Almost every Swiss company we speak to has run an AI pilot. Very few have one in production carrying real volume. The gap is not a modelling gap — the pilots usually work. It is that a pilot proves a model can produce an output, while production requires an organisation to accept that output as a decision.

Those are different problems, and only the first one is technical.

The four blockers, in the order they appear

No owner for the decision. The model recommends; someone must decide. If no named person owns the decision the model informs, the output has nowhere to go. This is the most common blocker and the least technical.

No path for the wrong answer. Every model is wrong sometimes. If there is no defined route for a human to catch, correct and feed back an error, the first visible failure ends the programme. Design the exception path before the happy path.

Data that only existed for the pilot. The pilot ran on an extract someone produced by hand. Production needs that extract on a schedule, with quality guarantees and an owner. That is a data engineering project nobody funded.

Unclear regulatory footing. Under the revised FADP, automated decisions with legal or significant effect trigger an information obligation and a right to human review. Discovering this at deployment stops the programme; designing for it costs almost nothing. We set out the detail in what Swiss companies actually have to do under the revised FADP.

How do you choose a first AI use case?

Pick one where the decision is already being made repeatedly, by a person, with a known cost of being wrong. That combination gives you a baseline, an owner, and a tolerance level — the three things a pilot usually lacks.

Good first candidates in the Swiss mid-market: invoice coding and matching, demand forecasting for a constrained product line, triage and routing of inbound requests, anomaly detection in operational telemetry. Each has an existing manual process to measure against.

Bad first candidates: anything where the value depends on a customer-facing experience being flawless, anything where the training data would have to be created from scratch, and anything whose sponsor cannot say what a wrong answer costs.

What does it cost to put an AI use case into production?

For a mid-market Swiss company, the model is rarely the largest line. In the programmes we see, the distribution is roughly: 15% modelling, 40% data engineering, 25% integration into the operational workflow, and 20% governance, monitoring and change management.

Budgets that assume the reverse are the ones that stall at 60% spent with a working notebook and nothing in production.

Build, buy, or use an API?

Three honest defaults.

Use an API when the capability is general — language, transcription, extraction, classification of common document types. Building this yourself in 2026 is almost always a mistake, and hosted inference at the edge has made the latency argument largely obsolete.

Buy when a vendor has domain-specific training data you cannot obtain. Their data advantage is the product; your version will underperform for years.

Build when the model encodes something proprietary — your pricing behaviour, your process physics, your customer graph. This is where the defensible value is, and it is a much narrower set than most roadmaps assume.

What about data leaving Switzerland?

It is a legitimate concern and a solvable one. Inference can run in-region; several providers offer Swiss or EU processing, and edge inference platforms let you keep the request path inside a chosen geography. The revised FADP does not require residency, but a supervisory expectation or a client contract may.

Decide this early, because it constrains the provider shortlist, and it is expensive to retrofit once an integration exists.

A twelve-week path to something real

  • Weeks 1–2. Name the decision, the owner, the baseline and the cost of a wrong answer. If any of the four is missing, stop here — that is the work.
  • Weeks 3–5. Build the data path first, on a schedule, with quality checks. Not the model. The data path.
  • Weeks 6–8. Model, evaluated against the human baseline on data it has not seen.
  • Weeks 9–10. Exception path, monitoring, and the human-review route the FADP expects.
  • Weeks 11–12. Shadow mode in production. The model runs, humans decide, you measure the disagreement. Ship when the disagreement rate is understood, not when it is zero.

Shadow mode is the step most often skipped and the one that most reliably prevents an embarrassing first week.

Working on this right now?

Tell us where you are in two questions. A consulting partner reviews every enquiry within 2 business days.