The mess we started with
Hundreds of spreadsheets, holding the business together. Forecasts, orders, stock, suppliers, lead times — all of it lived in spreadsheets that fed each other through fragile links. Numbers were keyed in by hand, copied from one sheet to the next, and reconciled by people who just knew which cell mattered. Every sheet had the same backbone and its own quirks: a local fix here, an exception there, a formula nobody dared touch.
It mostly worked — until a link broke, a paste went one row out, or the one person who understood a sheet was away. The real process wasn't written down anywhere. It was scattered across the formulas, hidden in the links, and carried in people's heads.
Pulling the knowledge out of the sheets
We used AI to read the spreadsheets the way the team couldn't. Not just the numbers — the process inside them. The AI traced every link to see how data really flowed, unpicked the formulas to recover the rules they encoded, and worked alongside the team to surface the implicit ones: the manual workarounds, the "always check this first", the exceptions that never made it into a formula.
Instead of leaving that as another flat list, we captured it as a knowledge graph — every forecast, order, supplier, lead time and rule held as a connected node, with the links between them made explicit. That turned scattered cells into an honest, navigable map of how supply and demand were actually managed, quirk by quirk — and gave the AI a structure it could reason over rather than just read.
Combining it with the wider picture
Process knowledge on its own isn't enough. We combined what we'd extracted with the broader organisational context — suppliers and their lead times, logistics and costs, demand patterns, and the goals behind the numbers. The result was a single, contextual knowledge base: not a frozen snapshot, but living knowledge that stays current as the business changes. This is the board the AI plays on — the full picture, in one place, owned by the client.
Checked by the people who knew the work
Before the agent touched any of it, the client's own experts signed it off. We took the reconstructed knowledge — the process we'd traced, the rules we'd recovered, the workarounds we'd surfaced — and put it back in front of the people who ran supply and demand every day. They read it, corrected what we'd misread, and confirmed the parts we'd got right. Only the knowledge they validated went on to ground the agent.
This step was deliberate, not a formality. It's a human-in-the-loop check that keeps people in charge of their own knowledge — and it's exactly what makes the automation trustworthy. The agent doesn't act on a guess about how the business works; it acts on a picture the experts have confirmed is right.
One agent, plain language
Then we put an agent in front of it. With the whole, validated picture to ground it, the agent automated the end-to-end process the spreadsheets used to carry. The hundreds of sheets, the links, the copy-paste — all of that complexity now sits behind it. The person in charge doesn't open a spreadsheet. They just say what they want, in plain English, and the agent does the work and shows its reasoning.
You: “What should we reorder this week, and where are we about to run short?”
The agent reads current stock and open orders, applies the lead times and reorder rules it learned from the sheets, weighs it against the demand forecast, and comes back with a costed plan — and the reasoning behind every line, traceable to its source.