What does AI automationreally cost an SME?

The four cost blocks

If you want to estimate the cost of AI automation, do not start with the model price, but with the whole process. In practice, costs consist of four parts:

1. AI usage: providers charge per token processed, split into input and output.
2. Automation platform: tools such as Make, Zapier or n8n connect systems and charge a monthly fee depending on volume.
3. Set-up: analysis, rules, interfaces, tests and quality checks. This is usually the largest one-off item.
4. Operation: monitoring, error handling, adjustments when systems change and new model versions.

On top of that comes your team’s time for coordination and approvals. It drops significantly after the introduction, but does not disappear entirely if control points are built in.

What the AI itself costs

A token is a building block of text. According to OpenAI, one token in English corresponds to about four characters or three quarters of a word; for other languages the ratio differs.^[5] German texts generally need more tokens than English ones. The following table shows the list prices of some well-known models.

List prices according to the providers’ websites, retrieved on 16.09.2026.^{[1, 2, 3, 4]} According to Google, Gemini 3.8 Flash will cost twice as much from 1.1.2027; DeepSeek prices depend on the time of day. Anthropic and OpenAI offer a 50 % discount for batch processing that is not time-critical.

Example: 500 product descriptions

Assume an automation creates 500 product descriptions. Each one sends around 1,000 tokens of instructions and product data and receives around 300 tokens back. That makes 500,000 input and 150,000 output tokens. The calculation is based on our own assumptions and only indicates the order of magnitude:

Even with the most expensive model in this example, the pure AI costs stay in single-digit dollars. In practice, checks, retries and translations are added and can multiply consumption. For many SME applications, the order of magnitude is still small compared with working time.

Automation platforms compared

The platforms bill differently, so it is worth looking closely at the unit.

Prices according to the providers’ websites, retrieved on 16.09.2026.

The difference is bigger than the entry prices suggest. A workflow with ten steps uses ten credits on Make, while n8n counts it as one execution. Self-hosted solutions, on the other hand, come with server and maintenance costs.

The bigger item: working time

The median wage in Switzerland was CHF 7,024 gross per month for a full-time position in 2024.^[9] For the cost of one working hour, labour cost statistics are more meaningful: for 2020 they show an average of CHF 63.62 per hour worked.^[10]

A simple estimate: if a manual process takes five hours a week, that adds up to around 260 hours a year. At 2020 labour costs, that equals around CHF 16,500. Today’s labour costs are likely to be higher. This figure is the benchmark against which the set-up and operation of an automation can be measured.

What studies say about the benefits

The best-known studies show clear effects, but not everywhere. In a study of 5,179 customer support agents, teams with AI assistance resolved 14 % more cases per hour on average; for new agents the figure was 34 %, while experienced agents saw hardly any difference.^[11]

A study of 758 Boston Consulting Group consultants shows the limits: for tasks that AI handles well, participants completed 12.2 % more tasks, 25.1 % faster and with more than 40 % higher quality. For tasks outside this frontier, however, the probability of a correct solution was 19 percentage points lower.^[12] The McKinsey Global Institute estimates the potential of generative AI at USD 2.6 to 4.4 trillion per year; this is an estimate, not a measured effect.^[13]

Where Swiss SMEs stand today

According to AXA’s 2026 SME labour market study, 74 % of Swiss SMEs use or test artificial intelligence: 35 % have integrated it and 39 % are experimenting. AI is used most often for translations (47 %), correspondence (42 %), advertising copy and process optimisation (35 % each).^[14] A year earlier, AXA found that only 33 % of SMEs had clear data protection rules for AI, and 23 % of micro-enterprises.^[15]

A survey by HWZ and Swisscom names obstacles including the shortage of skilled staff, regulatory uncertainty, doubts about accuracy and an unclear return on investment.^[16] These are exactly the points a cost calculation should answer.

Data protection and regulation belong in the calculation

The Federal Data Protection and Information Commissioner (FDPIC) states that the current Data Protection Act applies directly to AI applications. This includes transparency about purpose, functionality and data sources, and the possibility of a human review of automated decisions.^[17] Relevant provisions of the FADP include Art. 9 on processing by processors, Art. 16 f. on disclosure abroad, Art. 21 on automated individual decisions and Art. 22 on data protection impact assessments.^[18]

For US providers, the Swiss-U.S. Data Privacy Framework has applied since 15.09.2024, but only to certified companies.^[19] If you have customers in the EU, keep an eye on the EU AI Act: the transparency obligations under Art. 50 have applied since 2.8.2026, and the deadlines for high-risk systems have been postponed.^[20] In Switzerland, the Federal Council intends to ratify the Council of Europe AI Convention and to regulate AI by sector.^[23]

In practice, this means reviewing contracts with AI providers, documenting data flows and processing personal data only where necessary. This work is part of the set-up and should appear in the cost calculation.

Funding for getting started

Support is available for preliminary studies. The Innosuisse innovation cheque covers up to CHF 15,000 for a preliminary study with a Swiss research partner, for SMEs with fewer than 250 full-time employees. Pure implementation or digitalisation projects are not covered.^[21] Companies based in the Canton of Zurich can also join the free AI innovation programme of the Office for Economy, which supports them from introduction to prototype.^[22]

How to calculate your own case

1. Choose a process: a specific process with a clear volume, for example invoices, enquiries or product data.
2. Measure time: minutes per case multiplied by cases per month, valued at your labour costs.
3. Add error costs: corrections, follow-up questions, returns or missed deadlines.
4. Estimate running costs: tokens, platform, hosting and support per month.
5. Offset the set-up: one-off costs divided by monthly savings gives the payback period.
6. Start small: begin with one process, measure results, then expand.

Frequently asked questions