looks like | Intervention | Example action | What to measure |
|---|---|---|---|---|
| Low trust after errors | Employees stop using AI after a visible mistake | Human oversight and feedback loops | Let employees correct outputs and flag errors in one step | Repeat use after an error |
| Unclear rules | Employees avoid AI or use it privately | Role-specific guidance | Publish encouraged and off-limits uses for each role | Confidence in appropriate use |
| Low confidence | Employees cannot name a useful first task | Workflow-based training | Practice on real tasks from the employee's own role | First-task completion |
| Workflow friction | AI requires extra steps or switching tools | Defaults and embedded tools | Put AI drafting inside the existing workflow | Use at the target workflow step |
| Negative incentives | Employees fear looking lazy or careless | Manager modeling and changes to performance review | Judge AI-assisted work on its quality | Willingness to use AI openly |
| Weak norms | AI use feels unusual or risky | Peer modeling | Run team challenges with role-relevant examples | Variation in adoption between teams |
Responsible AI adoption
Responsible AI adoption means employees understand why they are being asked to use AI and keep the ability to correct or decline it, without being monitored in ways they would object to if they knew. Behavioral interventions are effective because they change the setting people decide in, so we hold them to a stricter standard than an information campaign.
Transparency
Employees should know which tasks the organization wants AI used for, and what the tools and the organization can see about their use. When people do not have that information, many fill the gap with the worst plausible reading, and avoidance or hidden use follows. A one-page usage guideline written for each role answers most of these questions.
Autonomy
Every default and prompt we design keeps an easy way out. An AI-drafted first version that an employee can discard in one click respects their judgment. A workflow that cannot be completed without AI does not, unless the organization has decided openly that the task now requires it and has said so.
Fairness
Adoption gaps tend to follow existing inequalities unless a program is designed against them. In the Danish survey by Humlum and Vestergaard described above, women were 16 percentage points less likely than men to have used ChatGPT for work, and users had earned slightly more than non-users before ChatGPT arrived. A program that reaches mostly the employees who were already experimenting will widen those gaps, so we set adoption targets for each group, not only for the organization as a whole.
Privacy
Usage data should be collected at the level needed to improve the program, usually the team or role, and kept separate from individual performance records. Employees who believe every prompt is being read by their manager will either avoid AI or move to tools the organization cannot see.
Avoiding coercive implementation
Individual usage quotas and AI use counted in performance reviews can raise logins without raising useful work. Once usage becomes the target, employees have a reason to produce usage, and measured adoption stops telling the organization whether AI is helping. We aim interventions at making AI easier and more worthwhile to use, and we check whether the quality of the work improved along with the usage numbers.
How to measure AI adoption success
AI adoption success is measured by whether target behaviors persist and improve work, so we track six measures for each employee group, against a baseline and a comparison group where possible. Login counts alone cannot show any of the six.
| Signal | Measure | What it shows | How we usually measure it |
|---|---|---|---|
| Early signal | Activation | Whether employees complete a first useful task with AI | Share of each group completing a defined first task within a set period |
| Early signal | Employee confidence | Whether people feel able to use AI well | Short pre-post survey items on confidence and ease of use |
| Adoption signal | Repeat use | Whether use continues after the launch period | Share of each group using AI in at least a set number of weeks per month, tracked after prompts and campaigns end |
| Adoption signal | Workflow integration | Whether AI is used in the target task, at the target step | Usage logs or work samples tied to the specific workflow, not overall tool activity |
| Business and risk signal | Quality of decisions and work | Whether AI-assisted work is better, worse or the same | Blind review of work samples, error rates, rework, or outcome data where the task has one |
| Business and risk signal | Equity outcomes | Whether adoption and its benefits reach every group | All of the above, broken down by role, seniority, location, gender and other groups relevant to the organization |
Two timing rules matter. First, measure after the intervention stops, since the megastudy described above found that most effects ended with the program. Second, allow enough time for a habit to form, since the median in Lally's habit study was 66 days for simple daily behaviors.
Every measure needs a comparison group where one is feasible, because general enthusiasm for AI and other initiatives running at the same time move the same numbers. The Fortune 500 case study below shows how much a comparison group can change the reading of a pilot.
Case studies and practical applications
The same behavioral method applies across sectors, but the target behavior and the main barrier change with the setting. Below are examples from our own AI work and one large public-sector trial, followed by a non-AI project that shows where our diagnostic method comes from.
Workplace technology: a 20,000-person Fortune 500 workforce
A Fortune 500 HR technology company had deployed AI tools across its 20,000-person workforce, with training and internal communications behind them, and usage stayed low. We built an end-to-end map of how employees moved from first exposure to everyday use. It isolated five barriers: time constraints, poor workflow integration, unclear usage guidelines, low trust in AI accuracy and perceived complexity.
We then designed 20 candidate programs, each aimed at one barrier for one employee segment, and piloted eight of them for one month with more than 100 employees against a control group. The programs ranged from hands-on tasks embedded in real workflows to a mission-based challenge series run in Slack. The confidence results in the introduction come from this pilot. Because confidence fell in the control group over the same month, a count of participants alone would have understated what the programs did.
Each program came with the protocols and guides the client's teams need to run it, and the programs are now scaling to the full workforce. Read the Fortune 500 AI adoption case study.
Finance: defining a bank's next decade of AI
BDC, Canada's bank for entrepreneurs, asked what its service should become when its data and AI work as one system. Working with BDC's Senior Management Committee and AI Council, we defined a connected experience for entrepreneurs and specified ten AI capabilities in a phased roadmap. Every capability keeps human judgment, from account managers to community partners, as the point of trust. The blueprint is now being implemented. Read the BDC AI strategy case study.
Health: getting people to act on AI recommendations
A digital wellness platform serving 25 million members was running its recommendations on one behavior model applied to every member. A recommendation can be clinically correct and still be ignored, so we built a behavioral engine that models each member individually, scores each candidate action by how hard it will be for that person, and plugs into the platform's recommender. We tested the full engine on synthetic members before launch, and it now runs in production. Read the digital wellness behavioral engine case study.
Public sector: from technology access to use
The UK Government Digital Service gave Microsoft 365 Copilot licences to 20,000 government employees across 12 organizations from September to December 2024. In a survey answered by 7,115 users near the end of the trial, participants reported saving an average of 26 minutes a day, according to the cross-government findings report. The time savings were self-reported and the trial had no control group, so the figure shows perceived value more reliably than measured productivity.
Method evidence from outside AI: clean cookstoves in Uganda
For the World Bank in Uganda, our field research identified 26 specific barriers to clean cookstove adoption, including the weight of the cheapest stove, which most women could not carry on their own. While this was not an AI program, it illustrates why effective adoption work begins with diagnosing concrete, context-specific barriers rather than assuming a generic intervention will work.
Related resources from The Decision Lab
This guide sits at the center of our AI adoption material, and each resource below goes deeper on one step of the framework.
- AI adoption readiness assessment: how we use the SPROUT framework to measure whether an organization's people are ready to use AI, before or during a rollout.
- Behavioral barriers to enterprise AI adoption: the five barriers from our Fortune 500 diagnostic, with what removes each one.
- Why employees resist AI tools: the personal costs behind avoidance and hidden AI use, including the reputational penalty for visible use.
- AI adoption training: what makes training change behavior rather than only intentions, and where training needs support from managers and policy.
- How to increase AI adoption across your workforce: a seven-step sequence for leaders planning a program.
- Our case studies, including AI adoption, AI strategy and technology adoption work across the private and public sectors.
If you are planning or rescuing an AI rollout, our AI adoption team runs the diagnosis and the pilots, then hands over programs your teams can run themselves.
Frequently asked questions
What is the difference between AI access and AI adoption?
Access is a procurement milestone: employees can open the tool. Adoption is a behavior, defined for a specific role and task. A practical test is to imagine every reminder and incentive stopping tomorrow. The employees who would still use AI for the same tasks next quarter are the ones who have adopted it.
What is the SPROUT framework?
SPROUT is The Decision Lab's framework for measuring the behavioral conditions behind AI adoption. It covers social readiness, psychological readiness, role fit, organizational enablement, and understanding and technical capability. We use it at the diagnosis step to find which conditions are weakest for each employee group, then choose interventions aimed at those conditions.
How long does AI adoption take?
Plan in months for any single task. Habit research on simple daily behaviors found a median of about two months to near-automatic performance, and professional workflows are more complex. We recommend testing interventions for at least a month and tracking repeat use for at least a quarter after extra support ends.
What is the most effective intervention for AI adoption?
No single intervention works everywhere, because the main barrier differs between employee groups, even within one organization. Prediction is also unreliable: in the Nature megastudy, impartial judges failed to forecast which programs would work best. Testing a shortlist against a comparison group is more dependable than choosing one favorite.
How should organizations set AI adoption targets?
Set targets for a defined behavior in each employee group, rather than for licences or logins. A useful target names the task and how often AI should be used for it, such as weekly use for first-draft client summaries. Separate targets for each group stop a strong average from hiding groups that are being left behind.
What should leaders do when employees do not trust AI?
Start by making AI errors easy to spot and correct. People rely more on tools they can check and adjust, so leaders should keep a person responsible for every AI-assisted output and give employees a quick way to flag mistakes and see them fixed. Lower-stakes tasks are the best place to begin, because trust can then build on evidence.
What role do managers play in AI adoption?
Managers are influential and often underused. In Gallup's May 2026 data, just over a third of employees in AI-integrating organizations strongly agreed that their manager supported their team's AI use. We recommend giving managers specific actions, such as naming the tasks where AI is expected and judging AI-assisted work on quality, rather than asking for general enthusiasm.
Is it ethical to use nudges to increase AI adoption?
It can be, when the nudge makes a choice easier without hiding or removing it. A default that employees can discard in one click, backed by clear information about what the AI does, preserves choice. Tying usage to performance ratings or monitoring individual prompts adds pressure, and it also distorts the usage data the organization relies on.
