A sales assistant, built with their own team.

A manufacturer selling through distributors had years of sales data and no quick way to use it. We're building an assistant their sales team can ask questions, built in their own cloud with their own IT team.

What they want from it

  • See which customers drive most sales in each region
  • Spot customers who are buying less, before they leave
  • Know who pays late and by how much
  • Test what happens if a customer or product goes away

The journey so far

Five moments that shaped it.

  1. The pitch

    The first questions weren't about AI

    We showed a demo of a sales assistant. They asked where their data would go, what it would cost and why not just use ChatGPT.

    What we learned: Trust comes before features.

  2. Setup

    Everything in their own cloud

    Their IT team gave us a private space in their own cloud. Access took longer than planned. We waited instead of taking a shortcut.

    What we learned: Do it safely once, not quickly twice.

    Azure

  3. First weeks

    The data said one thing and meant another

    Years of sales history looked clean. Then came the real questions. Which date counts as the sale? What does a zero mean? The answers were in people's heads.

    What we learned: Ask what the data means, not only what's in it.

    PostgreSQL

  4. First result

    The first answer in the chat

    One analysis now runs end to end, from their data through tested code to an answer in the chat. Their sales team checks it against what they see in the field.

    What we learned: Make one thing work end to end before building the rest.

    AI Foundry

  5. Now

    Their team builds the next part

    Their IT team is building the next analysis themselves, in the same code. We review and help.

    What we learned: Hand over while you build, not at the end.

What they asked us

The questions that shaped it.

Can we test each piece before the next one?
Yes, and they were right to ask. We were building faster than their team could check. Now nothing new starts until they've tried the last piece.
Why not give the AI the whole database?
Because it would start guessing. It only gets a few approved tools that return checked numbers, so we control what it can see and say.
Where does the AI model run?
Inside their own Azure account, not in a public chatbot. They can switch to another model later without rebuilding.
How do we know the numbers are right?
Code calculates them and their team checks them against what they know. The AI only explains the result.
Can we trace an answer back?
Each answer points to the analysis it used, and to the order where the data allows. We log every question with its cost.

Have the data but no clear path?

We build AI systems with your team, in your own cloud, one tested piece at a time. Your team learns to run them, and we stay close.