AI applications for SMEs: assistants and chatbots

An AI application from Fox Computers starts from CHF 12,000, including concept, interface, model integration, testing and quality control. The price depends on how many data sources are connected, how many people use the application and whether it runs locally or in the cloud. Before we start, you get a quote with a fixed price for the agreed scope. If you want to check first whether the approach works, you can start with a prototype. For operation, there are later costs for hosting and model usage, which we estimate for you in advance.

At Fox Computers, an AI application takes from six weeks to the first production version, depending on scope and data sources. Early in the project there is a prototype that you try out with real cases from your daily work. That quickly shows whether the approach works before much budget is spent. Then we build out the interface, integrations and quality control and test thoroughly. Larger applications with several systems or many user roles take correspondingly longer. After launch, we fine-tune based on real use instead of planning on paper for months.

AI applications are useful wherever your team searches, reads, summarises or pre-sorts enquiries every day. Typical examples are a search assistant for internal knowledge that answers with a source, a chatbot on the website that answers questions from your content, e-mail pre-sorting by request and urgency, or a quote helper that creates a draft from earlier jobs. Reading invoices or analysing client feedback is part of it too. If the AI should work as a dedicated assistant in Teams, Slack, by e-mail or WhatsApp, the AI employee from Fox Computers is a better fit.

Depending on the task, Fox Computers uses cloud models from OpenAI, Anthropic (Claude) or Mistral, or open models such as Qwen locally via Ollama. The choice depends on quality, speed, cost and data protection. Mistral is a European provider, local models keep the data entirely in-house, and the big US models are often the strongest at complex tasks. We build every application so that the model can be swapped. If a better or cheaper model comes onto the market, you switch without redevelopment. That way your application stays independent of any single provider.

For important decisions, a human has the final say: Fox Computers builds AI applications so that the AI prepares and your team checks and approves. This principle is called human in the loop. A quote, a reply to a client or a booking only goes out once someone has confirmed it. The application flags uncertain cases and forwards them to the person responsible. Every output is logged, so errors can be traced and fixed in a targeted way. You decide how much the AI may handle on its own and extend it once the quality has been right for weeks.

We ground the AI in your real content: it answers from your documents and data and cites the source. Technically, this works through an upstream search, RAG for short (retrieval-augmented generation). If the application finds no matching information, it says so instead of guessing. On top of that come clear limits on which topics it may answer at all. Before launch, Fox Computers tests with a set of typical questions and known correct answers, and repeats these tests after every model change. Errors can never be ruled out entirely, which is why human review is a fixed part of sensitive tasks.

Yes, Fox Computers connects AI applications via interfaces to CRM systems, accounting software such as Abacus or Bexio, Microsoft 365 and Google Workspace. That way, the application reads client data, documents or e-mails right where they are and writes results back, for example a draft quote into the CRM or extracted invoice items into the accounting. Websites, WooCommerce shops, ticketing systems and your own databases can be connected too. We set permissions so that the AI only accesses what it needs for its task. Whether an interface exists, we clarify before the quote.

Fox Computers trains your team during rollout, briefly and hands-on, directly in the application and with cases from your business. Because we build the interface to be understandable without lengthy explanation, a short introduction is usually enough. We show what the AI can do, where its limits are and when a human needs to check. There is also written documentation for reference. If you wish, there is additional training in prompting with ChatGPT, Claude and the like, so your team can use AI safely outside the application too. We bill such training at CHF 160 per hour.

Yes, you decide which data the AI application uses, who has access and how it behaves, and if you wish, everything runs on your own server. Fox Computers plans every application in line with the revFADP and, where necessary, the GDPR. With cloud models, we use API access where, according to the provider, business data is not used for training, and we check where the data is processed. For especially sensitive content, the model runs locally or with a Swiss host. The application’s rules, sources and limits are documented and can be adjusted at any time.

Running costs consist of hosting, model usage and maintenance, and Fox Computers estimates them for your case before launch. Cloud models such as those from OpenAI or Anthropic are billed by volume. For an internal application with moderate use, this is often a single- or double-digit amount in francs per month; with many thousands of requests, correspondingly more. Local models have no fees per request, but there are costs for hardware or Swiss hosting. We bill monitoring, updates and adjustments at CHF 160 per hour or agree on fixed support. We keep an eye on costs at all times.