Why AI adoption stalls: the common mistakes and what the evidence says to do instead

A team runs an impressive AI demo. Leadership approves a pilot. Six months later the pilot is still a pilot, nobody can say whether it saved money, a few employees quietly use their own chatbot accounts instead, and the project is shelved. If that sounds familiar, you are in large company. Surveys of executives, interviews with engineers and controlled experiments all point to the same uncomfortable pattern: AI adoption usually goes wrong for organizational reasons, not technical ones.

This article sorts the most common adoption mistakes into four groups, shows the evidence behind each, and is honest about what that evidence can and can't tell you. By the end you should be able to look at an AI initiative, your own or someone else's, and spot the warning signs early.

What "failure" means, and why the headline numbers disagree

You will often hear that most AI projects fail. The numbers behind that claim are real, but they measure different things, so it helps to read them carefully.

• "More than 80 percent." A 2024 RAND report states that by some estimates more than 80% of AI projects fail, about twice the rate of IT projects that don't involve AI¹. RAND is citing an earlier estimate here, not measuring the rate itself. Its own contribution is qualitative: interviews with 65 experienced data scientists and engineers, conducted between August and December 2023.
• "95 percent." MIT's Project NANDA reported in mid-2025 that just 5% of integrated generative AI pilots were "extracting millions in value," with most showing no measurable effect on profit and loss². Its evidence is a review of more than 300 public AI initiatives, 52 interviews and 153 survey responses from senior leaders. That is a small, non-random sample, and "no measurable P&L impact" is a high bar that early projects rarely clear.
• "42 percent." In a survey of more than 1,000 respondents in North America and Europe, S&P Global Market Intelligence found that the share of companies abandoning most of their AI initiatives jumped to 42% in 2025, from 17% the year before. The average organization scrapped 46% of its proofs of concept before they reached production³.
• "Over 40 percent" of agent projects. Gartner, the research and advisory firm, predicted in June 2025 that more than 40% of "agentic AI" projects (systems that take multi-step actions on their own) will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls⁴. That is a forecast, not a measurement.

None of these figures is a precise failure rate for "AI." Some count abandoned pilots, which can be a sign of healthy experimentation rather than waste. Some rely on executives' self-reports. Taken together, though, they agree on direction: a large share of AI efforts don't make it from demo to durable value. The more useful question is why.

Mistakes before anything is built

RAND's interviews produced five root causes of AI project failure, and three of them happen before a model is ever deployed¹:

1. Solving the wrong problem. Stakeholders "misunderstand" or "miscommunicate" "what problem needs to be solved using AI." A team builds a forecasting model when the business actually needed faster approvals, or optimizes a metric nobody upstairs cares about.
2. Chasing the technology. The organization "focuses more on using the latest and greatest technology than on solving real problems for its intended users." This is the "we need an AI strategy" project that starts from a tool and goes looking for a use.
3. Picking problems too hard for AI. Some tasks are simply beyond what current systems can do reliably, and no amount of effort fixes that.

The second mistake is especially easy to make now, because vendors label almost everything as AI. Gartner calls the agent version of this "agent washing": rebranding existing assistants, robotic process automation (rule-based software that mimics clicks and keystrokes) and chatbots as autonomous agents. It estimates that only about 130 of the thousands of vendors selling "agentic AI" are offering the real thing, and that many use cases sold as agentic "don't require agentic implementations"⁴.

The other two RAND root causes are about foundations: the organization lacks the data needed to train or ground an effective system, or lacks the infrastructure to manage that data and run models in production¹. Data quality came up again and again: 30 of the 50 industry interviewees described persistent data-quality problems. Generative AI does not make this go away. MIT NANDA's interviewees described tools that "forget context," don't learn from feedback and don't improve over time, which made them hard to trust for anything beyond simple tasks².

It is also fine to conclude that AI isn't the answer. In Statistics Canada's survey for the second quarter of 2026, 40.0% of Canadian businesses said AI was not relevant to their business¹². For many of them, that is probably a sound judgment rather than a mistake.

Mistakes in the rollout: bolting AI on

The most consistent finding in the business data is that value comes from changing the work, not from switching on a tool.

Not redesigning workflows. McKinsey's 2026 global survey of 1,719 participants found that only 6% qualify as "AI high performers," meaning they attribute 5% or more of their earnings before interest and taxes (EBIT) to AI and report significant value from it⁵. Nearly three-quarters of those high performers say they have fundamentally redesigned workflows because of AI, compared with about a quarter of everyone else. This is survey data and correlation, not proof that redesign causes the gains, but the gap is large and consistent with the other evidence here.

Not encouraging or training people. A study of about 25,000 Danish workers in 11 occupations exposed to chatbots, by economists Anders Humlum and Emilie Vestergaard, found that employer encouragement nearly doubled take-up for the typical worker, from 47% to 83%, and that benefits were 10% to 40% larger when employers encouraged use⁷. Training is uneven in Canada: among businesses that used AI, 32.0% trained existing employees on it, and that share fell to 24.0% for businesses with one to four employees¹².

Ignoring "shadow AI." When organizations don't provide approved tools, employees bring their own. MIT NANDA found workers at more than 90% of companies regularly using personal AI tools for work, while only 40% of companies said they had bought an official subscription². That is a data-security risk, and also a missed signal: the informal use shows where people already find AI useful.

Rewarding output volume. Pushing staff to "use AI more" without guidance can create what researchers at Stanford and BetterUp call "workslop": AI-generated work that "looks good, but lacks substance," shifting the real thinking onto whoever receives it¹⁰. In their online survey of 1,150 full-time US desk workers in August and September 2025, 40% said they had received workslop in the previous month, and each incident took an average of 1 hour 56 minutes to deal with. The researchers estimate this costs about $9 million a year in lost productivity for a 10,000-person company. Treat that dollar figure as a rough, self-reported estimate.

Mistakes in measurement: trusting the feeling

If you can't tell whether AI is helping, you can't fix it or defend it. The evidence shows that our impressions are a poor guide.

Perceived speed is not real speed. In early 2025, METR, a nonprofit that evaluates AI systems, ran a randomized controlled trial with 16 experienced open-source developers working on 246 real tasks from their own projects⁸. Each task was randomly assigned to allow or disallow AI tools. With AI, the developers took 19% longer. Yet they had expected AI to speed them up by 24%, and even afterwards believed it had sped them up by 20%.

That result needs two caveats. It covered a narrow, demanding setting, and METR now labels it out of date. A follow-up with 57 developers, published in February 2026, estimated a speedup instead, but with wide uncertainty: about 4% faster for newly recruited developers, with a confidence interval running from 15% faster to 9% slower⁹. METR describes the new data as an unreliable signal, partly because many developers now refuse to work without AI at all. The lasting lesson is not the size of the effect but the gap between what people felt and what the clock showed.

Averages hide who benefits. In one of the best-known field studies, Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,172 customer-support agents and found that an AI assistant raised issues resolved per hour by 14% on average, by 34% for novice and low-skilled workers, and had minimal impact on the most experienced and skilled⁶. A rollout judged only on the average misses where value actually lands, and which staff need something different.

Small gains at the level of the whole job. The Danish study found users reported average time savings of just 2.8% of work hours, and no measurable effect on earnings or recorded hours, with confidence intervals ruling out effects larger than 1%⁷. The data run only to 2024, and tools have improved since. But it is a useful counterweight to vendor claims of hours saved per day: task-level speedups don't automatically add up to job-level change.

Mistakes in accountability: "the AI said it"

A quieter mistake is treating AI output as someone else's responsibility.

In November 2022, after their grandmother died, a traveller named Moffatt asked Air Canada's website chatbot about bereavement fares. The chatbot said a reduced bereavement rate could be claimed retroactively, within 90 days of the ticket being issued. The airline's own policy page said the policy didn't apply once travel was complete. When Mr. Moffatt applied, Air Canada refused, and argued before British Columbia's Civil Resolution Tribunal that it could not be held liable for information from its agents or representatives, including the chatbot.

The tribunal called this "a remarkable submission" and noted that, in effect, Air Canada was suggesting the chatbot was "a separate legal entity that is responsible for its own actions." It ruled that "it should be obvious to Air Canada that it is responsible for all the information on its website," and that "it makes no difference whether the information comes from a static page or a chatbot." It found negligent misrepresentation and awarded $812.02 in damages, interest and fees¹¹.

The sum was small, but the principle travels: in the eyes of a Canadian tribunal, a customer-facing AI speaks for the company that deploys it. Treat it as you would any other employee-facing or customer-facing channel, with review, testing and an owner.

What the evidence can't tell us yet

A few honest limits are worth keeping in mind.

• Much of the evidence is self-reported. McKinsey, S&P Global, MIT NANDA and the workslop study all rely on what respondents say. Executives may overstate or understate value, and people who answer AI surveys may not represent everyone.
• The tools are moving fast. METR's own reversal shows how quickly a 2025 finding can age. Studies of 2023 and 2024 tools may understate what 2026 tools can do.
• Correlation isn't causation. High performers redesign workflows, but they may also have better data, more money and stronger leaders. No randomized trial has yet shown, at company scale, that redesign alone delivers the gains.
• "Failure" is sometimes the right outcome. Killing a pilot that doesn't work is cheaper than scaling it. A high abandonment rate can mean organizations are learning, not just wasting money.

What this means for you

Whether you lead a team, run a small business or are simply asked to "start using AI," these questions will catch most of the mistakes above:

1. Can you state the problem without the word "AI"? If the goal is "use agents" rather than "cut invoice processing from five days to one," start again.
2. Do you have the data, and is it good enough? Check before you buy. Ask a vendor to run on a sample of your real, messy records.
3. What will change in the workflow? Write down which steps disappear, who reviews the output and what people do with the time saved.
4. How will you measure it, and against what? Record a baseline before rollout. Time real tasks rather than asking people how fast they feel.
5. Who benefits, and who doesn't? Look at results by role and experience, not just the average.
6. Who owns the output? If it reaches a customer, someone in your organization is accountable for it, as Air Canada learned.
7. What are people already using? Ask staff which AI tools they use on their own. That tells you where demand is and where your data may be going.

Sources

1. Ryseff, J., De Bruhl, B. F., & Newberry, S. J. (2024, August 13). The root causes of failure for artificial intelligence projects and how they can succeed: Avoiding the anti-patterns of AI (RR-A2680-1). RAND Corporation. rand.org/pubs/research_reports/RRA2680-1.html
2. MIT Project NANDA. (2025, July). The GenAI divide: State of AI in business 2025. As reported in Ramel, D. (2025, August 19). MIT report finds most AI business investments fail, reveals "GenAI divide." Virtualization Review. virtualizationreview.com/articles/2025/08/19/mit-report-finds-most-ai-business-investments-fail-reveals-genai-divide.aspx
3. Wilkinson, L. (2025, March 14). AI project failure rates are on the rise: report (S&P Global Market Intelligence survey data). CIO Dive. ciodive.com/news/AI-project-fail-data-SPGlobal/742590
4. Gartner. (2025, June 25). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 [Press release]. gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
5. Tinkoff, D., Van der Veken, L., Chui, M., & Balakrishnan, T. (2026, August 25). The state of AI in 2026: On the road to ROI. McKinsey & Company. mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
6. Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. doi.org/10.1093/qje/qjae044
7. Humlum, A., & Vestergaard, E. (2025). Large language models, small labor market effects (BFI Working Paper No. 2025-56). Becker Friedman Institute, University of Chicago. bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-3.pdf
8. Becker, J., Rush, N., Barnes, B., & Rein, D. (2025, July 10). Measuring the impact of early-2025 AI on experienced open-source developer productivity. METR. arXiv:2507.09089. metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study
9. Becker, J., Rush, N., Cunningham, T., Rein, D., & Mahamud, K. (2026, February 24). We are changing our developer productivity experiment design. METR. metr.org/blog/2026-02-24-uplift-update
10. Niederhoffer, K., Rosen Kellerman, G., Lee, A., Liebscher, A., Rapuano, K., & Hancock, J. T. (2025, September 22). AI-generated "workslop" is destroying productivity. Harvard Business Review. hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity, survey figures from BetterUp Labs, Workslop. betterup.com/workslop
11. Moffatt v. Air Canada, 2024 BCCRT 149 (Civil Resolution Tribunal of British Columbia, February 14, 2024). decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
12. Do, V., Sood, S., & Johnston, C. (2026, June 11). Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026. Statistics Canada. www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2026010-eng.htm