AI has arrived in the mainstream of business: according to the ifo Institute's May 2026 survey, 54.5 per cent of German companies now use AI, up from 40.9 a year earlier. That shifts the risk. The question is no longer whether you use AI, but whether your project ends up among those that actually get something out of it. Because alongside adoption, the abandonment rate has climbed - faster than the analysts predicted. We laid the credible studies of the past two years side by side. The result is uncomfortable and comforting at once: it is always the same five patterns, and every one of them is recognisable before the budget flows.
First, the scale of the problem. Gartner predicted in July 2024 that at least 30 per cent of GenAI projects would be abandoned after proof of concept by the end of 2025 - due to poor data quality, inadequate risk controls, escalating costs or unclear business value. The measurement then came in harder than the forecast: in the 2025 enterprise survey by S&P Global Market Intelligence, 42 per cent of companies reported abandoning the majority of their AI initiatives - up from 17 per cent the year before. On average, 46 per cent of proof of concepts were scrapped before reaching production.
Why does nearly half get cancelled? The most thorough answer comes from the RAND Corporation, which interviewed 65 experienced data scientists and engineers about failed projects. Their five root causes align strikingly with what Gartner, McKinsey and Deloitte measure on the business side. From that alignment come the five patterns.
Pattern 1: the project has a tool, but no problem
The most frequently named cause in the RAND interviews is not technical: stakeholders misunderstand or miscommunicate which problem the AI is supposed to solve in the first place. In third place comes its mirror image: a focus on the latest technology rather than a real problem. Together they produce the classic project that starts with "we need to do something with AI" and ends with a tool nobody misses when it breaks.
That focus is measurably worth money is shown by BCG's analysis of 1,803 C-level respondents: companies that extract significant value from AI pursue 3.5 use cases on average. The laggards spread themselves across 6.1 - and the focused group achieves 2.1 times the return. More projects do not mean more results; they mean thinner budgets, thinner attention and thinner data per project.
Early warning sign: the project name states the technology instead of the process. "Roll out ChatGPT" is not a project. "Cut quote preparation from four days to one" is. Whether that even requires an AI agent or a simple workflow will do is the second question - the boundary is drawn in AI agent or automation workflow.
Pattern 2: the data is not ready
Root cause number two at RAND: the organisation lacks the data to carry the project. Gartner named poor data quality first among its four abandonment reasons. And in Informatica's CDO survey from January 2026, 57 per cent of the 600 data leaders surveyed name a lack of data reliability as the key barrier on the way from pilot to production - though it should be said that Informatica, as a data management vendor, makes its living from precisely this narrative. The direction still holds; it matches the independent sources.
The treacherous part of this pattern: it shows up late. In the proof of concept, the team works with hand-picked examples and everything looks fine. Production then brings the real data - duplicates, stale entries, missing fields, three spellings per customer. Why that ruins any automation, and how to assess the honest state of your data in an afternoon, is covered in CRM hygiene before AI.
Early warning sign: nobody in the room can say where the data lives, how much of it there is and how old it is. If those three questions require research, that research is the project's first working week - before the business case, not after it.
Pattern 3: the tool arrives, the process stays
The best-evidenced single finding in the entire body of research comes from McKinsey: across 25 tested organisational attributes, redesigning workflows has the biggest effect on the bottom-line impact of generative AI. Not the model, not the budget, not the tool choice: the process. At the same time, Deloitte's State of AI 2026 (3,235 executives across 24 countries) shows that only 30 per cent of organisations redesign their processes around the AI. The other 70 per cent lay the tool on top of the old workflow and wonder why the time savings evaporate in daily use.
For the German-speaking market there is a staffing problem on top: in Bitkom's representative survey of 604 companies, 53 per cent name a lack of technical know-how and 51 per cent a lack of staff capacity as the biggest obstacles. A new process needs someone who builds it, explains it and repairs it in the first weeks. If that person is not named, nobody takes the role.
Early warning sign: the project plan contains no sentence about what changes in the participants' workflow. If everyone works exactly as before after the rollout, it was not a project - it was a licence purchase.
Pattern 4: the pilot stays a pilot
This is where everything condenses. McKinsey's State of AI from November 2025 (1,993 respondents from 105 countries) shows the gap: 88 per cent of companies now use AI in at least one function, but nearly two thirds are stuck in the experimentation or pilot phase. Only 7 per cent report running AI fully at scale. Deloitte measures the same thing from the other side: only 25 per cent of organisations have brought 40 per cent or more of their pilots into production.
The pilot is the most comfortable place in an AI project: small budget, curated data, well-disposed users, no responsibility for the worst case. Production, by contrast, demands answers to uncomfortable questions - who runs the operation, what happens on errors, which guardrails constrain the system, who is liable for the output. Gartner explicitly names inadequate risk controls as an abandonment reason, and according to Deloitte only 21 per cent of organisations have a mature governance model for AI agents. In the pilot, all of these questions are deferred; in production, all of them fall due. This is exactly the threshold where the 46 per cent from the S&P measurement die.
Early warning sign: the pilot has no end date, no named operations owner and no criteria for when it goes to production or gets switched off. A pilot without a kill criterion is not a test - it is a permanent condition.
Pattern 5: nobody measures whether it delivers
The quietest pattern, and probably the most expensive. McKinsey has quantified it: fewer than one in five companies track well-defined KPIs for their GenAI solutions - and that same KPI tracking is one of the strongest predictors of bottom-line impact. How fundamentally sound model evaluation is missing shows on the investment side too: in the BCG AI Radar 2026 (2,360 executives, including 640 CEOs), 94 per cent say they will keep investing in AI even if their initiatives deliver no payoff in 2026. You can call that confidence. You can also read it this way: the majority has no measurement that could stop them.
This pattern is also home to the most-quoted AI figure of the past year: the 95 per cent with zero measurable return from the MIT NANDA report. We took that figure apart in its own article - the short version: it measures organisations without a measurable P&L effect, not failed projects, and the study is small and preliminary. But even if you halve it, you land on a finding that fits every other source: the normal case is not the AI project that causes damage. The normal case is the AI project nobody can say anything about.
Early warning sign: "success" is an adjective in the project proposal, not a number. If it says "more efficient", "faster" or "relieved" but shows no baseline and no target value, nobody will want to measure after the rollout - because by then, the result belongs to someone.
Three levers for your next project start
Write the one sentence before you choose the tool. "We cut process X from A to B, measured by Z." If that sentence does not exist, the venture is an experiment - which is legitimate, but then it gets an experimentation budget and an end date instead of a rollout plan.
Pull 50 real cases before you build the business case. Take fifty real records or transactions and work through them by hand. What you learn about duplicates, gaps and edge cases decides the project - and it costs an afternoon instead of a failed quarter.
Write the production criteria into the pilot. A named operations owner, one metric with a baseline, defined guardrails for what the system must not do, and a date on which the decision falls: scale or switch off. All of it fits on one page, and that page separates the 7 per cent from the two thirds.
If you have an AI project in front of you and want to know which of the five patterns is already built into it: we are happy to walk through the idea with you before any budget flows. It usually takes one conversation, not an audit. 🧭