AI in your business: from an impressive demo to something people use

Almost every company has an AI demo by now. A chatbot that answers from internal documents, an assistant that drafts proposals, automatic ticket summaries. The demo works, everyone is impressed, and three months later nobody opens it.

The problem is rarely the model. It is everything missing between the demo and production.

Why AI projects stall

Nobody knows if it’s good. The demo was tested on five questions picked by the person who built it. The first real user asks a sixth and gets a confident, wrong answer. From then on, they don’t trust it.

You can’t see where answers come from. Without sources, a correct answer and an invented one look the same. People check everything by hand, so the AI saves no time.

Cost shows up last. A long prompt sent thousands of times a day to the largest available model can end up costing more than the person it was meant to help.

What we do differently

1. The evaluation set comes before the prompt

Before writing any code, we collect real cases: old tickets, e-mailed questions, documents people search for often. For each one we agree with your team what a good answer looks like.

That set becomes the test. Every change — a new prompt, a different way of splitting documents, another model — has to beat the previous version on it, or it doesn’t ship. That’s how we know the system is improving instead of guessing.

2. Answers with sources

We use retrieval (RAG): the system first finds the relevant passages in your data, then writes the answer and points to the source. Users see where each statement comes from and can check it in a second.

When the sources aren’t enough, the system has to say “I don’t know”. A timely “I don’t know” builds more trust than ten good answers.

3. A person approves anything that leaves the company

The AI drafts, a person sends. For customer replies, proposals or anything with consequences, we keep an approval step. As the evaluation set shows that certain kinds of request consistently go well, we can automate just those.

4. Cost and latency measured from week one

Every answer has a cost and a response time. We track both from the start, so we can pick the right model for each step: a small, fast one for classification, a large one only where it really matters.

Your data

The question we hear most is “where does our data go?” In short:

• data stays in your own cloud accounts where possible;
• we use providers and settings that don’t train models on your data;
• access is scoped per user: the AI sees no more than the person asking.

Where to start

Don’t start with “we want a chatbot”. Start with a simple question: what does our team get asked 20 times a day, and find the answer to in documents? That’s the first AI project that actually pays for itself.

If you already have a demo gathering dust, we can look at why together. Most of the time it doesn’t need to be thrown away — it needs to be evaluated.