Why AI projects fail: an implementation problem, not a technology problem
Why AI projects fail can be answered clearly for industry: rarely because of the technology, mostly because of the implementation. What is missing is data access, a baseline before the start, clear usage rules and operational ownership after the pilot. According to Deloitte 2026, only one AI use case in five in manufacturing is scaled across the company.
For decision makers in brief
- Core finding: Industry has an implementation problem, not a technology problem. Representative surveys and our analysis of 52 published practitioner conversations (September 2025 to August 2026) agree on this.
- Effort: According to Schneider Electric, 80 to 90 % of the effort in an AI transformation goes into change management, training and workflow redesign (company statement, April 2026).
- Cost: The hidden item in the business case is the foundation, such as data architecture and interfaces. According to Capgemini 2025 (consulting study), 80 % of industrial initiatives first need such investment.
- Risk: If you start without a baseline, you cannot prove success later and will not get a second budget.
- Remedy: One narrowly scoped use case, one metric, regulated operations.
The core finding: why AI projects fail
In our analysis of the practitioner conversations, 206 individual pain point statements were coded. 42 of them concern data, IT/OT and integration, 32 leadership, culture and change. Technology as such hardly appears as a problem. The model generally works. It simply cannot reach the data, it does not fit into the workflow, or nobody has time to run it.
Philippe Rambach, Chief AI Officer at Schneider Electric, put it like this in April 2026: "80 to 90 percent of the work on an AI transformation is change management, training and workflow redesign, not technology." This is a company statement, yet it matches the picture from the conversations.
The barriers in figures
The representative surveys confirm the direction. Bitkom is the German digital industry association, IW Köln the German Economic Institute and VDMA the German mechanical engineering association. It is striking that money does not come first.
| Barrier | Share | Source |
|---|---|---|
| Data protection requirements | 77 % | Bitkom 2026, n=604 |
| Shortage of skilled workers | 70 % | Bitkom 2026, n=604 |
| Lack of time | 66 % | Bitkom 2026, n=604 |
| IT security | 61 % | Bitkom 2026, n=604 |
| Funding | 48 % | Bitkom 2026, n=604 |
| Unclear ROI, across industries | 62.7 % | IW Köln 2025, IW-Zukunftspanel |
| Lack of staff resources for implementation | 45 % | VDMA, position paper Industrial AI 2026 |
| Return on investment not proven | 44 % | VDMA, position paper Industrial AI 2026 |
| Insufficient data quality | 42 % | VDMA, position paper Industrial AI 2026 |
Two figures deserve a second look. First, the unclear or unproven return: it is often not a property of the technology but the consequence of a missing baseline. Second, the lack of time: a pilot that runs on the side has no owner as soon as project pressure eases.
According to the Federal Network Agency (Bundesnetzagentur) 2025, 6 % of companies tested AI and then discontinued its use. Almost half of the companies using AI invested zero euros in AI in the previous year. If you start initiatives without a budget for foundations and operations, disappointment is built in from the start.
Five causes and their remedies
If you want to understand why AI projects fail, the findings of the industry report reveal five recurring causes. Each has a proven remedy.
| Cause | How you recognise it | Remedy |
|---|---|---|
| Selection by data availability | The use case was chosen because data happened to be available there | Select equipment and processes by business criticality |
| No baseline | Nobody knows the starting metric from the system | One metric per use case, baseline from the previous year, defined before the start |
| One-off solution | The pilot runs on only one line and only with special effort | Build the pilot so that it can be transferred to plants two and three |
| Technology without people | Process, training and metrics only follow after the project | Process redesign and capability building belong in the project |
| No operations | Unclear who monitors and retrains model quality | Define operational ownership before the rollout |
The fifth cause is the one most often underestimated. Models deteriorate over time. If you do not define operational ownership, you lose the value again after about twelve months. In the practitioner conversations, this very topic is a blind spot: there is a lot of talk about use cases and almost none about where the model runs and who operates it. Regulated AI operations as a managed service make monitoring, retraining and documentation binding.
The governance vacuum
A spring 2026 survey by the ZEW Leibniz Centre for European Economic Research of around 1,500 companies shows how unregulated the use of AI in manufacturing often is: only 30 % actively provide AI applications to their employees, 39 % merely tolerate their use, 8 % expressly prohibit it.
Tolerance is not governance. Where employees use AI without rules, approved tools and data classification, design data, cost calculations and customer information migrate into public services. A ban does not solve the problem, it shifts it into shadow IT. A short, understandable policy and an approved tool are more effective.
Co-determination is part of this. In practice it has proven worthwhile to involve the works council before the pilot rather than only before the rollout, because a works agreement negotiated after the fact is regularly more expensive than one agreed in advance.
What successful projects do differently
Only 14 of the 52 conversations analysed mention a result with a figure at all. These documented cases from industry show two patterns. They are company statements from our own analysis, not projects we were involved in.
Pattern 1: speed is the real currency
Remarkably often, the figure does not prove the technology but the speed of implementation: six weeks at VW in the Gläserne Manufaktur, ten months to start of production in a platform project, under 24 months per plant at Progroup. If you are championing an initiative internally, the time to the first result convinces more than percentages. This is why an implementation framework with a fixed duration pays off, as described on the page on our AI project fixed price, with 72 hours to the first pilot and 4 to 10 weeks of implementation.
Pattern 2: the lever lies before the AI
Several of the strongest figures do not come from AI projects but from process work: the switch from push to pull at RUWAC, standardisation at Progroup, the move from paper to digital boards at AbbVie. Several interviewees say the same thing: first clarify the physical process, then automate.
Three findings from the studies
The industry report summarises three further differences. Solutions developed through strategic partnerships reach production significantly more often than purely in-house developments. Systems embedded in existing workflows are effective, not generic assistants that only improve individual productivity. In addition, the majority of industrial initiatives first need a foundation of data architecture and OT/IT integration. If you hide this item in the business case, you produce the disappointment that is later booked as a failed AI project.
Our guide to choosing an AI partner shows how to assess an implementation partner against these criteria. For the German Mittelstand, mid-sized companies with roughly 50 to 1,000 employees, our page on AI implementation for mid-sized companies gives an overview of how implementation and operations fit together.
