Starting an AI pilot is easy. Getting it into everyday use is where most of the effort, and most of the failure, lies. Larger organisations have documented this problem extensively, and the lessons apply even more strongly to small firms, which have less room for projects that drift.

The pattern is well documented

In June 2025, the research firm Gartner predicted that more than 40% of so-called agentic AI projects would be cancelled by the end of 2027. The reasons it gave were rising costs, unclear business value and weak risk controls. It also warned of agent washing, in which existing chatbots and automation tools are relabelled as something more advanced, and estimated that only around 130 of the thousands of vendors in this space were genuine. Its analyst Anushree Verma put the underlying problem simply: "Many use cases positioned as agentic today don't require agentic implementations."

Swiss data show a similar gap between trying and embedding. An EY survey of 604 people working in Swiss companies, summarised on the federal government's SME portal, found 55% of companies had deployed AI strategically in at least one business area, 31% were still piloting and 14% had no concrete initiative. When UBS surveyed around 2,500 Swiss companies, its economists observed, as reported by Swissinfo, that many were using AI but few were doing so systematically.

Why pilots stall

Across these sources, the same handful of causes appear.

The value was never defined. A pilot that starts with let's see what AI can do has no finish line. Without a target number, nobody can say whether it worked, so it neither ends nor scales.

The data is not ready. In the EY survey, data quality and data silos were the most cited obstacle (20%), ahead of security and data protection concerns (19%) and a lack of qualified staff (18%). An assistant cannot answer reliably from price lists that exist in three versions.

Nobody wrote the rules. The OECD's survey of SMEs found that a third or fewer of those using generative AI had trained staff, set internal guidelines or looked into legal and regulatory questions. Without rules, cautious employees avoid the tool and enthusiastic ones use it in ways management cannot see.

The skills gap was underestimated. Eurostat's 2025 data, analysed by Euronews, show a lack of technical expertise as the most cited barrier among both medium and large European firms, ahead of privacy and legal uncertainty. Cost ranked lower down the list.

The project was too ambitious. Gartner's warning about agentic systems applies broadly. Integrating autonomous tools into older systems is complex and often disrupts the workflows it was meant to improve. Many tasks are better handled by plain, rule-based automation with a person approving the result.

The Swiss data protection angle

In Switzerland, governance is not optional. The Federal Data Protection and Information Commissioner has stated that the Federal Act on Data Protection applies directly to AI-supported processing. That means being transparent about the purpose, functioning and data sources of AI processing. People interacting with a chatbot have a right to know they are dealing with a machine, and whether their input is used to train the system. High-risk processing requires a data protection impact assessment.

None of this prevents a pilot. It does mean the questions should be answered at the start, not after a customer asks.

A pre-pilot checklist

Before committing time or budget, an SME owner should be able to answer yes to each of these:

  • One workflow, written down. Can you describe the current process step by step, including who does what?
  • One target number. Do you know the current time, volume or error rate, and what improvement would make the pilot worth continuing?
  • Clean source data. Is the information the tool will rely on current, in one place, and owned by someone?
  • A human checkpoint. Is there a defined point where a person reviews output before it reaches a customer or a system of record?
  • Data boundaries. Do you know which personal data enters the tool, where it is processed and stored, and what you will tell customers?
  • A named owner. Is one person accountable for the pilot and given time to run it?
  • A fixed end date. Will you decide to scale, adjust or stop after a set number of weeks?
  • The simplest tool that works. Have you checked whether a straightforward automation does the job before considering something more complex?

If any answer is no, fix that first. It is cheaper than discovering the gap six weeks into a trial.

The bottom line

Most stalled AI projects do not fail because the technology is weak. They fail because the goal was vague, the data was messy, the rules were missing or the design was more complex than the task required. A small firm that defines one measurable workflow, cleans the relevant data, settles the data protection questions and sets an end date has already avoided the most common reasons pilots go nowhere.

Want help scoping a pilot that can actually finish? The free AI Readiness Audit identifies a workflow, the data it depends on and the numbers to measure.

Sources

  • Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025. link
  • SME Portal (SECO), "AI is becoming widespread in Swiss companies", 15 July 2026. link
  • Swissinfo, "Six out of ten Swiss companies use AI", 19 May 2026. link
  • OECD, "Generative AI and the SME Workforce", 5 November 2025. link
  • Euronews, "Why European businesses are not using AI tools", 25 May 2026. link
  • Federal Data Protection and Information Commissioner, "AI and data protection", 24 September 2025. link