Service 03
AI that takes over work, not a bolted-on chatbot.
Most AI features in software are a text box that answers questions. That can be useful, but it does not take over work. We build the features that do — and that you can measure.
What we build
Document extraction that reads an invoice and prepares the entry, including a judgement on whether it is confident enough to proceed without a human. Agents that carry out a real action instead of describing one. And an MCP server, so an assistant outside your product can operate your product: look up, calculate, book.
That last one is less exotic than it sounds. Pilot-Next users already plan their flight in conversation with an assistant; the booking is then the logical next step in that same conversation, not in a separate screen.
Measure it, otherwise it is a guess
An AI feature without an evaluation set is a feature nobody can tell is getting better or worse. We record a set of real cases with the correct answer attached, and it runs on every change — to the prompt, to the model, to the context.
That comes with a cost line. An extraction costing ten cents per document is fine; the same extraction rerunning on every page is not. We measure both, and we build the cheap path first: pull the text out of the PDF, and only bring in the model when that fails.
What we say no to
AI where an `if` would do. An assistant that can only talk about data the user sees faster themselves. And any feature where a mistake silently flows into money or a legal obligation — a human belongs in that loop, and that is a product decision rather than a technical one.
If an AI feature does not measurably save work, we advise against it. That is cheaper for both of us than building it and removing it again a year later.
Frequently asked questions
Which model do you use?
It depends on the task, and it is deliberately replaceable: the call sits behind one layer, so switching is a configuration change rather than a rebuild. Models get better and cheaper every quarter; you do not want to be locked to one.
Does our data go to an AI vendor?
Only what the task needs, and only if you want that. We agree up front what does and does not leave the building, and we build the cheap path first — in document extraction we pull out the text ourselves, so the model is often not needed at all.
What is an MCP server?
A standard way for an AI assistant to use tools from your product: look something up, calculate, carry out an action. For the user it means operating your product from the conversation they are already in. Pilot-Next has one, so we have one in production.
Let's talk
What are we building?
One conversation is enough to know whether we fit. Tell us what you have in mind — we will tell you how fast it can happen, and whether we are the right people for it.