Four widely quoted studies put the enterprise AI failure rate somewhere between 30% and 95%. That spread comes from four different definitions of failure, measured across four different populations, and knowing which is which changes what a board should do about it.
The short answer
- The studies count different things: abandonment before production, completion without value, and absence of measurable P&L impact.
- Where they agree is the cause. Data quality, unclear business value and weak risk controls appear in every one of them; model choice appears in none.
- The failure is normally locked in at approval, when a pilot is funded without a named owner, a baseline measurement or a route into an operational workflow.
- The most useful board question is what has to be true in the operating model for this to survive contact with production.
The four numbers, and what each one counted
The figures circulate without their definitions attached, which is how the same organisation can be told in the same week that 30% of projects fail and that 95% do. Both are defensible. They measure different events.
| Source | Figure | What it counted | What it does not tell you |
|---|---|---|---|
| Gartner, July 2024 | At least 30% | Generative AI projects predicted to be abandoned after proof of concept by the end of 2025. | A forecast rather than a measurement, and abandonment is the only outcome counted. A project that shipped and did nothing is a success here. |
| RAND, 2024 | 80.3% | AI projects failing to deliver intended business value, split three ways: 33.8% abandoned before production, 28.4% completed with no value delivered, 18.1% delivering value too small to justify the cost. | Covers AI projects broadly rather than generative AI specifically, and draws on practitioner interviews rather than audited financials. |
| Gartner, June 2025 | Over 40% | Agentic AI projects predicted to be cancelled by the end of 2027, with cost, unclear value and inadequate risk controls named as drivers. | Also a forecast, and Gartner notes that a large share of vendor offerings marketed as agentic do not meet the definition. |
| MIT NANDA, 2025 | 95% | Generative AI pilots showing no measurable profit-and-loss impact, in The GenAI Divide: State of AI in Business 2025. | Built on 52 executive interviews, 153 survey responses and 300 public deployments. It is a small, non-peer-reviewed sample, and "no measurable P&L impact" is a demanding test that most IT investment of any kind would also fail. |
Rank them by how strict the test is and the spread disappears. Roughly a third of projects never reach production at all. Around 8 in 10 fall short of the value case that justified them. Almost none can point to a number in the accounts. Those three statements are compatible, and an executive committee that holds all three at once has a more accurate picture than one arguing about which figure is correct.
What the studies agree on
The causes converge far more tightly than the headline rates. Gartner names poor data quality, inadequate risk controls, escalating costs and unclear business value. RAND's practitioner interviews point at misunderstood or miscommunicated problem definitions, inadequate data, and organisations chasing the technology rather than a problem. The MIT work centres on the gap between tools that demo well and workflows that retain and act on what the tool produces.
Read together, the recurring causes are organisational. Data that cannot be trusted at the point of decision. A problem statement nobody senior owns. A business case whose benefit was never baselined, so its delivery cannot be tested. A workflow that has no place to put the output.
Model selection does not appear on that list in any of the four. Neither does compute, nor the choice of hyperscaler, which is worth noting given how much of the available advice is organised around exactly those two decisions.
The failure is usually locked in at approval
By the time a pilot is running, most of its outcome is already determined. Three commitments made at the approval gate account for a large share of what happens next.
1. Whether a named executive owns the operational outcome
Ownership of the outcome, which is a different thing from ownership of the project. A pilot sponsored by a function that cannot change the process it touches will produce a working demonstration and stop there, because adoption requires a decision its sponsor has no authority to take. Where the sponsor and the process owner are two different people, the handover between them is where the value is lost.
2. Whether the benefit was baselined before the work started
A pilot approved on "improve forecast accuracy" without a recorded starting accuracy cannot be evaluated, so it will be evaluated on enthusiasm. This is the single cheapest failure to prevent, and it takes a fortnight of measurement before the build starts. Once the pilot is live, the pre-intervention number no longer exists and cannot be recovered.
3. Whether the data it depends on has an owner and a service level
Pilots are built on a curated extract. Production runs on the live feed, with its outages, its schema changes and its unannounced upstream edits. A pilot whose data source has no named owner and no agreed refresh commitment has an expiry date it does not know about. This is the mechanism behind "poor data quality" in the Gartner list, and it is a governance gap rather than a technical one.
The constraint
In most organisations reporting stalled pilots, the binding limit is the absence of an operational owner with the authority to change the process the pilot touches. Adding models, budget or engineers to that situation increases spend without moving the outcome, because the decision that has to be taken is a management decision rather than a technical one.
Four questions worth asking at the next approval
- Who is accountable for the operational result, and can they change the process on their own authority? If the answer requires two directors to agree, the pilot has a governance dependency that should be resolved before funding, not after the demonstration.
- What is the number today, measured how, and by whom? A business case without a pre-measurement is a forecast with no way to be wrong.
- Which live data feed does this depend on in production, who owns it, and what have they committed to? If nobody has committed to anything, the pilot is being built on a courtesy.
- What happens to the workflow on the day this is switched on? If the answer is that people will use it when they choose to, adoption is being left to preference, which is where the MIT finding on retention comes from.
None of these questions require the board to understand the technology. All four are answerable in a paragraph each, and a proposal that cannot answer them is not ready for capital regardless of how strong the underlying idea is.
What this means for the investment case
Treating the failure rates as an argument against investing reads them backwards. What they describe is a category of spend where the variance between organisations is enormous and the differentiator is organisational readiness rather than technology selection. The same model, bought at the same price, produces a measurable margin improvement in one company and a dormant licence in another.
That makes readiness the thing worth diagnosing first, and it is diagnosable in weeks. Where decisions are made, what evidence they rely on, who owns the data underneath, and how long it takes an approved change to reach an operational process: those four things predict AI outcomes better than any assessment of the technology estate.
Sources
- Gartner, Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, July 2024.
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 2025.
- RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, 2024.
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, 2025.