- The disillusionment is real and measurable: 42% abandoned most of their AI initiatives in 2025 (2024: 17%), 95% of GenAI pilots show no measurable P&L contribution, over 80% of AI projects fail overall.
- But the causes are organisational, not technical: 73% of failed projects never agreed success criteria. A tool was bought; no process was built.
- Stock bubble ≠ technology error: after dot-com, stocks fell — the internet stayed. What matters operationally is not which share falls, but whom you depend on.
- For the mid-market the disillusionment is an opportunity: more mature tools, more honest prices, less hype competition — for everyone who starts properly.
“Is the AI bubble bursting?” is everywhere again — in the financial media because of valuations, in the trade media because of project track records. The two get mixed constantly, and that is exactly where the wrong conclusion is born. So first, the separation the debate usually skips: there is a stock-market question (are AI shares and valuations too high?) and an operational question (does AI deliver measurable value inside companies?). The market answers the first. The second now has solid data — and it rewards a close look.
The numbers behind the disillusionment
| Finding | Number | Source |
|---|---|---|
| Companies that abandoned most of their AI initiatives | 42% (prior year: 17%) | S&P Global, 2025 |
| GenAI pilots with no measurable contribution to the P&L | 95% | MIT (Project NANDA) |
| AI projects that fail overall — twice the rate of conventional IT projects | > 80% | RAND Corporation |
| Forecast: abandonment of AI projects lacking AI-ready data through 2026 | 60% | Gartner |
| German executives reporting higher revenue from AI | 11% | PwC executive survey |
| Failed projects that never agreed success criteria upfront | 73% | Industry analyses 2025/26 |
Compilation: Digital Maker, as of July 2026 — details in “Sources and context”
This is no longer a footnote; it is a pattern. And it matches the German picture: 59% of the mid-market does not use AI productively — the technology has arrived almost everywhere, value creation almost nowhere. Forrester therefore expects 2026 to be the year of re-sorting: roughly a quarter of planned AI spending slips to 2027 because proof of economic value is missing.
Why the projects actually fail
The obvious reading would be: the technology does not deliver. The data says otherwise. The documented root causes are almost entirely organisational:
- No defined success. 73% of failed projects had no agreed success criteria before starting. If you never define what should improve, in numbers, you can only fail — you just notice later.
- Data not AI-ready. The most common technical killer is not the model but the company’s own data: scattered, unstructured, outdated. That is why Gartner expects 60% of exactly these projects to be abandoned.
- Pilot without operations. The pilot impresses in the meeting, then nobody owns the daily grind: who maintains the prompts, checks quality, updates the model? Without an operating plan, every pilot dies on day three after the demo.
- Ten pilots instead of one process. Many companies have AI “a little bit everywhere” — and productive nowhere. The successful ones do the opposite: one frequent, expensive process, done completely.
In short: a tool was bought, but no process was built. It is the same diagnosis we know from our own work — and the reason we now only build pilots with acceptance criteria defined upfront.
What this has to do with a bubble — and what not
On the stock-market question, look back: after the dot-com crash of 2001, internet stocks lost over 80% — and yet every single thesis about the internet’s future was right. The bubble corrected valuations, not the technology. Whoever read falling prices as “the internet is dead” missed the decade that followed. The parallel is obvious: even if AI valuations correct sharply in 2026, a model that writes documentation and drafts quotes today will still do so tomorrow.
Operationally, a different question decides everything: whom do you depend on when things get turbulent? A valuation correction hits vendors first — and their customers with them: price hikes, discontinued products, acquisitions, switched-off models. 2026 has already provided the demonstrations, from the 19-day model shutdown by government order to the vendor that became its own partner’s competitor. If your processes are hard-wired to a single provider, a vendor crisis becomes your operational crisis. If you build on open, locally operable models behind an abstraction layer (in German), you read the bubble headlines as a spectator.
Why the disillusionment is an opportunity for the mid-market
It sounds counterintuitive, but the disillusionment is the best news for everyone who is serious. Three reasons:
- The tools matured while the hype fell. Open models have caught up to top-tier level on everyday tasks and run on a machine in your own building — unthinkable in 2023, at the peak of the hype.
- Prices are becoming honest. The hype premium is draining out of consulting offers and tool prices. What remains has to pay off — good for every buyer who does the math.
- The competition thins out. If 42% abandon, it also means: most of your competitors are not seriously implementing right now. Whoever gets one process properly into production now gains a lead nobody catches up cheaply later.
The four things to settle before the first euro are the failure list above, turned positive: success criteria in writing, one process instead of ten, an honest data assessment, operations planned from day one. Those are exactly the four points our sovereignty audit (€2,000 fixed price) checks before anything gets built — so a project never enters the 80% statistic in the first place.
Sources and context
42% abandonment rate (prior year 17%): S&P Global Market Intelligence, “Voice of the Enterprise” 2025. 95% of GenAI pilots without measurable P&L contribution: MIT Project NANDA (“The GenAI Divide”, 2025). Over 80% project failure: RAND Corporation. 60% abandonment forecast for projects lacking AI-ready data: Gartner. 11% of German executives reporting revenue effects: PwC executive survey, reported 2026. Deferral of roughly 25% of planned AI spending and the 2026 “re-sorting”: Forrester 2026 predictions. The 73% figure on missing success criteria comes from 2025/26 industry analyses of failed projects and should be read as an order of magnitude. All figures are snapshots of their respective surveys; assessments and recommendations reflect Digital Maker’s view based on our own project experience and are not investment advice.
Frequently asked questions: AI bubble and AI disillusionment 2026
Is the AI bubble bursting in 2026?
Two questions need separating. The stock-market question: AI valuations may correct — Forrester expects 2026 to be the year companies re-sort their AI strategies, and roughly a quarter of planned AI spending is being pushed back. The operational question: the technology itself is not going away. After the dot-com crash of 2001, stocks fell but the internet stayed — and the companies that used it soberly won the following decade. AI will stay the same way; only the hype premium disappears.
Why do so many enterprise AI projects fail?
Rarely because of the models. The documented root causes: 73% of failed projects had no agreed success criteria before starting, the data foundation was not AI-ready (which is why Gartner expects 60% of such projects to be abandoned through 2026), nobody owned day-to-day operations, and pilots never became production. In short: a tool was bought, but no process was built.
What does the AI disillusionment mean for the mid-market?
It is more opportunity than threat. The hype premium is disappearing from prices and expectations, the tools have matured, and open models are better and cheaper than ever. Whoever starts now with realistic expectations — one process, measurable criteria, a proper operating plan — gets better technology at more honest prices, with fewer competitors implementing seriously.
How do you start an AI project that does not join the 80% failure statistic?
Four things before the first euro: first, agree success criteria in writing (what exactly should improve, in numbers?). Second, pick one single frequent process instead of ten pilots. Third, assess your data honestly — poor data is the most common technical killer. Fourth, plan operations from day one: models and processes change, someone must be responsible. Exactly these four points are what a proper audit checks before any build starts.
Does sovereign AI protect against the AI bubble?
Nothing protects against falling share prices. But architecture protects against the operational fallout of vendor turbulence: an open model running locally or on your own EU infrastructure keeps working — no matter whether its vendor raises prices, gets acquired or disappears from the market. 2026 already delivered the demonstrations: a 19-day model shutdown by government order, and a vendor that became its own partner’s competitor.
Would your AI project survive the 80% statistic?
The sovereignty audit answers exactly the four questions most projects fail on: success criteria, data readiness, operating plan, dependencies. €2,000 fixed price, results in two weeks, the report is yours — even if you continue without us.