“Where do we start with AI?” is no longer a question about the future for mid-sized companies. The MIT NANDA report “The GenAI Divide: State of AI in Business 2025” (July 2025) evaluated more than 300 publicly documented AI initiatives, alongside 52 structured interviews and a survey of 153 executives. The finding is sobering: despite investments of 30 to 40 billion US dollars, 95 per cent of companies generate no measurable return from their generative AI projects. Only 5 per cent of integrated pilot projects actually unlock economic value.
The reason is rarely the technology. It is the absence of a roadmap. This article describes five steps that give the start a structure – from an honest baseline assessment through to scaling – and names the pitfalls that turn a promising pilot project into what the industry calls “pilot purgatory”: a project that works technically but never goes into production.
1. Why getting started with AI so often fails
According to the MIT NANDA report, the central bottleneck is not regulation, model quality or missing infrastructure. Most of the AI applications in use retain no feedback, do not adapt to the context of the company and do not improve over time. A pilot project runs technically – but nobody has decided in advance how success will be measured, or when it moves into productive operation. That is exactly what the term “pilot purgatory” describes: budget stays tied up, confidence in the leadership team drops, and new proofs of concept are approved without the old ones ever being closed out.
In short: success or failure is not decided by the model, but by the process around it – selection, a fixed time frame, measurement of success and clear ownership. Those are precisely the four points the following roadmap covers.
2. Step 1: Baseline assessment – where does your company really stand?
The use of AI in German companies has more than doubled within a year. The Bitkom study 2026 – a telephone survey (CATI) of 604 companies with 20 or more employees, conducted in early 2026 – puts the share actively using AI at 41 per cent; in 2024 it was 17 per cent. A further 48 per cent are planning or discussing adoption, while 11 per cent reject it.
Source: Bitkom study 2026, telephone survey of 604 companies with 20 or more employees.
The adoption rate on its own says little about the quality of the start, however. According to Bitkom, only 21 per cent of companies have a formal AI strategy, 43 per cent offer their employees no AI training at all, and 33 per cent report that adoption turned out more expensive than planned – at 19 per cent it has already led to job cuts. An honest baseline assessment therefore does not begin with the question “which tool?”, but with: which processes tie up the most time today? Where does the data sit that an AI system would need? Who in the business would own a pilot project? The free AI Readiness Check maps exactly these five dimensions in 12 questions and gives you an initial score plus a roadmap in two minutes.
3. Step 2: Choose one use case, not a wish list
The most common mistake after the baseline assessment is to kick off several use cases at once, because different departments have different wishes. Without shared prioritisation – and, according to Bitkom, only 21 per cent of companies have a formal AI strategy that would deliver it – you get exactly the fragmentation that nobody can evaluate later. A single use case works better, selected against three criteria:
- A tangible pain point: the process demonstrably costs time or money – not just as a gut feeling, but measured by a metric you can compare before and after.
- An available data basis: the information the system needs already exists – in a spreadsheet, a CRM or a shared drive. A use case whose data has to be laboriously assembled first is not a good first case.
- One owner: a single person carries the decision on the pilot – not a committee. How a process like this then connects with your website, CRM and automation to form one system is something we cover in our article on the AI marketing stack for SMEs.
4. Step 3: A time-boxed pilot project instead of an open-ended trial
A pilot project needs an end date – otherwise it becomes an open-ended trial that at some point nobody notices any more. In practice, a window of eight to twelve weeks works well: long enough to see robust results, short enough to spot wrong turns early. Before the start, two or three success criteria are put in writing – time saved per case, error rate or turnaround time, for example – along with a fixed date for the go/no-go decision.
A pilot without an end date is no longer a pilot – it is an open-ended trial that sooner or later nobody owns.
This discipline is the real difference between the 5 per cent of successful pilots in the MIT NANDA report and the other 95 per cent: not the model in use, but whether somebody decides at the end to close the pilot, scale it or stop it.
5. Step 4: Sort out data, processes and responsibilities
Before a pilot project moves into rollout, a short and honest check of the foundations pays off: is the data the system works with current and complete enough? Is the process that AI is meant to support documented at all – or does it exist only in one person’s head? And is it clear who reviews a decision made by the system before it has any effect outside the company?
- Data quality before data volume: a clean, small data set delivers better results than a large, unstructured one.
- Document the process: whoever handles the task manually today should write the sequence down once – which doubles as the best possible briefing for the AI system.
- Define sign-off: a named person reviews the output for as long as the pilot runs. That responsibility does not disappear later; it simply changes as confidence in the system grows.
6. Step 5: Train, assign ownership, scale
A successful pilot does not run itself from there. The Bitkom study 2026 shows that companies applying AI deliberately to specific problems report a stronger competitive position after one to two years in 77 per cent of cases. At the same time, 43 per cent of companies offer their employees no AI training whatsoever – a key reason why 33 per cent found adoption more expensive than planned.
Before you scale: decide who on the team will own day-to-day operation from now on, how new colleagues will be brought up to speed, and how often the output will be reviewed. Only once those three points are settled does a successful pilot become a dependable part of everyday business.
7. The three most common pitfalls in mid-sized companies
- Too many pilots without prioritisation: without a shared strategy – which, according to Bitkom, only 21 per cent of companies have – parallel silo solutions emerge that nobody consolidates.
- No end date, no measurement of success: this is precisely what “pilot purgatory” in the MIT NANDA report describes: projects that work technically but are never evaluated and never closed out.
- Rollout without training: according to Bitkom, 43 per cent of companies do not train their employees – a main reason why AI adoption turned out more expensive than planned at 33 per cent of companies and led to job cuts at 19 per cent.
The roadmap in practice
Baseline assessment, one use case, a time-boxed pilot, clean data, then training and scaling: five steps are enough to turn an AI experiment into a dependable result. Where your company stands today is something the free AI Readiness Check shows you in a few minutes. And if you would then like to set up your first pilot together with us: a paid initial analysis, or get in touch directly via contact.
8. FAQ: common questions about getting started with AI in SMEs
As a mid-sized company, how do I best get started with AI?
With an honest baseline assessment rather than with a tool. First establish where your company stands today, then pick a single use case with a genuine pain point and a sufficient data basis, and test it as a time-boxed pilot project with success criteria agreed in advance. The Bitkom study 2026 shows why this matters: only 21 per cent of German companies have a formal AI strategy at all – the rest are experimenting without coordination.
How many AI use cases should I start at the same time?
One. Several parallel pilots without shared prioritisation are one of the main reasons why AI projects in mid-sized companies never make the leap into productive operation. It is better to run a single use case with one clear owner, a fixed time frame and a defined way of measuring success – and only then move on to the next one.
Why do so many AI pilot projects fail?
The MIT NANDA report “The GenAI Divide: State of AI in Business 2025” (July 2025) evaluated more than 300 public AI initiatives, along with interviews with 52 organisations and a survey of 153 executives. The result: 95 per cent of companies generate no measurable return from their generative AI projects, despite investments of 30 to 40 billion US dollars. According to the report, the central bottleneck is not regulation or model quality, but the fact that most applications retain no feedback and do not adapt to the company context. The industry calls the outcome “pilot purgatory”: projects that work technically but never go into production.
Do I need a strategy before getting started with AI?
Not a complete strategy, but you do need prioritisation. The Bitkom study 2026 (604 companies with 20 or more employees surveyed) shows that companies applying AI deliberately to specific problems report a stronger competitive position far more often after one to two years (77 per cent). At the same time, 43 per cent say they offer their employees no AI training whatsoever – a main reason why 33 per cent of companies found adoption more expensive than planned.
