Why More Than 80% of AI Projects Still Fail — and What the Rest Do Differently

Two years ago we wrote about why enterprise AI projects fail. The reasons haven't changed. The numbers have gotten worse.

Global AI spending keeps climbing — IDC forecasts it growing from about $235 billion in 2024 to more than $630 billion by 2028. Companies are pouring money into AI at a pace that makes the dot-com era look restrained. And yet the failure numbers barely move. According to RAND Corporation, by some estimates more than 80% of AI projects fail — twice the failure rate of IT projects that don't involve AI. IDC research commissioned by Lenovo found that of the 33 AI proofs-of-concept an average organization launched, only 4 reached production. McKinsey's State of AI survey reports that while 88% of organizations now use AI in at least one business function, just 39% see any EBIT impact at the enterprise level. And here's the one that should keep every CTO up at night: in CloudBees' 2026 State of Code Abundance survey, 81% of enterprise technology leaders report production issues tied to AI-generated code.

The industry is spending more, shipping more, and failing at roughly the same rate as two years ago.

The numbers are brutal

Let's put this in perspective. More than 80% of projects failing, against spending that IDC sees passing $630 billion by 2028 — that is an enormous amount of work burned on demos that never reached production, pilots that never scaled, and "AI transformations" that transformed nothing except the consulting firm's revenue.

The stat that bothers us most, from the same CloudBees survey: 36% of organizations either track AI spend without measuring ROI or don't measure ROI at all. They spend six or seven figures on implementation and then never check if it actually worked. That's not a technology problem — that's an accountability problem.

What hasn't worked

Three things that companies keep trying, despite the evidence:

1. Throwing bigger models at bad problems

GPT-5 is remarkable. Claude Opus 4 is remarkable. But a state-of-the-art model pointed at the wrong problem is still a waste of money. We've seen companies spend €200K on a custom GPT-5 integration to automate a process that three people handle in 20 minutes a day. The AI worked perfectly — and saved roughly €800/month. The payback period was 21 years.

2. Spending more without measuring more

AI budgets keep growing year over year. Measurement practices haven't changed. Companies are scaling investments in systems they can't prove are working. We talked to a financial services firm last quarter that had 14 AI initiatives running simultaneously. When we asked which ones were profitable, the answer was "we think most of them are." Think. Not know.

3. Consultant-led transformations without internal ownership

The pattern: a Big Four firm comes in, builds a roadmap, delivers a pilot, and leaves. Six months later, the pilot is abandoned because nobody internal knows how to maintain it, the consulting firm's team has moved on, and the "AI Center of Excellence" they created is three people who've never deployed a production system.

What the successful minority actually does

Our team has been in this space for 13 years. The projects that succeed share three practices that the failures almost never have.

1. They measure before they build

Before writing a single line of code, they define what success looks like in business terms — not accuracy percentages, not F1 scores, but euros saved, hours recovered, error rates reduced. They measure the current baseline. They set a target. And they build a dashboard that tracks progress from day one.

We now include a measurement framework in every project proposal. Not because clients ask for it — most don't — but because it's the single biggest predictor of project success. If you can't measure the impact, you can't prove the value, and a project that can't prove its value gets killed in the next budget cycle.

2. They start small and prove value before scaling

The successful minority doesn't launch with a company-wide AI transformation. They pick one process, one team, one measurable outcome. They prove it works. Then they expand. It's slower, it's less impressive in board presentations, and it's dramatically more likely to succeed.

One of our manufacturing clients started with a single production line. AI-assisted quality inspection on one machine. It took 6 weeks and cost €35K. When it cut defect escape rates by 40%, the conversation about scaling to 12 lines was easy — because we had data, not promises.

3. They build governance alongside the system, not after

Who reviews the model's outputs? What happens when it's wrong? How do you retrain when performance drifts? Who owns the data pipeline? These questions get answered before launch, not six months later when something breaks.

The projects that skip governance tend to work fine for 3-6 months and then slowly deteriorate. Nobody notices because nobody is watching. By the time someone does notice, the trust is gone and the project gets shut down.

The measurement problem

That measurement stat — the 36% who track AI spend without measuring ROI, or don't measure it at all — points at the root cause of most AI failure. Not bad models, not bad data, not bad engineering. Bad accountability.

Here's what we've started doing: every system we build includes a simple measurement dashboard that tracks three things. The time or cost saved compared to the pre-AI baseline. The error rate compared to the pre-AI baseline. The total cost of running the system including API costs, maintenance, and human review time.

It takes about 2-3 days to build. It adds maybe 5% to the project cost. And it's the reason our clients renew — because they can see, in actual euros, what the system is doing for them.

If you can't show the CFO a dashboard that proves the AI is saving money, the AI is one budget cut away from being shut down.

What this means if you're planning an AI project in 2026

The technology isn't the bottleneck anymore. The models are extraordinary. The tooling is mature. The infrastructure exists. What's missing is discipline.

If you're planning an AI project this year, start with measurement. Define what success looks like before you write an RFP. Pick one process where the ROI math is obvious. Build governance from day one. And don't scale until you've proven value on a small scope.

The minority who succeed aren't using better models or spending more money. They're asking better questions and measuring the answers. That's it. That's the whole secret.