Generative AI: moving from experimentation to value creation
- 1Individual time savings are not economic gains: value exists only when freed capacity is redeployed or removed from the cost base.
- 2Pick few use cases, tie each to a financial indicator and a business owner, and rank them like any other investment.
- 3Data protection (revised FADP) and the evolving European framework should be addressed from the start, with simple, proportionate rules.
- 4The real cost lies in integration, training and human review, far more than in licences.
Over the past twelve months, most leaders of SMEs and mid-sized companies have allowed some experimentation with generative AI. A few employees draft faster, an assistant has been connected to internal documentation, an impressive demonstration has been shown to the executive committee. Then the CFO asks the obvious question: what does this change in our results? The answer is often awkward, not because the technology fails, but because the experiment was never designed to answer that question.
This article offers a reading of that gap and a decision framework suited to companies of a few dozen to a few hundred people. It follows our approach to strategy and value creation: start from the expected result, then choose the means, including AI when it is the right lever.
Why pilots do not turn into value
A time saving is not an economic gain
The first misunderstanding is about how value is counted. An employee who saves twenty minutes a day generates no additional franc or euro until those minutes are converted: into higher volume, shorter lead times, a position not backfilled, overtime avoided. In an SME, where teams are small and roles broad, freed time dissolves easily into the rest of the day. It is real for the individual and invisible to the company.
Use cases chosen by the technology
The second misunderstanding concerns where ideas come from. Companies start from what the tool can do, rarely from a quantified problem. The use cases selected are those that are easy to demonstrate (summarising, rewording, translating), not those that weigh on margin, growth or risk. They are useful, but rarely sufficient to justify an organisational effort.
Diffuse ownership
A pilot without a business owner remains an IT project or a project for enthusiasts. Nobody has a target to meet, so nobody gives up a resource or changes a process. Without that decision, a pilot stays a pilot.
A risk framework addressed too late
In Switzerland, the revised Federal Act on Data Protection (FADP) has been in force since 1 September 2023. It strengthens information duties, requires processing to be documented and governs transfers of data abroad. Yet many trials have been run on consumer-grade tools with no clear rule about what may be entered. At European level, a political agreement on the AI Act was reached in December 2023 and the text still has to be finalised. Swiss companies that sell into the European Union or operate systems there will need to follow the timetable closely. We do not recommend waiting: a few simple rules now cost little and avoid having to stop, later, uses that have already taken root.
A worked example, with its assumptions
Take an industrial mid-cap of 250 people. Its sales administration has 12 people who spend about 30% of their time drafting quotes and customer replies. Assume generative AI, with human review, cuts that drafting time by a third.
- Current drafting time: 12 × 0.30 = 3.6 full-time equivalents.
- Time theoretically freed: about 1.2 full-time equivalents.
- Average fully loaded cost of a position: CHF 100,000, so a theoretical value of CHF 120,000 per year.
On the cost side, assume around CHF 5,000 for licences, and roughly CHF 40,000 in the first year for integration with existing tools, training, preparing reply templates and governance. In total, about CHF 45,000.
If the company captures all of the freed time, the first-year net result is in the order of CHF 75,000. If it captures only half, because the time dissolves, the gain falls to CHF 60,000 and the net result to about CHF 15,000. Same tool, same use, yet a difference of a factor of five depending on the management decision that accompanies it. These figures are illustrative, but the mechanism is general: value depends less on the technology than on the organisation's ability to convert freed capacity.
Four decisions to scale
1. Choose few use cases and rank them as investments
Two to four use cases are enough. Each must answer three questions: which financial or operational indicator moves, by how much, and who is accountable? They compete with the company's other projects for the same scarce resources. We describe this discipline in our use case on project portfolio rebalancing: an AI initiative has no special status, and is ranked on value, risk and the effort it requires.
2. Decide in advance what happens to the freed time
Before deployment, leadership sets how the recovered capacity will be used: absorbing growth without hiring, shortening response times, redeploying towards sales or service, not replacing natural departures. This is a managerial decision, not a technical one. It belongs in the budget and in the objectives of the team concerned.
3. Redesign the process, not just equip the person
Inserting an assistant into an unchanged process yields marginal gains. Durable gains come from revisiting the sequence: who prepares, who validates, which data feeds the step, which quality control applies. This is operations and process work before it is AI work, and it is often possible to find gains without any model at all. We readily tell clients that AI is not always the answer, and that clarifying a process is sometimes enough.
4. Set a proportionate risk framework
For an SME, a useful framework fits on one page: the categories of data that never leave the company, the approved tools, the cases where human review is mandatory, minimal traceability, a named owner. It reflects the FADP today and can be adjusted once the European framework is settled. Companies that also target the US market must deal with rules that are more fragmented, by state and by sector, which makes a clear internal baseline all the more valuable. We discuss this in our use case on preparing US market entry.
What executives can do this quarter
Take an inventory of current trials and of the tools actually in use, including those teams adopted without announcing it. Select two or three use cases anchored to an indicator and an owner. For each, estimate the full cost, not only licences, and the share of value you commit to capture. Set a review point at three months with an explicit stop criterion. A pilot that does not clear that bar is not a failure: it is information, obtained at low cost.
Generative AI is a real lever, but one among several, and it creates value only when inserted into a strategy, a process and a steering system. If you would like to assess where your trials stand and which deserve deployment, our diagnostic is a simple starting point with no commitment.
A personalised first reading in three minutes.