The pilot trap: why AI in mid-sized companies stays on the surface

Many companies use AI, but mainly as a freely available tool for individual tasks – without their own data, without changed workflows and without evidence of what has improved. According to the available studies, companies that achieve measurable value with AI more often redesign their workflows, measure the impact, anchor responsibility in management and focus on a few core applications. These studies come mainly from large corporations and show correlations, not causes.

Adoption is not the problem. According to Bitkom, 57 per cent of German companies with 20 or more employees use AI, up from 20 per cent in 2024. The real question is therefore: why does this use so rarely have a noticeable effect, and what would need to change in the approach?

Broad use, little of the potential realised

The companies themselves see the gap. Of the AI users surveyed (n = 343), only 3 per cent say they exploit the potential "rather strongly"; by their own assessment, 87 per cent exploit it little or not at all (Bitkom presentation 2026, slide 9). It is striking that this figure lies between 84 and 87 per cent in all size classes. Larger budgets alone evidently do not solve the problem.

One reason lies in the type of use. Germany's Institute for Employment Research (IAB) found for 2025 that 90 per cent of establishments using generative AI use publicly available models. Only 16 per cent use models further trained on their own data. Intensive use is limited to a relatively small proportion of establishments. The fields of application fit this picture: customer contact (72 per cent) and marketing (54 per cent) are far ahead according to Bitkom, with production (22 per cent) and research and development (16 per cent) well behind. KfW accordingly recommends that mid-sized companies use AI "not just selectively but intensively".

Where tools without process get stuck

A freely available chatbot speeds up a draft. But it neither changes the workflow in which the draft is produced nor knows the data that make the difference in the company. The benefit therefore remains individual and hard to measure.

This becomes particularly clear with so-called shadow AI, i.e. the use of private AI accounts for work tasks. According to Bitkom (21 October 2025), it is widespread, occurs in individual cases or is suspected in around four in ten companies. Only 26 per cent officially provide generative AI, and 23 per cent have rules for its use. Bitkom President Ralf Wintergerst advises avoiding "AI sprawl". Shadow AI is less misconduct on the part of employees than a symptom: where the company offers no approved tool and no rules, private use arises without company data and without connection to processes. That does not produce a scalable use case.

What companies that scale do differently more often

The most robust comparisons come from international surveys. In McKinsey's "State of AI" 2026, around 6 per cent of respondents are "high performers" who attribute at least 5 per cent of their operating profit to AI. Almost three quarters of them have redesigned workflows, compared with about a quarter of the rest. High performers are also twice as likely to have defined processes for measuring impact and visible commitment from leadership.

BCG (2024) arrives at a related picture: 74 per cent of the companies surveyed do not yet achieve tangible value. The leading companies pursue about half as many use cases as others but expect more than twice the return. 62 per cent of their AI value comes from core functions such as operations, sales and R&D. BCG places around 70 per cent of the effort in people and processes, and only 10 per cent in algorithms.

Two caveats belong here. Both studies mainly survey large corporations and are based on self-reporting. And they show what successful companies do more often – not that this approach causes the success. For mid-sized companies, there is no comparable comparison between those that scale and those that get stuck in the pilot.

What should be settled before the next pilot

From these patterns, test questions can be derived that should be clarified before the next budget step:

• Baseline measurement: How long is the lead time today, how high the error rate, how large the processing effort? Without a baseline, no improvement can be demonstrated.
• Process, not tool: Which work step disappears, who checks what in future, which company data does the solution need? These questions determine the benefit more than the choice of model.
• Scaling and termination criteria: Which effect justifies expansion, and below which threshold do you stop? Both belong before the project starts, not in hindsight.
• Responsibility: Who in the business unit is responsible for the result, and which manager decides on the transition into operation?
• Focus: Which one to three applications affect a core process? Ten parallel experiments create activity, but rarely evidence.

This list does not automatically make AI economically viable. It ensures that a pilot answers a question – even if the answer is that an improved manual process or conventional software is the better choice. If you are planning a structured feasibility assessment with fixed decision points at this stage, you can talk to capinori about it.

Frequently asked questions about the AI pilot trap in mid-sized companies

Sources (7)

• Bitkom: press release "Erstmals nutzt die Mehrheit der Unternehmen KI", 14 September 2026 (German). Website
• Bitkom Research: Künstliche Intelligenz in der Wirtschaft 2026, presentation (603 companies with 20 or more employees) (German). PDF
• Bitkom: press release "Beschäftigte nutzen vermehrt Schatten-KI", 21 October 2025 (German). Website
• IAB: Friedrich/Kagerl, "Jeder vierte Betrieb nutzt mittlerweile generative KI", IAB-Kurzbericht 8/2026, 5 May 2026 (German). PDF
• KfW Research: Fokus Volkswirtschaft No. 554, 29 July 2026 (Dr Volker Zimmermann) (German). PDF
• McKinsey & Company: The state of AI in 2026, 25 August 2026. Website
• Boston Consulting Group: Where's the Value in AI?, October 2024. PDF