AI Implementation in an SME: One Workflow First

Short answer

• Where to start: with one workflow that costs your team hours every week. You do not need a licence for everyone or a strategy paper to begin.
• What to measure first: how often the workflow comes up, how long one case takes and how many errors happen today. Without these numbers you have nothing to compare against later. One study found that AI helped with some tasks and made the result worse on others. The participants in another believed AI had saved them time when it had slowed them down.
• What you need from day one: a person who checks the results and a list of the errors they find.
• Which rules apply: according to the Federal Data Protection and Information Commissioner (FDPIC), the Swiss Federal Act on Data Protection (FADP) applies directly to AI. Switzerland has no AI-specific law. The Federal Council wants to adapt existing law where needed, with a bill to be drafted by the end of 2026. It supports Switzerland joining the Council of Europe's AI convention and plans a consultation on that in 2027, before parliament decides.
• How to stay independent: make sure the accounts, data, code and documentation belong to you. That way you can switch AI vendors, and the workflow keeps running without the consultant.

The fear of becoming dependent

In June 2026 the Zurich daily Tages-Anzeiger ran the headline "AI costs for companies are exploding, while the measurable benefit often stays small". In the comments under it, a reader quoted the article's line that the cheap subscriptions had been "classic offers to build a market" and added: "a less polite word for it would simply be getting people hooked, the principle is exactly the same." (Tages-Anzeiger, 10 June 2026, in German, my translation)

The commenter sees the cheap offers as a way to make companies dependent on the AI vendor. With a consultant, the fear is that the workflow ends up depending on the one person who understands how it works.

I think the dependency rarely comes from the AI itself. It usually starts on day one, when the accounts are not in your company's name and nobody measured how things ran before. Your own accounts keep you in control, and the numbers from before let you judge an AI vendor or a consultant. Settle both from the beginning and you can switch vendors and let the consultant go.

Finding the first workflow

Rank recurring work by hours

Before you pick a tool, write down for one week which work repeats, how often it comes up and how long it takes. Typical entries are following up on quotes, typing in delivery notes, answering the same customer questions and putting together reports.

Then sort by hours per week. At the top you get what costs the most time in total. That is not always what annoys people most.

What makes a good first workflow

From the top of the list, take a workflow that comes up often and where the cases look alike. You should be able to check its output and catch an error before it reaches a customer.

A workflow like that shows you within a few weeks whether AI works for you. With ten workflows at once, most of your time goes into coordination.

What to measure before you start

The Swiss government's SME portal advises measuring the impact once AI is in place (kmu.admin.ch, in German). I also measure before AI is in place. To begin with, record four values, once today and once after four weeks with AI:

The most important row is minutes per case. It includes the time of the person who checks.

I would rather measure minutes per case than estimate them. The research organisation METR ran a randomised study in 2025 with 16 experienced developers, who completed 246 tasks, some with AI tools and some without. The developers expected before each task that AI would make them about 24% faster, and after the study they estimated it had made them 20% faster. They were actually 19% slower (Becker et al., arXiv 2507.09089).

You count errors because AI makes the results worse on some tasks. Dell'Acqua and colleagues published an experiment with 758 knowledge workers as a Harvard Business School working paper. About half of them got 18 tasks that suit AI. On those, the participants with AI completed 12.2% more tasks and were 25.1% faster. The others got one task that does not suit AI, and with AI they were 19 percentage points less likely to reach the right answer than the group without it (Dell'Acqua et al., HBS Working Paper 24-013, 2023).

Brynjolfsson, Li and Raymond analysed data from 5,179 customer support agents. Those with access to an AI assistant resolved 14% more issues per hour on average, new and less skilled agents 34% more, experienced and highly skilled agents hardly any more (Brynjolfsson, Li, Raymond, NBER Working Paper 31161, 2023). So count cases per person, or the average hides who the AI helps.

Who checks the results

A small program builds a weekly report for my own online shop every Monday at 6 am. It pulls data from three sources: visitor numbers, Google Search and the shop's order and stock data. I wrote the code for it in May with AI assistance, and it first ran on its weekly schedule in October. The report itself calculates by fixed rules, without AI.

On 12 October, before the report was due to go out by email for the first time, I had every line recalculated by a second AI system that had not written the code. It ran its own queries against the same sources.

Of the report's 95 lines, 65 were right. In 20 lines the number was right, but the label, the period or the selection of data counted was wrong. Another 9 lines were simply wrong, and one line was incomplete.

The raw data was right, but the report misread it:

• A fully refunded order counted as a sale.
• The line "change from previous week" did not compare last week with the week before. It compared two periods of 480 days each that almost completely overlapped. It showed 0.9% more visits, while the real week was 1.3% below the week before.
• The report listed 9 product variants, each one colour in one size, as sold out, while 27 could not be ordered.
• It listed products that were already in the shop as gaps in the range, because it never looked them up in the catalogue.

The automated tests for the report all passed anyway. None of them checked a refund or a product where only some of the variants had no stock left. Almost all of the 30 lines that were not right went back to 9 different causes in the code the AI had helped write. Each cause has since got its own tests that would catch the old error.

The older content in my shop came from earlier AI models. On 11 October I had 27 of these articles and pages checked claim by claim against sources, again by AI systems that had not written the texts. They found 248 wrong claims and 181 more they could not support.

Several pages said, for example, that shipping was always free. In Switzerland it is free only from CHF 50.

I only let a correction go live once it had passed a second check, and a third if the second check had changed or added it. Corrections that failed a check never went live. I held back the corrections that needed a decision from me or a new translation, and also two lines for one page that no third check had seen.

The rest went live the same day in 26 of the 27 articles and pages, in German, English and French. The 27th had nothing that needed correcting.

I find an AI system works as a checker only when there are sources it can compare every number and every claim against, as in my shop. Your first workflow often has no such sources yet, and nobody knows which errors are typical. So at the start, have a person from your team who already knows the workflow check every result and keep a list of the errors they find.

That list shows you when checking a sample of cases is enough instead of every case. A result that looks plausible cannot tell you that. A passing automated test cannot either. The weekly report passed all its tests too. I had this article put through the same kind of check before it went live, with every claim checked against its source in a step separate from the writing. I describe how in AI Blog Writing: How I Check Every Claim.

When you don't need AI

Readers in Swiss comment sections keep writing that some workflows need no AI. The automatic acknowledgement my shop sends when a customer withdraws from a purchase is one of them. Behind it are a fixed template and a check for a marker word that the shop's withdrawal form puts into each message, and an AI language model would only add risk there.

For workflows like that, a script or the scheduler in the software you already use costs less. A fixed rule also decides the same way every time for the same input.

Mascha Kurpicz-Briki, professor of computer science at the Bern University of Applied Sciences, says on the government's SME portal (in German, my translation): "Sometimes the processes have to be adapted so that the data is generated systematically in a format that artificial intelligence can evaluate." If a workflow runs five different ways today, make it uniform before AI comes in.

Rules for data and tools before the team starts

The FADP is worded to be technology-neutral. That is why the FDPIC states that it applies directly to AI-supported data processing (FDPIC, AI and data protection). The FDPIC also says you must make the purpose of the processing, how it works and its data sources transparent if you process personal data with AI. Where the risks to the people concerned are high, the FADP requires a data protection impact assessment.

The Federal Council is relying on changes to existing law. On 12 February 2025 it came out in favour of Switzerland joining the Council of Europe's AI convention and of adapting laws where needed, sector by sector wherever possible. The bill for consultation is to be drafted by the end of 2026 (admin.ch). In a consultation, cantons, parties and associations comment on a draft before it goes to parliament. The Federal Council plans to open one in 2027 on Switzerland joining the convention, an international treaty on AI and human rights (its goals for 2027, in German). Parliament then decides whether Switzerland joins, and its decision can still be put to a popular vote.

For your team, a one-page internal policy is enough to start. It says which tools are allowed, which data never goes into them and who is responsible. It does not replace the data protection duties above.

How to stay independent of the AI vendor and the consultant

Answer these five questions before you start:

• Are all accounts and subscriptions in your company's name, with your payment method and your admin access?
• Can you export your data at any time in a common format, without asking the AI vendor?
• Is it written down what the workflow does, which rules apply and where it reports errors?
• Can someone on your team operate the workflow, stop it and restart it?
• Do you know what happens if you cancel your contract with the AI vendor tomorrow?

When I work as a consultant, the code, data and access always belong to my clients. I document every automation and hand it over.

If you only want to buy licences and train the team, you can do without outside help with AI implementation and get by with the AI vendor and an internal policy.

What happens with your list

With the list from your week, I go through the concrete next steps with you in a first call. After that, AI consulting covers the steps from the ranking of recurring work to a prototype on your data.

Frequently asked questions