In short: almost every company is trying something with AI, but few see money from it. A widely cited MIT report found that as many as 95% of organisations get no measurable return on generative AI - a figure that is methodologically contested, but whose direction other independent sources confirm. The reason is almost never technological - the model works; what’s missing is a redesigned process, a clear goal and good data around it. The good news is that the same research shows what the few who succeed do differently. This post breaks the causes of failure down and shows how to avoid them in your company.
Do AI rollouts really fail that often?
Yes - and the gap between “we use AI” and “we make money on AI” is huge. Adoption is everywhere: according to McKinsey, 88% of companies now use AI in at least one function - up from 78% a year earlier. But using it isn’t the same as getting a result. A report from MIT (the NANDA project, 2025) found that 95% of organisations adopting generative AI see no measurable return, and only 5% of pilots extract real value (these are preliminary, non-peer-reviewed, self-reported findings - treat the figure itself with caution, though other studies echo the direction).
The profit side tells the same story. McKinsey reports that 39% of organisations see any enterprise-level EBIT impact from AI, while only about 6% are leaders with a truly significant effect. Other studies - from Gartner, S&P Global and RAND - reach similar conclusions on how many projects get abandoned; we gathered them in our post on what AI adoption actually is. The picture is consistent: almost anyone can buy and switch on AI, but only a handful turn it into a return.
Wondering which side of that statistic your AI idea is on? We’ll test it on a concrete process during a free consultation - on your company’s own data, not in theory.
If it’s not the technology, what actually fails?
Usually everything except the model: a badly framed goal, missing data and a process nobody changed. After interviewing experienced engineers and data scientists, RAND identified five most common causes of AI project failure:
- A misunderstood problem - the team jumps on a tool before settling what should actually improve and how they’ll know.
- Missing data of the right quality and quantity - without it, even the best model has nothing to work with.
- Chasing the technology instead of the problem - AI “because we need AI”, not because it solves a concrete pain.
- Weak infrastructure around data - the information exists, but scattered and unconnected.
- A problem too hard for today’s technology - a goal that’s out of reach by design.
Four of those five causes have nothing to do with the model itself. One of RAND’s interviewees put it bluntly: “80 percent of AI is the dirty work of data engineering”. That rarely looks impressive on a slide - which is exactly why the step gets skipped and the project stalls.
Why isn’t buying the tool enough?
Because the value isn’t in the tool, it’s in the process around it - and that has to be deliberately rebuilt. This is the hardest finding in the data: out of 25 attributes McKinsey tested, redesigning workflows was the one most strongly linked to AI’s real impact on profit - yet only 21% of companies have fundamentally redesigned even some of their processes. Most buy the tool, bolt it onto the old way of working, and then wonder why nothing changes.
That’s exactly what the BCG 10-20-70 rule describes: in an AI transformation about 70% of the value comes from people and processes, 20% from technology and data, and only 10% from the algorithms themselves. In other words: buying the model buys you 10% of success. The other 90% is work no licence covers - and it’s what separates the winning 5% from the rest.
Why does a pilot look great, then fail to reach production?
Because the demo runs on the happy path, and production is every other path at once. A pilot shows something is possible. A rollout requires it to work every day, on real data, integrated with the rest of the company, with people who actually change their habits. Most projects collapse between the two - which is why Gartner predicts at least 30% of generative AI projects will be abandoned right after the pilot stage.
We know the same pattern from building software: AI tools deliver the first 70% surprisingly fast, and all the risk sits in the last 30% - integration, bugs and maintenance. We broke that down in our post on why the spec matters in AI software. A pilot with no planned path to production usually stays there.
Build AI yourself or with a partner?
The data is fairly clear: with a good partner, twice as often. In the MIT study, tools bought from specialised vendors reached production roughly twice as often as internally built ones - about 67% of successful deployments versus roughly half that. The reason isn’t magic: a partner who has done it a dozen times knows the typical pitfalls from RAND’s list and doesn’t hit them for the first time at your expense.
That doesn’t mean handing everything off - quite the opposite. The model that works best is one where the partner brings experience and leads the first rollout, while the know-how and data stay with you. The goal is a company that can stand on its own after the project, not one permanently dependent on a single vendor.
How do you improve the odds that a rollout pays off?
Focus on what the successful 5% do - not on picking a model. In practice it comes down to a few rules:
- Start with one process and a measurable goal - not “adopting AI” in the abstract. How to choose that first one, we cover in where to start.
- Redesign the process, don’t just swap the tool - ask how this work should look with AI, not where to wedge a model into the old flow.
- Get the data in order for a narrow scope - not the whole company at once, just what this one process needs.
- Measure “before” and “after” - without a baseline you can’t prove a return. How to calculate it, we show in how much adopting AI costs.
- Engage leaders and train the team - because that’s the “people” pillar, where by the 10-20-70 rule most of the value sits.
The difference between a rollout that fails and one that pays off fits in a single table:
| A rollout that fails | A rollout that pays off | |
|---|---|---|
| Starting point | ”let’s buy AI” | one process with a measurable goal |
| What changes | the tool, the process stays | the process rebuilt around AI |
| Data | ”we’ll sort it out someday” | ready for a narrow scope |
| Proof | a feeling that it works | a “before” and “after” number |
| Team’s role | gets a tool | learns a new way of working |
Frequently asked questions
If 95% of companies see no return, is AI even worth it? Yes - because that 95% is usually the result of a bad approach, not a lack of potential. The leaders in the 5% extract real value from the same tools. The difference is in the process, the goal and the data, not in access to a better model.
Where do I start so I don’t end up in that 95%? With one narrowly chosen process and a number you can improve - and with measurement that shows early whether it works. We laid out the whole first step in where to start adopting AI.
Does a smaller company stand a better chance? Often yes - it has a shorter decision path and fewer systems to reconcile, so it rebuilds one process and sees the result faster. Scale isn’t the advantage here; the discipline of a narrow start is.
Key takeaways
- AI adoption is widespread, returns are rare - 88% of companies use AI, yet by a widely cited (if contested) MIT report as many as 95% see no measurable return from it.
- The cause is almost never technological - the goal, the data and the process fail, not the model.
- The value is in redesigning the process - that, not the tool, is where profit impact comes from; only 21% of companies do it.
- A pilot is not a rollout - most projects collapse on the road from demo to production.
- Twice as often with a partner - rollouts led with a specialised partner reach production roughly twice as often as builds done purely in-house.
- The recipe is repeatable - a narrow process, a measurable goal, good data, a redesigned way of working and an engaged team.