Most AI Pilots Never Ship
Summary: Every major study of enterprise AI reaches the same conclusion: most efforts do not deliver. The exact failure rate depends on what each study counts. MIT’s NANDA initiative reported that only about 5% of generative AI pilots achieve rapid revenue acceleration, while the rest stall with little or no measurable impact on profit and loss [1]. BCG found only about 5% of companies achieving AI value at scale [2]. S&P Global found 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier [3]. A Prosigns survey put the share of initiatives that never reach production at 82% [4]. The numbers differ because the definitions differ. The direction does not.
Why the numbers disagree
“AI failure” is not one measurement. The studies below count different things, so none of them can be averaged or compared directly.
| Source | What it measured | Result |
|---|---|---|
| MIT NANDA, 2025 | Generative AI pilots achieving rapid revenue acceleration | About 5%; the rest stall with little or no measurable P&L impact [1] |
| BCG, 2025 (1,250+ companies) | Companies achieving AI value at scale | About 5%; about 60% report little or no material value [2] |
| S&P Global Market Intelligence, 2025 (1,000+ respondents) | Companies abandoning most AI initiatives | 42%, up from 17% in 2024 [3] |
| Prosigns, 2026 (1,200 enterprise leaders) | Initiatives that never reach production | 82%, up from 73% in 2024 [4] |
Read them as a range, not a rate. “No financial return,” “no value at scale,” “abandoned,” and “never in production” are four different failures. What they share is a direction: most organizations that have started with AI have not yet gotten a return from it.
A few cautions apply. The MIT findings draw on about 150 leader interviews, a survey of 350 employees, and an analysis of 300 public AI deployments [1], a modest base for a headline number. The Prosigns study comes from a company that sells AI development services. S&P also found that the average organization scrapped 46% of its proof-of-concept projects before they reached production [3]. We cite these studies because they point the same way, not because any single figure is exact.
Where pilots actually fail
RAND interviewed 65 experienced data scientists and engineers about why AI projects fail. The most common reason was not technical. It was misunderstanding or miscommunication about what problem the project was meant to solve [5]. The second root cause RAND lists is that organizations often lack the data needed to train an effective model. RAND also notes that the commonly cited figure of more than 80% of AI projects failing is an estimate from other sources, twice the failure rate of IT projects without AI [5].
MIT’s authors reached a similar conclusion from a different angle. They located the barrier not in model quality but in a “learning gap”: generic AI tools stall in enterprise use because they do not learn from or adapt to how a business actually works [1]. S&P’s respondents named cost and data privacy and security risk as top obstacles [3].
Put together, the evidence says the model is rarely the problem. The problem is the work around it.
What the surviving minority does differently
Prosigns reports that organizations with a system in production for 12 months or longer saw a median annualized ROI of 287%, while pilot-only organizations reported a median of -34% [4]. Those are medians for two self-selected groups, and they come from a vendor survey, so treat them as a signal, not a promise. They do show how wide the gap is between a pilot and a system people rely on.
In our experience, the teams that cross that gap do four things before they write much code:
- Name one problem in business terms. Not “use AI in operations,” but ”cut the time to answer a coverage request.” This addresses the failure RAND found most often.
- Check the data first. Pilots run on a clean sample. Production runs on messy, siloed, real data. Find out which one you have before you commit.
- Decide how you will know it is right. Build a test set from real cases and agree on what “good enough” means before the demo, not after.
- Price it at production volume. Model cost per request, then multiply. A cost that looks small in a pilot can look very different at 100 times the traffic.
None of this is exotic. It is discipline, and it is mostly done before the build starts, which is why we start every engagement with a diagnosis.
Questions to ask before funding the next pilot
- Who owns this system on day 100, and who is on call when it breaks?
- How will we know when it is wrong?
- What does it cost at ten times today’s volume?
- What happens to the business if we turn it off?
If those have no answers, that is the work to do first.
Sources & References
[1] MIT NANDA: The GenAI Divide, State of AI in Business 2025, as reported by Fortune (August 2025)
[2] BCG: The Widening AI Value Gap (2025), 1,250+ firms worldwide
[3] S&P Global Market Intelligence AI survey (2025), as reported by CIO Dive
[4] Prosigns: State of Enterprise AI 2026, vendor survey of 1,200 enterprise leaders
[5] RAND: The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (Ryseff, De Bruhl, Newberry, 2024), based on 65 interviews