Supply as a signal, not a spreadsheet

Between what a person took at a location and what we purchased sits a chain of retellings. Every link adds delay and loses detail.

A spreadsheet is a week late

The classic scheme: a shift notices something is short, the manager adds it to a request, the request goes to purchasing, purchasing merges it with the rest. By the time a decision is made the original observation is stale, and the reason it ran out has been lost.

The problem is not people’s discipline. The problem is that data travels through memory and retelling instead of being recorded at the moment of the event.

What counts as a signal

We call a signal a structured event: what exactly, where, when, in what quantity and in what context. Not “we ran out of milk”, but a consumption event tied to a location, a time and a remaining stock level.

• Consumption and stock are recorded as events, not as a weekly total.
• A deviation from the normal consumption rate is its own signal, not a row in a shared table.
• Returns and write-offs are signals too: without them the demand picture is wrong.

Forecasting does not remove the human

Demand forecasting is the obvious application of the AI layer, which is exactly why it is easy to overrate. We work from the assumption that automation prepares a decision and a person confirms it: purchasing stays an area of responsibility, not an autopilot.

“Automation that cannot be challenged is not a system — it is someone else’s opinion with an interface.”

What breaks today

The honest picture: event-level recording does not cover every item, part of the data is still entered manually, and signal quality depends on how carefully a shift is closed. While that holds, the supply layer stays in roadmap status rather than being a working mechanism.