AI integration and automation for your business systems

Where does AI help a business?

AI helps where people spend time reading, sorting and rewriting text: incoming emails, forms, documents, support questions, CRM notes. A language model can classify, extract and draft in seconds, and a person checks the result where it matters. The value comes from placing that step correctly inside a process, with clean data going in and validated output coming out.

That is why we treat AI integration as a software project, not a prompt-writing exercise. The model is one component. Around it sit data access, rules, a review step, logging, cost limits and the interface people actually use. Most of our work on AI features happens there. For the delivery side, see our chatbot, CRM and sales automation and API integration services.

Examples from systems we built

• Reply classification in outreach platforms. In the outreach platforms we built for a SaaS client, a reply arriving over IMAP stops the sequence immediately. A language model then classifies it, and positive replies go to the CRM or become a task for a manager. A label set by a person always wins over the AI label.
• Personalized first lines with guardrails. The same platforms generate a personalized opening line for each contact. The output is sanitized to one line with no links and falls back to a neutral sentence when validation fails.
• Provenance for every model call. In an affiliate recruitment engine, each classification or generated text is stored with the model, prompt version, token count and cost. The model never decides a score; scores come from rules over collected evidence.
• Streaming generation with a fallback provider. For an email design tool we worked on, the AI service streams generated content into the editor, enforces cost ceilings per customer and can switch to a second provider.
• AI agents as readers. PingMyUsers exposes its data through a read-only MCP server with ten tools, so AI assistants can query providers and rankings directly.

When should you not use AI?

Knowing where AI does not belong is part of the job. The MCP server mentioned above parses free-text requests with a deterministic intent parser, not a language model. The questions follow a predictable pattern, and a parser is faster, cheaper and never makes things up.

We advise against AI when:

• a clear rule or a lookup table gives the right answer every time,
• the result must be exact, such as prices, deadlines, tax or contract terms (we compute those in code),
• the volume is small enough that a person handles it in minutes a day,
• the data cannot leave your systems and no suitable hosting or contract is available,
• a decision affects customers, money or legal rights and nobody will review the output.

How do data protection rules and the EU AI Act affect AI projects?

Data protection law already applies. According to the FDPIC, the revised Swiss Federal Act on Data Protection (FADP) is directly applicable to AI-supported processing: purpose, functioning and data sources must be transparent, and people can ask for a human review of automated individual decisions. In the EU, the GDPR applies to AI processing of personal data as to any other. We plan for that from the start: we send only the data a task needs, check each provider’s terms on data retention and training use, and keep logs that show what the system did.

Switzerland has no AI act of its own yet. The Federal Council decided in February 2025 to ratify the Council of Europe’s AI Convention and plans sector-specific amendments, with a consultation draft due by the end of 2026.

The EU AI Act matters for any company whose AI systems are used in the EU, including companies based outside it. Prohibited practices have applied since February 2025, and transparency duties, such as disclosing that a user is talking to an AI, since 2 August 2026. Under the Digital Omnibus adopted in 2026, obligations for high-risk systems listed in Annex III now apply from 2 December 2027. We check which category a system falls into during discovery. This is technical guidance, not legal advice; for a binding assessment, involve your legal counsel.

How does an AI project start?

With one process and one measurable question: how many hours per week does this task take, and what would a good result look like? We collect real examples, build a small test set, try the approach on it and show you the error rate before anything touches production. If the numbers do not justify AI, we say so and propose the simpler solution. Describe your use case and we will reply within one business day.