AI automation is not about “putting AI everywhere”. For an SME, the most robust approach consists of selecting a repetitive and observable task, defining precisely what the system can consult and prepare, then testing the result under supervision.
A Swiss SME should start with a frequent, unambiguous and reversible process: prioritize emails, prepare a draft quote, offer slots or identify prospects to follow up. Legal, medical, financial or contractual decisions require a reinforced framework and appropriate human intervention.
What is AI automation in an SME?
Classic automation applies predefined rules: when an event occurs, a specific action is triggered. AI-powered automation can also interpret text, summarize context, classify a request, or prepare content. It nevertheless remains dependent on the data available, the instructions and the controls put in place.
The Confederation's SME portal emphasizes that application cases must be evaluated individually and that blind confidence in the results constitutes a risk. Human supervision therefore remains a component of the system, not a failure of automation.
1. Choose a process, not a technology
The right starting point is not “we want to use a big language model”. It is rather: “this request arrives several times a week, follows a known logic and unnecessarily occupies a qualified person”.
A good first process has four characteristics
- Frequent:it comes back often enough to be observed.
- Delimited:the start, the expected result and the exceptions are identifiable.
- Reversible:an error can be corrected before producing damage.
- Measurable:we can compare the situation before and after the pilot.
Preparing drafts is often more suitable for a first pilot than independent execution. It makes it possible to measure quality without immediately granting broad power of action.
2. Define data, permissions and responsibilities
Before any connection, the company must establish what information is necessary, who can consult it and how long it is useful. The practical principle is simple: do not give access to a box, a folder or an entire CRM if a smaller scope is sufficient.
The Federal Data Protection and Transparency Officer recalls that the LPD applies directly to processing using AI and insists on transparency regarding the purpose, operation and sources of data. This does not mean that a tool is automatically “compliant” because it is sold in Switzerland: compliance also depends on use, configuration, responsibilities and contracts.
Questions to document
- What data is consulted and for what purpose?
- Who authorizes the connection and who can revoke it?
- Which actions remain obligatorily subject to validation?
- How is an error or exception reported?
- Where are applicable processing operations and subcontractors documented?
3. Build a truly testable pilot
A pilot must operate on a sufficiently real scope to produce learning, but sufficiently limited to remain controllable. For example, you can choose a single shared box, a quote template or a type of appointment.
- Describe the expected result with examples.
- Define the cases that the assistant must refuse or transmit.
- Maintain human validation on driver output.
- Note the requested corrections and their cause.
- Review the rules before expanding access.
4. Measure utility, not just speed
A quick demonstration is not proof of value. The measurement must cover a defined period and compare a starting point to an observed result.
| Indicator | Question | To avoid |
|---|---|---|
| Volume | How many actions have been prepared? | Extrapolate from a single demo |
| Quality | Which part requires significant correction? | Counting a draft as a completed task |
| Time | How much real time is saved, including validation? | Use a theoretical rate without method |
| Risk | What exceptions or errors appeared? | Hide rejected cases |
| Business | Does the process improve a useful outcome? | Confusing activity and value |
5. The most frequent errors
Automate an already confusing process
AI also amplifies inconsistencies. If no one knows when to follow up with a lead or how to calculate a quote, connecting a model doesn't solve the missing business rule.
Giving too much access upfront
A driver should apply a principle of least privilege. Access can be widened after observation; the opposite is more difficult to correct.
Measure only successful exits
Rejections, corrections and exceptions are essential to understand the real reliability of the process.
Promise total autonomy
Autonomy is not an end in itself. For some tasks, the best product is the one that prepares correctly and requires a decision at the right time.
30-day implementation plan
- Week 1:map a task, its data and its exceptions.
- Week 2:connect a limited perimeter and configure validations.
- Week 3:run the pilot, note each correction and observe the rejected cases.
- Week 4:compare the measurements, correct the rules and decide to extend, maintain or stop.
To see how this method applies to emails, quotes, appointments and reminders, visit the pageAI assistant for Swiss SMEsor test thelocal demonstration of Sania.