Short answer
- An SME's AI strategy has five steps, from tools each person uses alone to processes that have a responsible person and need no particular laptop.
- If your key process runs only on one person's laptop, you have not reached step 4, and it is too early for an AI agent.
- The table shows your step and your next move.
An agent that depends on one laptop
Take a Treuhand office with 15 staff, a kind of Swiss firm that many SMEs hand their bookkeeping and tax returns to. The office is invented for this example. It is collecting quotes for an AI agent. An AI agent is a program that uses AI to choose the next step on its own and operates other programs to carry it out. This one would check invoices in the accounting software and ask the client for missing details.
The office's most important automated process is a macro that fetches the invoices from the mailbox and exports them into the accounting software. It runs on one accounts clerk's laptop. During her two weeks of holiday, the export waits because nobody else starts it, and the clients' invoices are booked up to two weeks late.
The invoices reach the accounting software only through this macro, so while she is away the agent would have nothing to do either. The office is on step 1 today and needs steps 2 to 5 before the agent can help.
Where you stand and what to do next
The five steps are my own wording and follow no standard. Your company is on the highest step that fits it, provided every step below fits too.
| Step | You are here when | What to do next |
|---|---|---|
| 1. Tools per person | Staff use AI tools for their own work. | Set a data rule. |
| 2. Data rule and limited access | A data rule says which data may go into which tool. An AI system with its own data access sees only what its task needs. | Build a test set for one task. |
| 3. Your own test set | Each new AI model is tested on your test set of real cases with the right answer. | Move the key process off the laptop. |
| 4. Processes off the laptop | Every important process keeps running during holidays. | Start a list of all automated processes. |
| 5. Processes with a responsible person | Each process has a responsible person, a next review date and someone for exceptions. | Put an AI agent on one of these processes. |
You can move an important macro without AI onto a server before steps 2 and 3, and it counts toward step 4 once those two steps fit.
Step 1: Tools per person
A reader commented on an article in the Zurich daily Tages-Anzeiger: "In my experience, employees are pushed into using AI without a strategy and without proper training." (Tages-Anzeiger, 10 June 2026, in German, my translation)
The clerk drafts client letters with her private AI chat subscription. A colleague summarises accounting documents in a different tool. Both save time, but the office does not know which client data goes into which tool.
The 2025 SME labour market study by the insurer AXA (in German) found that 71% of the 300 SMEs surveyed use AI. Some have it built into their work, others only try it out. Of these AI users, only a third have a clear rule on which company data staff may enter into AI tools. In AI Implementation in an SME: One Workflow First, I show how to introduce one process and measure its benefit once this rule is in place.
Step 2: Data rule and limited access
Next, the office sets its data rule. Client data goes only into an AI tool under a company contract. Even under that contract, staff who type into an AI chat replace client details with placeholders. The clerk may still use her private subscription for texts without client or company data. ChatGPT and Data Protection in Switzerland sets out which data may go into an AI tool.
The office is also planning an AI assistant for client questions. It would read information sheets, deadline lists and template letters. The client files, the office's own payroll and its staff files stay closed to it, even though they sit on the same drive.
The Swiss Data Protection Ordinance requires that authorised persons can access only the personal data they need for their tasks. I apply the same limit to an AI system.
Step 3: Your own test set
A program will sort incoming client emails into receipts, questions and deadlines and pass every email with a deadline to a person. Before the office picks a model (the actual AI behind a tool), it builds a test set of 50 real emails already handled, each with the right category next to it.
The office sets in advance how many of the 50 emails must land in the right place, say 48. If an email with a deadline does not reach a person, the model fails, however well it does on the rest.
Under a company contract, the sorting software will send each email to a model through the model provider's API, the direct connection for programs. Anthropic, for example, regularly retires older models and announces this at least 60 days ahead for publicly released models. Requests to a retired model then fail. The office tests the successor model on the same 50 emails before the retirement date. OpenAI announced on 14 October 2026 that it will remove several widely used models from its API on 14 January 2027.
Step 4: Processes off the laptop
I count a process as important if two weeks without it during a holiday would affect clients or lead to missed deadlines.
The invoice export counts toward step 4 once it moves to a server, which for an SME means a cloud service or an office computer that stays on around the clock with nobody signed in. There the export runs under an account of its own, not a person's. It starts on a schedule, for example every Monday at 7 am, or as soon as an invoice arrives in the mailbox.
The macro has to be rebuilt for this, because it depends on the clerk's personal mailbox and her office programs. The rebuilt macro reads a shared invoice mailbox.
Error messages go to a separate shared mailbox instead of to one person. Two people know how to stop the process and restart it.
Step 5: Processes with a responsible person
At every handover of the invoice export into the accounting software, an automatic check confirms that the client number and amount are there. The automated steps do most of the work, and anything unclear goes to the clerk or her deputy.
The federal minimum standard for information and communication technology (ICT) recommends an inventory of all applications and defined responsibilities for security. The office keeps a list of all its automated processes, the invoice export included. For each process, it records:
- the responsible person and the purpose
- the data it reads
- the next review date
- the person who handles exceptions
If nobody needs a process any more on its review date, switch it off.
I think the office's planned AI agent pays off only at this step. It takes over the unclear invoices and hands the clerk or her deputy its draft questions for the client. The limits from step 2 decide what it reads, and it has passed a test set like the one in step 3, made of real invoices. It is on the list with a responsible person, and the clerk and her deputy handle its exceptions. AI Agents in an SME: Uses, Limits, Guardrails covers what such an agent needs.
Sources for further reading
- Art. 3 para. 1 letter a of the Data Protection Ordinance (status as of 1 December 2025) requires data access control (Fedlex).
- The federal ICT minimum standard 2023 is a recommendation aimed mainly at operators of critical infrastructure and, in its own words, "essentially applicable to any business". It lists an inventory (ID.AM-2), security roles (ID.AM-6), managed access for people, devices and programs (PR.AC-1) and no more permissions than needed (PR.AC-4).
- Anthropic and OpenAI each keep a page on retired models.
The AXA study ran an online survey in March 2025 among 300 SMEs with at least five employees in German- and French-speaking Switzerland.
First the data rule, then the key process
In the AXA study, only a third of the SMEs that use AI have a clear rule for company data. My table puts at most that third past step 1. The most important gap I see before an agent is step 4, where processes stop depending on one person's laptop.
If you have no data rule yet, set one first. Then write down the key process that stops when a particular person is away. You can move this process to a server without a test set if it uses no AI. If it does, test its AI as in step 3 before you move it. On the AI consulting page, I explain how I help with both.
Frequently asked questions
Who is responsible when an automated process makes a mistake?
For personal data, your company stays responsible under the Swiss Federal Act on Data Protection (FADP) as long as it decides what the data is processed for and how, even when a program does the work. Internally, each process has its responsible person on the list from step 5.
Do you need your own server or an IT department to run processes off the laptop?
You can start with a cloud service or an office computer where nobody is signed in. You do not need an IT department, but you need two people who can look after the process.
Is it worth it if a person still checks every result in the end?
Yes, if checking takes less time than the work itself. The test set from step 3 shows how many cases a model gets right. If a model fails, a person checks every case, which is not enough for an agent. Buy one only for a process on the step 5 list, once a model has passed.
Should an SME train the team first or change its processes first?
Train first, and put the data rule into the training. Training does not replace the rule. Move your key process to a server after that, with a test set beforehand if it uses AI.