AI agent development for Swiss businesses

What an agent is, and what it is not

A chatbot answers questions. An agent acts. It receives a goal, such as sorting incoming supplier invoices or preparing a quote, decides which tools to call, checks its own results and stops when the job is done or when it needs a person. That extra freedom is useful and also the source of new risks.

We compare the two in AI agent vs chatbot. For answering customers on your website, see AI chatbot development for Switzerland, and for fixed workflows without model decisions see AI automation for Switzerland.

Autonomy levels and the controls that match

Most Swiss firms start at the first or second level and move up only after weeks of logged results. A higher level is earned by data, not by enthusiasm.

Privacy, decisions and human review

The revised Federal Act on Data Protection has applied since 1 September 2023 and is supervised by the FDPIC. It contains rules on automated individual decisions, which can matter when an agent decides about people, for example in hiring, credit or customer eligibility. Where a decision has legal effect or similarly significant effect on someone, plan for human review and for telling the person. Not legal advice; your lawyer should confirm what applies. If you handle EU residents' data, the GDPR can apply too.

• Keep personal data out of prompts unless the task needs it.
• Record which model, provider and region process which data.
• Check provider terms on data retention and training use before sending real records.
• Agree who can read logs, since logs contain the same data.

Safety by design

An agent connected to email or a database can be tricked by text hidden in a document or a web page, a problem known as prompt injection. We design for this: tools get the smallest permission that works, external content is treated as data and not as instructions, risky actions need approval, and every step is logged with inputs and outputs. Our AI agent security checklist lists the controls we review before launch.

We also test with realistic and hostile examples in each language, including messy scans, odd abbreviations and mixed-language emails, which are normal in Swiss offices.

Cost control is part of safety. We set spending caps with the model provider, limit the number of steps per task and alert the owner when a run loops or fails, so a faulty agent cannot run up a bill overnight.

Typical first agents

• Inbox triage: classify, label and draft replies for a shared mailbox in several languages.
• Document intake: read invoices, order confirmations or forms and prepare accounting entries for review.
• Sales support: research a lead, summarize history from the CRM and draft a first message.
• Internal helpdesk: answer staff questions from policy documents and create tickets when needed.
• Reporting: gather figures from several tools and draft a weekly summary.

We build with OpenAI and Anthropic models, workflow tools such as n8n, and custom code where needed. The choice depends on your data rules, languages and budget.

Pick the first agent by asking three questions. Does the task repeat often? Are the inputs already digital? Is a mistake easy to spot and undo? Tasks that pass all three make good pilots. A pilot that saves staff a few hours each week and leaves clear logs gives you the evidence to decide on the next step, and it gives your lawyer and auditor something concrete to review.

How we work and what it costs

We start with one process and a measurable target, build a pilot, run it beside the people who do the work today, and widen it only when results hold. We work remotely from India. Switzerland is 3.5 to 4.5 hours behind, so reviews take place in the Swiss morning, and test runs continue during the Swiss night.

Projects start at $1,500 and the rate is $30 an hour. Running costs include model usage and hosting, which we estimate in the quote. The AI agent cost guide lists the drivers, and our AI agent development service page covers deliverables.