Almost every company has now started AI projects, introduced tools or run first pilots. Expectations are high: more efficiency, better decisions, new business models.
Reality looks more sober. A widely quoted report by an initiative at the MIT Media Lab concluded in the summer of 2025 that ninety-five per cent of the organisations surveyed drew no measurable return from their AI initiatives — measured against a “marked and sustained” effect within six months. The figure is not peer reviewed and is frequently misread as a failure rate (see the correction above). From my own work I know: The reasons rarely lie in the technology. They lie in the organisation.
Three core problems that reinforce each other
Inflated expectations
AI is no cure-all. Without a clear use case, without precisely stated target figures and without measurability, you are left with big promises and small effects. How far expectation and evidence diverge shows in a McKinsey survey of 1,719 respondents across 97 countries: eight in ten say AI has improved their personal productivity — but only 37 per cent attribute any effect on earnings to AI. Both figures come from the same survey. The reason is usually the same — expectations are not synchronised with strategy and reality.
Data and process chaos
“Garbage in, garbage out” applies to AI more than ever. Data sits in silos, is incomplete, has grown historically or is simply inaccessible. How much that weighs was examined by the RAND Corporation in 2024 across 65 interviews with experienced practitioners. Data problems were named by 30 of the 50 practitioners — but the most frequent cause was a different one: 84 per cent named leadership and problem framing, that is, projects where what needed solving was misunderstood or miscommunicated from the outset. The biggest obstacle is therefore neither computing power nor the data strategy alone.
Missing enablement
Even when the use case and the data are right, AI stays ineffective if people are not brought along. This is a people topic, not an IT topic. How much decision-making authority people grant a machine depends strongly on what is at stake. The Pew Research Center surveyed just over eleven thousand US adults in 2022: 71 per cent oppose AI making the final hiring decision — for promotions it is 47 per cent, for performance reviews 39 per cent. The resistance is therefore not against AI in management as such, but against the finality of the decision. For Germany, a representative Bitkom survey from spring 2025 shows: 70 per cent of employees are offered no AI training by their employer, and only 20 per cent have been trained. Adding the 6 per cent who do not take up an existing offer, roughly three quarters have had no training.
The result: shadow IT, mistrust, uncertainty. Employees have long been reaching for AI tools privately; at work they lack orientation.
What the companies where it works do differently
They start with the business, not the tool. The question is not “which model do we take?” but “what do we want to achieve for our customers, and what can we improve internally?”
They treat data as a product. More and more organisations create roles such as AI data stewards or compliance specialists in order to manage data as a strategic asset rather than a by-product.
They combine stability with adaptability. Deloitte calls this “stagility”. Years of iterative working pay off here — making processes pragmatically more flexible without losing focus.
Three examples from practice
Automotive group: self-service instead of a central project
An international manufacturer faced hundreds of potential use cases from its business units. Instead of central projects, a platform was built on which teams could develop their own applications — from quality checks in production to automated documentation. Alongside it, clear guard rails: data access, compliance rules, governance for traceability. Within a few months more than ten solutions were running in daily operations, not as pilots.
Logistics: routes in real time
A European group wanted to make its supply chains more resilient. The existing route planning worked on fixed rules — too slow for today's dynamics. An agent was built that combines traffic, weather and customer data. The result: measurably shorter delivery times, a double-digit fall in fuel consumption. Importantly, the dispatchers kept the decision; the system only proposed options.
Mid-sized manufacturer: spotting faults before they get expensive
A machinery manufacturer wanted to reduce scrap and energy consumption. Previous systems only detected problems after the fact. With sensors and continuous analysis, scrap fell by almost twenty per cent and energy use per unit dropped. Decisively, the skilled workers on the shop floor helped determine which indicators were monitored and which thresholds were set.
All three examples have the same thing in common: AI only takes effect when organisation, data and people work together.
What leadership, HR and teams must deliver now
Leadership sets the frame instead of steering the detail. Successful organisations show that it matters less to take every decision yourself than to set the frame. Define priorities, create transparency, connect use cases to company goals — and take responsibility for data quality, compliance and ethics.
HR becomes the architect of enablement. Standardised mass training is not enough. It takes reskilling, short learning units and individual learning paths — and new role profiles that become indispensable alongside AI.
Teams become designers instead of users. Teams allowed to develop their own ideas, work with data and reshape processes secure acceptance and the capacity to innovate. Autonomy in action, embedded in a stable frame.
Fazit
AI does not run by itself. It only adds value once organisation, data and people work together. Leadership must set guard rails, HR must enable capability, teams must be allowed to shape.
AI is a change topic, not a question of tools. Everything else stays a demo.
Sources
- Ryseff, J. / De Bruhl, B. / Newberry, S.: The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation, RR-A2680-1, 2024 — 65 interviews, 50 of them with practitioners from more than 50 organisations. rand.org
- McKinsey & Company: The State of AI in 2026: On the Road to ROI, August 2026 — 1,719 respondents across 97 countries. mckinsey.com (PDF)
- Pew Research Center: AI in Hiring and Evaluating Workers: What Americans Think, 20 April 2023 — 11,004 US adults, American Trends Panel, fieldwork December 2022, margin of error ±1.4 percentage points. pewresearch.org
- Bitkom e. V.: Ein Fünftel wurde im Job zu KI geschult, press release of 7 July 2025 — Bitkom Research, telephone survey of 1,005 people aged 16 and over, including 513 in employment, spring 2025. Industry association source. bitkom.org
- Challapally, A. / Pease, C. / Raskar, R.: The GenAI Divide: State of AI in Business 2025. MIT NANDA, July 2025 — not peer reviewed; 52 interviews, 153 questionnaires. The claim is “zero return”, not “failed”.
- Deloitte Insights: Stagility: Creating stability for workers for organizations to move at speed. 2025 Global Human Capital Trends, 24 March 2025. deloitte.com
- Project experience of CALADE GmbH — anonymised examples from 2021 to 2025
Correction, 10 September 2026. The figure in the original text was reported imprecisely: the underlying report does not say that 95 per cent of pilots fail, but that 95 per cent of the organisations surveyed obtained zero return. It is also not peer reviewed and rests on 52 interviews and 153 questionnaires. Text and heading are corrected accordingly. On the same occasion three further figures were checked; two did not survive and have been replaced by the RAND study of 2024 and the figures of the Pew Research Center. The article’s observation is unaffected.
This article offers a professional assessment and does not replace legal or management advice.
