Headlines say that 80% or even 95% of AI projects fail. Those figures are real but narrower than they sound, and most of the causes behind them are well known and avoidable. This guide explains what the studies measure, why the catalog gives a per-use-case failure risk (low / medium / high) instead of a percentage, and what to do about each of the nine failure reasons the catalog uses. It also shows how to read the cost bands, data-flow values and lock-in levels, and how to keep a way out.
Why it matters
The international figures, and what they actually measure
| Figure | What it measures | Limits |
|---|---|---|
| MIT NANDA, "The GenAI Divide" (2025): about 5% of GenAI pilots achieve rapid revenue acceleration; the other ~95% show little or no measurable P&L impact [1] | Whether a pilot shows a measurable effect on profit and loss, based on about 150 leader interviews, a survey of 350 employees and an analysis of 300 public deployments | Mostly large US companies. "Failure" means no measurable P&L impact, not that the tool was useless (time saved without lower costs counts as no impact). Preliminary report, not peer reviewed, and the full text is only available on request. The same report finds that bought tools from specialised vendors succeed about 67% of the time, and in-house builds about one third as often [1]. |
| RAND (2024): "by some estimates" more than 80% of AI projects fail, twice the rate of non-AI IT projects [2] | RAND quotes the 80% from other estimates. Its own research is 65 interviews with experienced data scientists and engineers about why projects fail | The 80% is not RAND's own measurement. The value lies in the root causes (see below). Written mainly for the US defence sector and for machine-learning projects that train their own models. |
| Gartner (July 2024): at least 30% of GenAI projects abandoned after proof of concept by end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value [3] | Analyst forecast | A prediction, not a count. Gartner sells advisory services, and its cost examples (USD 5–20 million for business model innovation) refer to large enterprises, not a 50-person company [3]. |
| Gartner (June 2025): over 40% of agentic AI projects cancelled by end of 2027, due to rising costs, unclear value or insufficient risk controls. Only about 130 of thousands of "agentic" vendors offer real agent technology; the rest is "agent washing" (rebranded chatbots and RPA) [4] | Analyst forecast, plus a poll of 3,412 webinar participants in January 2025 | Self-selected webinar audience. Still a useful warning about hype in the agent market. |
| BCG (September 2025): 5% of companies achieve AI value at scale, 35% are starting to, 60% report hardly any material value despite substantial investment [5] | Executives' self-assessment of value from AI across the whole company | Consultancy survey with a commercial interest in transformation programmes. Measures company-wide value, not individual projects. |
| McKinsey "State of AI" (2025): 88% use AI in at least one function, but nearly two thirds have not scaled it beyond experiments and pilots; only 39% report any effect on company-wide EBIT, mostly below 5%; about 6% are "high performers" [6] | Self-reported survey of organisations worldwide | Self-assessment; "no EBIT effect" is not the same as "failed". High performers are more likely to have redesigned workflows around AI [6]. |
German and Mittelstand figures
- Use is now widespread, success less so. Bitkom (603 companies with 20+ employees, September 2026): 57% use AI, up from 36% a year earlier [7]. ifo (May 2026): 54.5% of companies use AI, but only 47.2% of medium-sized companies, compared with 67.2% of large ones [8]. KfW's Mittelstand panel shows a lower 20% for 2022–2024, because it covers all SMBs, including the smallest, and uses an earlier period [9]. Simple, broadly usable tools lead: text generation is used by 14% of the Mittelstand, text recognition by 10%, data analytics by only 3% [10].
- The barriers that companies name. IW (IW-Zukunftspanel, n = 951, 2024): 63% say the benefit of AI is hard to assess, 62% that staff must be trained first, 50% that adapting AI to their needs is laborious, 46% lack IT specialists, 46% lack access to data, 42% that the time saved is lower than expected, 40% name data protection, 39% costs that are too high, and 34% that AI does not meet quality expectations [11]. Bitkom AI users name data protection requirements (66%), legal uncertainty (56%), costs that are hard to calculate (54%) and lack of staff acceptance (42%) [7].
- But German users mostly report a positive balance. Among AI-using companies in the IW panel (n = 408), 48% see the benefit as greater than the cost, 27% as equal, and only 16% say the cost is greater [11]. That does not fit a "95% failure" story. Most German use is bought standard software, which the MIT study also finds succeeds more often [1].
- The Mittelstand-Digital view. About 60 AI trainers and experts from the Mittelstand-Digital network again rank missing know-how and skilled staff as the main barrier, followed by a missing data basis and low digital maturity [12]. KfW finds that digital maturity, digital know-how and innovativeness predict AI use better than company size does. Finding a first practical use case is itself a key hurdle [10].
Why the catalog gives low / medium / high instead of a percentage
No study measures failure rates per use case, all of them define "failure" differently, and none is representative of German SMBs. A single percentage would suggest a precision that does not exist. Each use case therefore has a failure_risk estimate based on maturity, complexity, data needs and the evidence above:
| Level | Meaning (vocab.yaml) |
|---|---|
| low | Well understood and productised; failure is rare when the basics are covered |
| medium | Works when data and process are prepared; pilots often stall without an owner |
| high | Often stalls in the pilot or misses its ROI; needs strong data, a capable team and a committed sponsor |
Each use case also lists 1–3 failure_reasons, explained below. Read the level as "how much care this needs", not as a forecast for your company.
When you need it
- When choosing use cases (GUIDE-001): prefer low-risk use cases for your first project.
- Before the pilot (GUIDE-009): turn the listed failure reasons for your use case into go/stop criteria.
- Before signing a contract (GUIDE-010): check cost bands, data flow and lock-in.
- When a running project stalls: use the nine reasons as a diagnosis list.
Who is involved
| Role | Part |
|---|---|
| Management / owner | Sponsor; sets a budget cap and decision dates; stops projects that do not deliver |
| Process owner (department lead) | Owns the problem, the baseline and the KPIs; most "unclear goal" failures start here |
| End users | Test early; their acceptance decides whether the tool is used |
| IT / IT security | Assess integration effort, operation and data flows |
| DPO, works council, legal | Clear compliance points early, not at rollout (GUIDE-004, GUIDE-005) |
| Data owners | Confirm that the data exists, is accessible and is good enough (GUIDE-003) |
Step by step
The nine failure reasons in the catalog, what typically happens, and what helps. RAND's five root causes (misunderstood problem, missing data, technology before problem, weak infrastructure, problems too hard for AI) appear in several of them [2].
-
unclear-goal: no clear process, owner or success metric. RAND names misunderstanding the problem as the most common root cause [2]; IW's top barrier is benefit that is hard to assess (63%) [11]. Countermeasures: a one-page charter with a named process owner, a measured baseline and 3–5 KPIs with go and stop thresholds (GUIDE-009). Start from a process problem, not from a tool (GUIDE-001). RAND also advises committing to a problem for at least a year [2]. -
data-quality: data missing, scattered or of poor quality. Named by Gartner [3], RAND [2], IW (46% lack data access) [11] and the Mittelstand-Digital experts [12]. Countermeasures: run the data readiness check before the pilot (GUIDE-003), budget data preparation as its own work package (seedata_prep_effortin each use case), and look at the PREP family for the preparation steps. If the data does not exist, the project starts with collecting data, not with AI. -
adoption: low acceptance by staff or customers. Bitkom: 42% of AI users name lack of staff acceptance [7]; IW: 62% say staff must be trained first [11]. Countermeasures: involve users and at least one sceptic from the start, use AI champions, communicate honestly about effects on jobs (GUIDE-008), train staff (GUIDE-006, also a duty under Art. 4 AI Act), and measure actual use, not licences. -
accuracy: the model is not accurate enough for the task. 34% of German companies say AI does not meet quality expectations [11]; RAND warns that some problems are too hard for AI [2]. Countermeasures: test on your own test set of 50–200 real cases including hard ones (GUIDE-009), keep a human in the loop where errors have external effects, and choose tasks where a good draft is already useful. Monitor quality after go-live, because models and data drift (GUIDE-011). -
integration: integration into existing systems is harder than expected. MIT attributes much of the gap to brittle workflows and tools that do not fit daily operations [1]; IW: 50% find adapting AI to their needs laborious [11]. Countermeasures: prefer AI built into the systems you already use (seeintegrationsin each use case: DATEV, Lexware, M365, Business Central and others), check APIs and interfaces in the pilot, and include integration effort in the production cost estimate (GUIDE-002, GUIDE-009). -
cost-overrun: running or scaling costs are higher than the benefit. Gartner names escalating costs in both forecasts [3][4]; 54% of Bitkom AI users say costs are hard to calculate [7]. Countermeasures: calculate cost per case at full volume, not at pilot volume; set usage limits and alerts for API spending; compare against measured savings, not vendor claims; and check funding (GUIDE-012). -
compliance: legal, data protection or works council issues block rollout. Data protection is the most frequently named barrier among German AI users (66%), followed by legal uncertainty (56%) [7]. Countermeasures: clarify AVV, legal basis and DPIA with the DPO before real data flows (GUIDE-005); involve the works council early, since a system suited to monitoring staff triggers co-determination under §87(1) no. 6 BetrVG, also in a pilot (GUIDE-004); check the AI Act risk class (GUIDE-014). -
skills: missing in-house skills to run and maintain it. The top barrier in the Mittelstand-Digital expert survey [12]; 46% lack IT specialists [11]. Countermeasures: match the implementation path to your skills: buy (path A) if you have no IT team, self-host (path C) only if someone can run it (GUIDE-002). Name an internal owner even for bought tools, and use the free Mittelstand-Digital offers and funded consulting (GUIDE-012). -
vendor-change: the vendor changes the product, pricing or terms. Fast-moving market, agent washing [4], models withdrawn or repriced. Countermeasures: keep prompts, test sets and source data on your side, prefer contracts with notice periods for price and model changes, and plan your exit (see the next section and GUIDE-010).
Cost and vendor lock-in: how to read the catalog
Cost bands. Each use case gives an estimate per implementation path (A buy, B cloud API, C self-host, D custom) for a typical SMB deployment with about 20–50 users. These are bands, not quotes, and your own staff time is not included, although it is often the largest cost item.
| Setup (one-off) | Running (per month) | ||
|---|---|---|---|
| S0 | < 1,000 € | R0 | < 100 € |
| S1 | 1,000–10,000 € | R1 | 100–500 € |
| S2 | 10,000–50,000 € | R2 | 500–2,000 € |
| S3 | 50,000–250,000 € | R3 | 2,000–10,000 € |
| S4 | > 250,000 € | R4 | > 10,000 € |
Read them like this: compare the running band × 36 months plus setup with the measured savings over three years. A use case with S1/R2 costs roughly 20,000–80,000 € over three years before staff time. Path C often moves cost from running to setup (hardware); path A the other way round.
Data flow (paths A and B):
| Value | Meaning | What to check |
|---|---|---|
on-prem |
Stays on your own hardware | Your own security and operation |
eu-provider |
EU-headquartered provider, processing in the EU | AVV, subprocessors |
us-eu-region |
Mostly US provider, processed in an EU region | The US CLOUD Act still applies; AVV, transfer clauses, DPIA (GUIDE-005) |
third-country |
Typically processed outside the EU | Transfer mechanism (DPF, SCC); often unsuitable for personal or confidential data |
Lock-in (on the recommended SMB path):
| Level | Meaning | Typical exit effort |
|---|---|---|
| low | Data, prompts and models are portable | Days to weeks |
| medium | Some proprietary configuration or training data | Weeks to months |
| high | Deep integration, proprietary models or formats | A new project |
Exit strategy, before you sign: - Data export: contractually guaranteed export of your data, configurations, prompts, custom assistants and logs, in documented formats. - Open formats: keep source documents, test sets and labelled data in open formats (CSV, JSON, PDF/A, Markdown) on your own storage, not only inside the tool. - Swappable models: for paths B and C, put an OpenAI-compatible gateway between your application and the model so the model can be changed without rebuilding. Re-run your test set whenever a model changes. - Contract clauses: notice periods for price and model changes, deletion confirmation, and switching support. Under the EU Data Act, cloud and SaaS customers (including existing contracts) can switch provider: at most two months' preparation, a 30-day transition (up to seven months if technically complex), data retrievable for at least 30 days, and no switching charges from 12 January 2027 [13]. Details and contract points: GUIDE-010.
Typical risk by family
Qualitative assessment for a typical SMB use; individual use cases can differ.
| Family | Typical risk | Main reasons |
|---|---|---|
| LANG – Language & Text | low–medium | accuracy (wrong facts in drafts), adoption; higher for HR texts (compliance) |
| KNOW – Knowledge, Search & RAG | medium | data-quality (outdated, duplicate documents, access rights), accuracy |
| DOC – Document Intelligence | low–medium | integration (ERP, DATEV); data capture is productised, complex document review is harder |
| CONV – Conversational AI | medium–high | accuracy towards customers, adoption, integration with back-end systems |
| AGENT – Agents & Process Automation | high | accuracy over multi-step tasks, integration, cost-overrun, unclear-goal [4] |
| CODE – Software Engineering & IT Ops | low–medium | skills to review output, compliance (code and secrets leaving the company) |
| VIS – Computer Vision | medium | data-quality (labelled defect images, lighting), integration into the line |
| SPEECH – Speech & Audio | low–medium | accuracy (dialect, noise), compliance (recording consent, works council) |
| GEN – Generative Media & Design | low–medium | compliance (copyright, labelling), vendor-change |
| PRED – Predictive Analytics | medium–high | data-quality (history, sensors), unclear-goal, skills |
| ANOM – Anomaly, Fraud & Risk | medium | data-quality, false alarms reduce adoption |
| REC – Personalisation & Recommendation | medium | too little data or traffic in small shops, cost-overrun |
| OPT – Optimisation & Planning | medium–high | integration (ERP, MES), master data quality, planners' acceptance |
| ROBO – Robotics & Physical AI | high | integration, cost-overrun, skills, safety requirements |
| SCI – Science, Engineering & R&D | high | skills, data, long time to value |
| HLTH – Healthcare & Clinical | high | compliance (medical device rules, health data), integration |
| SEC – Cybersecurity | low–medium | skills, vendor-change; mostly bought as part of security products |
| DATA – Data & Analytics | medium | accuracy (wrong figures from generated queries), data-quality, skills |
| EDU – Education & Training | low–medium | adoption; compliance where AI evaluates learners (high-risk under the AI Act) |
| PERS – Personal & Everyday AI | low | vendor-change, data protection with consumer apps |
| AUTO – Everyday Automation & Micro-Tools | low | skills to maintain scripts and flows, shadow AI (compliance) |
| PREP – Data Readiness for AI | medium | unclear-goal (data clean-up without a use case), underestimated effort |
Checklist
- [ ] Failure risk and failure reasons of the chosen use case read and discussed
- [ ] First project chosen with low or medium failure risk
- [ ] Process owner, baseline and KPIs with go and stop thresholds defined
- [ ] Data readiness checked; data preparation budgeted
- [ ] Users, DPO and works council involved before real data flows
- [ ] Accuracy tested on your own test set
- [ ] Integration and full-volume running costs calculated, including staff time
- [ ] Data flow and lock-in level accepted consciously
- [ ] Exit plan: export, open formats, swappable model, contract clauses
- [ ] Decision date set; stopping is an accepted outcome
Templates & tools
- Failure-reason pre-mortem (30 minutes): before the pilot, ask "it is one year later and the project has failed: why?" Map each answer to one of the nine reasons and assign a countermeasure and an owner.
- Three-year cost sheet: setup band, running band × 36, internal staff days × daily rate, against measured savings.
- RAND report with five principles for success [2].
- IW-Trends 2/2025 with the full list of barriers by company size [11].
- Mittelstand-Digital Zentren: free, vendor-neutral support for finding use cases and first implementations (see GUIDE-009 and GUIDE-012).
- Bundesnetzagentur Data Act fact sheet on switching cloud providers [13].
Common pitfalls
- Taking the 95% at face value. It measures P&L impact in large companies within a short period. Use it as a warning, not as a forecast for your invoice-processing project.
- Choosing the hardest use case first. An agent or a custom model as the first project combines several failure reasons. Start with a productised use case and build skills.
- Technology before problem. "We need an AI agent" is not a goal. RAND lists this as one of the most frequent paths to failure [2].
- Counting licence costs only. Staff time, data preparation and integration usually cost more.
- Believing agent labels. Ask what the product does on its own, with which tools and which approvals [4].
- No exit plan. Custom assistants, RAG indexes and fine-tuning data are often not exportable. Keep the source material on your side.
- No stop decision. A project that never ends is the most expensive kind of failure.
Sources
- Fortune: "MIT report: 95% of generative AI pilots at companies are failing", 18 August 2025 (on MIT NANDA, "The GenAI Divide: State of AI in Business 2025"). https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- Ryseff, De Bruhl, Newberry: "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed", RAND, RR-A2680-1, 13 August 2024. https://www.rand.org/pubs/research_reports/RRA2680-1.html
- Intelligent CIO: "Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025", 5 August 2024 (on the Gartner press release of 29 July 2024). https://www.intelligentcio.com/eu/2024/08/05/gartner-predicts-30-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025/
- Silicon.eu: "Gartner: Over 40 per cent of agentic AI projects will be discontinued by the end of 2027" (on the Gartner press release of 25 June 2025). https://www.silicon.eu/gartner-over-40-per-cent-of-agentic-ai-projects-will-be-discontinued-by-the-end-of-2027-18146.html
- BCG: "The Widening AI Value Gap: Build for the Future 2025", 30 September 2025. https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap
- IT Brief UK: "McKinsey report shows AI interest but slow scaling" (on McKinsey, "The state of AI in 2025: Agents, innovation, and transformation"). https://itbrief.co.uk/story/mckinsey-report-shows-ai-interest-but-slow-scaling
- Bitkom: "Erstmals nutzt die Mehrheit der Unternehmen KI", press release, 14 September 2026 (n = 603 companies with 20+ employees). https://www.bitkom.org/Presse/Presseinformation/Erstmals-nutzt-Mehrheit-Unternehmen-KI
- ifo Institut: "Mehr als die Hälfte der Unternehmen nutzt Künstliche Intelligenz", 5 June 2026. https://www.ifo.de/fakten/2026-06-05/mehr-als-die-haelfte-der-unternehmen-nutzt-kuenstliche-intelligenz
- KfW Research: press release of 11 February 2026 on AI use in the Mittelstand (KfW-Mittelstandspanel). https://www.kfw.de/%C3%9Cber-die-KfW/Newsroom/Aktuelles/Pressemitteilungen-Details_880896.html
- Zimmermann, V.: "Künstliche Intelligenz im Mittelstand: einfach zu handhabende Technologien mit breiten Einsatzgebieten überwiegen", KfW Research Fokus Volkswirtschaft Nr. 554, 29 July 2026. https://www.kfw.de/PDF/Download-Center/Konzernthemen/Research/PDF-Dokumente-Fokus-Volkswirtschaft/Fokus-2026/Fokus-Nr.-554-Juli-2026-KI_Mittelstand.pdf
- Scheufen, Engels, Schmitz: "Ziele, Nutzen-Kosten-Verhältnis und Hemmnisse Künstlicher Intelligenz in der deutschen Wirtschaft", IW-Trends 2/2025 (IW-Zukunftspanel, wave 49, 2024). https://www.iwkoeln.de/fileadmin/user_upload/Studien/IW-Trends/PDF/2025/IW-Trends_2025-02-04_Scheufen_et_al.pdf
- WIK-Consult: "KI-Kurzstudie" im Rahmen von Mittelstand-Digital, Kurzfassung (survey of about 60 experts from the Mittelstand-Digital Zentren). https://wik.org/fileadmin/user_upload/Unternehmen/Veroeffentlichungen/Kurzstudien/2024/WIK_Kurzfassung_KI-Studie.pdf
- Bundesnetzagentur: Data Act fact sheet "Switching between data processing services". https://www.bundesnetzagentur.de/EN/Areas/Digitalisation/DataAct/DL/Factsheet4.html
Accessed 2026-11.