Custom AI applications that take work off your plate.

Search first, then answer (RAG)

A language model knows nothing about your company. With retrieval-augmented generation (RAG) the tool first finds the relevant passages in your documents and hands them to the model. The answer then comes from your sources, with a reference. Explained in preventing hallucinations with RAG.

A person decides where it matters

For quotes, customer emails or decisions about money, the tool makes a proposal and a colleague approves it. The model is allowed to say it does not know, instead of guessing.

Testing with real questions

Before launch we build a test set of real questions with the right answers, and measure how often the tool gets it right.

Model and data

We pick the model per task and budget, for example from Anthropic, OpenAI or Google, or via Azure OpenAI if you work in Microsoft. Your data goes through those providers’ business APIs, and we host the application itself in the EU. Also read Azure OpenAI or OpenAI: where does your data go?

Costs under control

Language models charge per use. We set limits and show what the tool costs per month, so there are no surprises.

It pays off when a task comes back often and takes a lot of time, when the information is recorded somewhere (in documents, emails or a system), and when a mistake can be checked before it does harm.

It does not pay off at low volume, for decisions that must be flawless without a check, or when your data is messy; then start by getting your data in order. Sometimes a standard AI subscription for your team is enough. Then we will say so.

What you can see from us: our advisor tool is a wizard we built ourselves, giving advice and a ballpark price in a few questions. A public case of an AI tool for a client is not on the site yet.

We price a custom AI application after a call, with a fixed price up front. What determines the price:

• the number of sources and documents, and how messy they are;
• integrations with your other systems;
• the number of users and roles;
• whether a human approval step is needed;
• requirements for security and logging.

The language model’s usage costs come on top per month and depend on how often the tool is used. We make them clear up front.