Why AI projects don't reach production: what the failure numbers actually measure
Depending on the study, 30 to 95% of AI projects don't make it: RAND measures over 80% failure, Gartner at least 30% abandoned after POC, MIT 95% of pilots with no measurable return, S&P Global 42% of companies giving up. These numbers don't measure the same thing, but they point to four causes you can address at scoping. A sourced breakdown and a go/no-go method for CIOs, CTOs and innovation leaders.
Transparency note. This article follows the IgnitionAI editorial policy. Each failure rate is attributed to its primary source, with its date and methodology. The scoping recommendations are IgnitionAI estimates based on our 3 engagements in 2024-2025, tagged as such.
Depending on the study you look at, 30 to 95% of AI projects don't make it. RAND measures over 80% failure, Gartner at least 30% abandoned after POC, MIT 95% of pilots with no measurable return, S&P Global 42% of companies giving up on most of their initiatives. These numbers don't contradict each other: they measure different things. And they point to the same four causes, all addressable at scoping.
Four numbers, four definitions
A failure rate means nothing without its denominator. Here is what each study actually counts.
| Source | Date | Figure | What it measures | Methodology |
|---|---|---|---|---|
| RAND | Aug 2024 | over 80% | AI projects that "fail", twice the rate of non-AI IT projects | 65 interviews with experienced engineers and data scientists |
| Gartner | Jul 2024 | at least 30% | generative-AI projects abandoned after POC by end of 2025 | analyst forecast |
| Gartner | Jun 2025 | over 40% | agentic-AI projects cancelled by end of 2027 | analyst forecast |
| MIT NANDA | Jul 2025 | 95% | organisations with no measurable return on their GenAI pilots | 300+ deployments analysed, 52 interviews, 153 leaders |
| S&P Global | 2025 | 42% | companies that abandoned most of their AI initiatives, vs 17% in 2024 | over 1,000 companies (North America, Europe) |
These numbers measure distinct objects: "failing" (RAND), "being abandoned after the POC" (Gartner), "delivering no measurable return" (MIT), "giving up on most initiatives" (S&P Global). None is wrong. Read together, they tell the same story: the move from pilot to value-creating production is where most projects stop.
The trend is worsening. S&P Global measures a jump in the abandonment rate from 17% in 2024 to 42% in 2025. Gartner raised its own post-POC abandonment forecast over the year. Initial enthusiasm fades once the bill and the real complexity show up.
The four causes that keep coming back
The studies converge on the same factors. They are structural, and each can be addressed before the first line of code.
1. Miscalibrated expectations at the top. RAND's leading cause is leaders' misunderstanding of what AI can and cannot do. Goals are poorly communicated, time and resources underestimated. Only 14% of organisations say they are ready to integrate AI, while 84% of leaders consider it strategic.
2. The data isn't ready. Data quality is RAND's second cause, and one of the four abandonment reasons Gartner cites. A model that performs on clean demo data degrades on real data, which is heterogeneous and incomplete.
3. No measurable success criterion. MIT attributes failure less to the model than to a gap in integration and organisational learning. With no value metric defined upfront, a pilot "works" without anyone being able to say whether it creates a return. It ends up a permanent POC.
4. Cost and compliance discovered too late. S&P Global cites cost, data privacy and security as the top obstacles. When the real inference cost or GDPR constraints surface after development, the project stops instead of reaching production.
IgnitionAI estimate based on our 3 engagements in 2024-2025: on the projects we audit after a first internal failure, these four causes are almost always present together, and always detectable in a two-week scoping.
The answer: decide before you spend
What the four causes share is that they are visible before development, not after. The method that separates the projects reaching production from the rest comes down to one decision: a written go/no-go, ahead of the build budget.
Our scoping takes two weeks and addresses the four causes one by one:
- Expectations: a quantified business objective and a success metric, validated by the sponsor before any development.
- Data: an audit of the available sources, their quality and their sensitivity, on your real data.
- Feasibility: a map of the technical options, sometimes the conclusion that a business rule or a better tool fits better than AI.
- Cost and compliance: an estimate of inference and infrastructure costs, and the regulatory perimeter, set from the start.
The deliverable is a written opinion: go, or a reasoned no-go. A no-go after €8,000 of scoping costs a thousand times less than a project abandoned after six months of development. We detail these cost lines in our article on the cost of a RAG project, and the format in our engagement models.
What the 5% who succeed do
MIT describes a divide between a minority that extracts value from AI and a majority that stalls. The organisations on the right side share three traits: a narrow, precise scope rather than a sprawling ambition, careful integration with the existing information system, and the choice to buy or partner rather than rebuild everything in-house.
That is exactly the model we apply: a scoping that narrows the work to a defensible use case, a production rollout integrated with your IS and monitored, and full code transfer to your teams at delivery. Governance and access control are built in from the design stage, as detailed on our governance page.
FAQ: AI project failure
-
What percentage of AI projects actually fail?
There is no single number, because the studies don't measure the same thing. RAND finds over 80% of AI projects failing (August 2024). Gartner forecasts at least 30% of generative-AI projects abandoned after POC by end of 2025. MIT reports 95% of pilots with no measurable return (July 2025). S&P Global measures 42% of companies abandoning most of their initiatives in 2025.
-
Why don't AI POCs reach production?
Four causes recur across the RAND, MIT and S&P Global studies: miscalibrated expectations at leadership level, insufficient data quality, no measurable success metric, and cost or compliance constraints discovered too late. All four are detectable before development, during a scoping.
-
How do you know whether an AI project is worth it before investing?
Through a written go/no-go scoping, ahead of the development budget. Two weeks are enough to quantify the expected value, audit the real data, compare the technical options and estimate inference and compliance costs. The deliverable is a written opinion that authorises the next phase, or not.
-
What sets apart the AI projects that succeed?
According to MIT (2025), the organisations that extract value from AI share three traits: a narrow, precise scope, careful integration with the existing information system, and the choice to rely on a partner rather than build everything in-house. Success comes down to execution and integration, rarely to the raw performance of the model.
Methodology and sources
Studies cited (failure rates):
- RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (J. Ryseff, B. De Bruhl, S. Newberry), August 2024, based on 65 interviews with engineers and data scientists: rand.org
- Gartner, press release of 29 July 2024, 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025: gartner.com
- Gartner, press release of 25 June 2025, Over 40% of Agentic AI Projects Will Be Canceled by End of 2027: gartner.com
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (A. Challapally, C. Pease, R. Raskar, P. Chari), July 2025, 300+ deployments, 52 interviews, 153 leaders: nanda.media.mit.edu
- S&P Global Market Intelligence, Voice of the Enterprise 2025, over 1,000 companies (North America, Europe): spglobal.com
IgnitionAI estimates: the mission observations (the four causes appearing together, detection at scoping) rest on 3 AI audit or design engagements delivered in 2024-2025 for French mid-market companies in regulated sectors. Possible variation depending on context. See our editorial policy.
Sources last reviewed: 2026-06-13.
Related IgnitionAI articles:
- How much does an enterprise RAG project cost in 2026
- Enterprise RAG in production: 5 critical decisions your first POCs hide from you
- The European AI Act: what mid-market CTOs must prepare after the Digital Omnibus
Has an AI project already cost you dearly without reaching production, or do you want to avoid that? Our two-week scoping ends with a written go/no-go, with no commitment to continue. Request a conversation.