AI made the prototype free. Not production.

AI has made building a tool almost free. It has not made what comes next any cheaper. A demo can be put together in an afternoon. Running it every day, with reliable data, in line with the Swiss Federal Act on Data Protection (FADP), and with someone able to repair it, still costs as much as before.

The result, in Switzerland: nearly two-thirds of AI pilot projects stop before they go into production, according to a study published in June 2026. SMEs, which have neither a data team nor an in-house lawyer, are the most exposed.

The same pattern comes up in the requests I receive. An AI tool has been tried, sometimes even built in-house. It half works, nobody really knows where the data goes, and the only person who understands it is overwhelmed or has left.

What the figures say in Switzerland

AI is already inside Swiss companies, but rarely at the heart of how they work. A UBS study run with the Intervista institute, covering about 2,500 companies, finds that six in ten use AI tools, but only 5% make AI a core part of their business model.

The study does something rare: it includes small organisations and tells them apart. The gap is clear. About 10% of large companies use AI in at least five of the six areas studied, against 2% of micro-businesses.

The real bottleneck is the move into production. Adesso's GenAI Impact Report 2026 estimates that nearly two-thirds of AI projects are dropped after the pilot. More than half of the companies surveyed (53%) report a revenue increase of at least 5% thanks to AI, but 17% invest with no measurable benefit and 5% cannot quantify it.

A caveat matters here. The survey rests on 100 senior executives, with no public information on the size of their companies, and it is published by a firm that sells AI projects itself. Read it as an order of magnitude, not as a measure of SMEs in French-speaking Switzerland.

The barriers people cite tell the same story. In the UBS study, data protection and wrong decisions caused by inaccurate data outweigh every expected benefit.

The 2026 Swiss Observatory of Data and AI, run with HEG Geneva, reaches the same conclusion: the hard part is no longer believing in AI, but scaling it, even as data quality declines.

The cost does not vanish. It moves.

Building has never been this fast. In the Adesso survey, software companies report an average productivity gain of about 35% from AI. A motivated employee can now assemble a small internal tool in a few hours, without calling a developer.

Building was never the most expensive part of software. As The New Stack points out, maintenance generally costs several times more than the initial build. AI has cut the price of the first step, not of the ones that follow.

That is why I see the same thing in almost every request: internal tools multiplying faster than the company's ability to understand them. When their creator changes role or leaves without a handover, they become "zombie applications": they still run, but nobody really knows how, or on which data.

Shadow AI makes it worse. In the DACH region, 92% of the security leaders surveyed believe that at least some employees use tools such as ChatGPT or Copilot without formal permission. Again, in plain terms: the survey covers companies with more than 250 employees. It would be surprising if SMEs, with fewer rules and often no internal IT department, did any better.

Five things that separate a demo from a tool in production

A demo proves that an idea can work. A tool in production has to work every day, for everyone, without surprises. Between the two are five things nobody sees in the meeting.

1. Reliable data. A demo runs on a clean sample. In production, the tool receives the real data: duplicates, empty fields, formats that change. If nobody looks after quality, the answers degrade without warning.
2. Compliance with the FADP. Switzerland does not yet have a general AI law, but the Federal Data Protection and Information Commissioner (FDPIC) points out that the FADP applies directly as soon as an AI system processes personal data. That means being transparent about the tool's purpose and how it works, and carrying out an impact assessment when the risk is high.
3. A named owner. Every tool needs a person who knows what it does, which data it uses, and who to call when it breaks. Without that person, the tool is orphaned at the first departure.
4. Maintenance and hosting. Models change, external services evolve, libraries age. An unmaintained tool does not stop all at once: it degrades, then it breaks at the worst moment.
5. Measuring the result. In the Adesso survey, 17% of companies invest with no measurable benefit. Deciding before launch which indicator will show whether the tool is actually useful (hours saved, errors avoided, shorter lead times) means you do not discover the answer a year later.

What this means for an SME in French-speaking Switzerland

For an SME, each of these five things weighs more. There is often no data team, no lawyer and no IT department: one person carries the tool on top of their real job. The UBS study confirms it. SMEs specifically cite the lack of people and technical resources as a barrier to implementation.

The figures also need an honest reading. Most of the available studies cover larger companies, or do not say how big the respondents are. For SMEs in French-speaking Switzerland with 20 to 100 employees, reliable data is missing. What follows therefore also rests on what I see in practice.

The good news is that an SME has an advantage large groups do not: management sits close to the tools. That is a decisive factor. According to Adesso, in 76% of the companies surveyed, management is actively involved in AI projects, and among those, 55% have already built AI into their essential processes. An SME owner can sort through their tools in a few weeks, where a large group takes years.

The aim is not to use less AI, but to have fewer tools, better looked after. One AI tool that runs properly is worth more than ten abandoned demos. And it is better to start now: the federal AI regulation project is due to go out for consultation by the end of 2026, and it will be simpler to respond with a set of tools that is already in order.

This article gives general pointers and is not legal advice. For a specific compliance question, speak to a lawyer who specialises in data protection.

Where to start: six questions to ask yourself

Take your AI tools and your internal tools one by one, and answer these six questions honestly. Every "no" is a piece of work.

1. Do you know how many AI tools or internal tools are used in your company, including those built by your employees?
2. For each one, is a named person responsible, and does someone else know how to take over?
3. Do you know where the data these tools process is hosted, and under which law?
4. Are your clients and your employees told when an AI system processes their data?
5. Is there minimal documentation: what the tool does, which data it uses, how to restart it?
6. Can you say, with a figure, what each tool saves you?

If you answer no to several of these questions, do not start another project just yet. Begin with an inventory: list your tools, keep the ones that really help, document them and name an owner. It is less spectacular than a new demo, but it is what turns AI into a lasting gain.

If you want to do this with an outside view, Zelif offers a one-hour diagnostic call, free and without obligation. What comes next is described in the method and in support and hosting.

Sources

• ICTjournal, En Suisse, deux tiers des projets IA pilotes restent au stade expérimental, 9 June 2026 (Adesso GenAI Impact Report 2026, 100 senior executives and 750 consumers).
• ICTjournal, L'IA progresse dans les entreprises suisses, mais sans adoption systématique, 19 May 2026 (UBS / Intervista study, about 2,500 companies).
• cmm360, KI-Studie 2026: Potenzial trifft Skalierungslücken, June 2026 (2026 Swiss Observatory of Data and AI, Colombus Consulting, Oracle and HEG Geneva).
• ICTjournal, L'IA accentue les cyberrisques dans la région DACH, 20 August 2026 (Fastly / Sapio Research survey, companies with more than 250 employees).
• FDPIC, IA et protection des données (page in French).
• ICTjournal, La nLPD est directement applicable à l'intelligence artificielle, 14 November 2023.
• Federal Chancellery, Réglementation de l'IA (page in French).
• The New Stack, Vibe Coding: The Shadow IT Problem No One Saw Coming.