A survey of 18,000 Danish workers, published by Anders Humlum and Emilie Vestergaard in the Proceedings of the National Academy of Sciences, asked workers who believed ChatGPT could cut a task's time in half whether they planned to use it. Fewer than 1 in 4 of them planned to use it in the next two weeks.
Behavioral barriers to enterprise AI adoption are the human and organizational obstacles that stop employees from using AI tools consistently, even when they believe the tools would save them time. In our AI adoption work at The Decision Lab, the five most common are time constraints, poor workflow integration, unclear usage guidelines, low trust in AI accuracy, and perceived complexity.
Adoption improves when an organization identifies which barrier is holding back each group of employees and removes that barrier specifically. Awareness campaigns on their own did not move usage in either the Danish study or our Fortune 500 work. We identified the five barriers in a 20,000-person Fortune 500 workforce, and we diagnose them with our SPROUT framework, validated with data from more than 20,000 employees.
The five behavioral barriers at a glance
- Time constraints. Learning the tool costs time now, while the saving arrives later, if it arrives at all.
- Workflow integration. AI sits outside the tools and decisions people already use.
- Unclear usage guidelines. Employees do not know what is permitted or expected.
- Low trust in AI accuracy. Checking every output cancels the time the tool was supposed to save.
- Perceived complexity. Confidence and training needs vary widely across roles and employee groups.
Why believing AI saves time does not drive adoption
The intention-behavior gap is the distance between what people believe would help them and what they actually do. In the Danish study, the researchers randomly told some workers how much time ChatGPT could save them, and usage did not shift. Our Fortune 500 client had already run training and internal communications before we started, and employees knew about the tools and still did not use them.
The five barriers in detail
1. Time constraints: the cost of switching arrives before the payoff
Present bias is the tendency to weigh an immediate cost more heavily than a larger benefit that arrives later. At our client, employees were weighing the effort of changing a workflow against an uncertain payoff, and time constraints came up as a barrier in their own right. The programmes we piloted embedded hands-on AI tasks in work employees already had to do.
2. Workflow integration: tools that sit outside the work get abandoned
MIT's Project NANDA reported in 2025 that 19 in 20 of the enterprise generative AI pilots it reviewed produced no measurable impact on profit and loss. The report is preliminary, and its headline figure has been criticised. Its diagnosis still matches what we see: pilots stall when a tool does not fit how people already work or does not learn from their feedback.
We have seen the same pattern outside generative AI. When one of the world's leading vehicle manufacturers asked us to build a data dashboard for its sales team, the design problem was fitting the data into decisions the team already made. The client has since scaled the tool across its operations.
3. Unclear usage guidelines: people hold back when they do not know the rules
The 2025 Trust, Attitudes and Use of Artificial Intelligence study was led by Nicole Gillespie and Steve Lockey of Melbourne Business School at the University of Melbourne, in collaboration with KPMG. It surveyed more than 48,000 people across 47 countries. Only 4 in 10 employees said their workplace had a policy or guidance on generative AI, and fewer than half said they had received any AI training.
Danish workers named employer restrictions and a perceived need for training as their main obstacles. Unclear guidelines were one of the five barriers at our client too. They are among the cheapest barriers to fix, since writing down what is allowed costs far less than buying a new tool.
4. Low trust in AI accuracy: use has outrun trust
The Melbourne and KPMG study found that 2 in 3 people use AI with some regularity, while fewer than half are willing to trust it. Employees who use a tool they doubt end up checking every output and losing the time saving, or accepting output they are unsure of. At our client, a skeptical segment of employees distrusted AI's value and stalled at first contact.
5. Perceived complexity: the barrier falls unevenly across a workforce
Humlum and Vestergaard found that women were 16 percentage points less likely than men to have used ChatGPT for work. Workers in less IT-focused occupations, such as teachers, more often said they needed training before they could use it.
Complexity is partly a matter of confidence, and in our pilot, confidence moved quickly in both directions. The pilot section below gives the figures.
Why barriers differ by employee segment
Our client's workforce split into two groups with opposite failure modes. Enthusiasts were already experimenting but kept running into friction, while skeptics stalled at first contact. An organisation-wide average would blend the two into a middling score that describes neither group, so every programme we designed targeted a specific barrier for a specific segment.
Seniority matters too. Employees can only judge where AI fits their job once leadership has decided what it is for. With BDC, Canada's bank for entrepreneurs, we worked with the Senior Management Committee and AI Council to specify ten AI capabilities in a phased roadmap. Each capability keeps human judgment from account managers and community partners as the point of trust.
How the SPROUT framework diagnoses adoption barriers
SPROUT is The Decision Lab's framework for measuring the conditions that sustainable AI adoption depends on. We developed it and validated it with data from more than 20,000 employees. It scores five pillars:
- Social Readiness. Measures how peer norms and the wider team environment shape willingness to use AI. Closest barrier: none of the five directly. It captures whether visible AI use feels socially safe.
- Psychological Readiness. Measures how individuals feel about AI, including openness and fear of replacement. Closest barrier: low trust in AI accuracy.
- Role Fit. Measures how relevant and useful AI feels for day-to-day work. Closest barrier: workflow integration and time constraints.
- Organizational Enablement. Measures whether policies, leadership, incentives, and culture support AI use. Closest barrier: unclear usage guidelines.
- Understanding and Technical Capability. Measures confidence and competence in using AI, from basic literacy to effective prompting. Closest barrier: perceived complexity.
The mapping is approximate. SPROUT measures conditions, and a single barrier can show up across more than one pillar.
Team members answer 17 anonymous questions in our free AI Adoption Diagnostic, and results appear once at least five people have responded. The output includes pillar scores and a ranked list of barriers, each paired with a recommended next step.
As an illustration, suppose a team scores well on Role Fit but poorly on Organizational Enablement. That team already sees where AI fits its work, and rules or incentives are what hold it back, so publishing clear usage guidance should come before any new training.
Results from our Fortune 500 pilot
We designed 20 candidate programmes and tested eight in a one-month controlled pilot with more than 100 employees. The programmes ranged from hands-on tasks inside existing workflows to a mission-based challenge series run in Slack, and we compared them against a control group on before-and-after measures of ease and confidence.
The published case study reports a 35% rise in organisation-wide AI adoption and a 23% rise in use of AI in employees' core responsibilities. Confidence in exploring AI tools rose 41% in the pilot group and fell 26% in the control group over the same month. The programmes are now scaling to the company's full 20,000-person workforce.
These are results from one proprietary pilot. The case study does not publish:
- Starting levels for adoption or core-responsibility use
- How core-responsibility use was measured
- A breakdown by programme or segment
- Whether the gains lasted beyond the pilot month
Read the figures as what moved in one organisation rather than as a benchmark.
Four actions enterprise leaders can take
- Give employees protected time to rebuild one real workflow with AI, and compare usage against a team without that time.
- Publish a one-page statement of permitted tools and uses. After 90 days, ask anonymously whether people know the rules.
- Run the diagnostic and split the results by segment before reading any organisation-wide average.
- Design first contact separately for skeptics, and track 30-day use against a single-track control group.
Humlum and Vestergaard also found that Danish workers using ChatGPT were already earning slightly more before it arrived, even though the tool could help workers with less expertise the most. If behavioral barriers fall hardest on employees with lower confidence or less technical exposure, enterprise rollouts may be widening existing performance gaps. Organizations can test this by comparing adoption and task outcomes across employee segments before and after a rollout.
Frequently asked questions
What are behavioral barriers to enterprise AI adoption?
They are the human and organizational obstacles that keep employees from using AI tools consistently: time constraints, workflow integration, unclear usage guidelines, low trust in AI accuracy, and perceived complexity. They are separate from technical barriers such as data or infrastructure gaps, and they persist after the technology itself works.
Why do employees resist using AI tools at work?
Beyond the five barriers, visible AI use can cost people standing with colleagues. A 2025 study in the Proceedings of the National Academy of Sciences by Jessica Reif, Richard Larrick, and Jack Soll found that people who used AI at work were rated lower on competence and motivation. Our guide on why employees resist AI tools covers this in detail.
What is the intention-behavior gap in AI adoption?
It is the difference between believing AI would help and actually using it. In AI adoption, the gap means employee surveys showing enthusiasm for AI can overstate how much the tools will be used. Telling workers about time savings did not change their behavior in the Danish study.
How can organizations increase employee adoption of generative AI?
Diagnose which barriers affect which employees, then design interventions for each segment instead of one programme for everyone. Embed AI in tasks people already do, publish clear usage rules, and test each change against a comparison group. In our Fortune 500 pilot, this approach raised organisation-wide adoption within a month.
Why do AI pilots fail to produce business impact?
MIT's Project NANDA found that most enterprise generative AI pilots stall when tools fail to fit existing workflows or learn from user feedback. It also found that solutions bought from specialized vendors, and genuine partnerships, tended to outperform ambitious internal builds. The report is preliminary, so treat its exact figures with care.
How should leaders address low trust in AI accuracy?
Let employees edit AI output instead of accepting or rejecting it whole. Berkeley Dietvorst, Joseph Simmons, and Cade Massey found in Management Science in 2018 that people used an imperfect algorithm far more when they could adjust its forecasts, even slightly. Showing the tool's limits early also prevents a later loss of trust.
What AI usage guidelines should employers provide?
At minimum, a short document naming the approved tools, the permitted uses, the data that must stay out of AI tools, and how AI-assisted work will be evaluated. Keep it to about a page, and check after 90 days whether employees can say what the rules are.
How can companies measure barriers to AI adoption?
Survey employees anonymously, since many will not admit unapproved AI use under their own name. Our free AI Adoption Diagnostic uses 17 anonymous questions to score the five SPROUT pillars and rank a team's barriers. Comparing those results with logged usage shows how much AI use is going unreported.
