The Picture Nobody Has Time to Assemble

Where It Goes Wrong

To answer one ordinary question — why did that customer get a short delivery twice in a month — somebody opens the order system, then the warehouse system, then the carrier's portal, then a spreadsheet that one person maintains. Four logins, four ways of naming the same site, four ideas of what a week is.

It can be done. It takes a morning, and it is done by the one analyst who knows which column in which export means what. So it gets done for the monthly review, and it gets done when something has already gone badly enough to warrant a morning.

The picture is always assembled after the point where it would have been useful.

The second cost is that the joining lives in somebody's head. The mapping between a site code in one system and a depot name in another is not written down anywhere, it is remembered. When that person is on leave, the question cannot be answered at all, and nobody quite says so out loud.

This belongs to the supply chain director, and it is carried day to day by the operations manager and the one analyst everybody relies on.

What Happens Instead

The first thing that happens is the boring thing: your data is read and sorted into the parts of a business it describes. Shipments, stock, carriers, suppliers and freight are one area. Invoices, reserves and settlements are another. Sensor readings are another. Vehicles and drivers another again. Records are placed by what they are about, so the same kind of fact lands in the same place whether it arrived from an ERP, a warehouse system, a transport system or a file.

Where the columns are too vague to place a table, it is placed by what the source is instead. A warehouse, yard, transport or order-management extract is logistics. A sensor or tag feed is telemetry. A claims system is finance. The point of that rule is to avoid the thing every catalogue tool does, which is to sweep the awkward half of your estate into a bucket called "other" and then never mention it again.

Then the joins. Where two areas share the ground they stand on, the link is drawn. Where an area shares no column name with anything else, it still gets a relationship rather than being left floating on the edge of the diagram looking irrelevant — a domain that appears unconnected is a domain nobody asks a question about, and that silence is usually wrong.

All of that is worked out by fixed rules from the shape of your files. No model is asked, no search goes out, nothing crosses the network for that step, and the same files produce the same map every time. A model is used afterwards to add commentary, and what it adds is clearly the commentary rather than the structure. The structure is something you can re-derive and check. Our own configuration files, which sit in the same place as your data, are deliberately excluded, because a tool that reports its own scheduler to you as your operations domain is not describing your business.

On top of the map, each area is reviewed: what state it is in, where the findings are, and for each finding its category, how serious it is, the rule it relates to and a suggested remedy, along with which system each part of it came from. When an area cannot be assessed, the answer is that it was not assessed and why — stated in those words, because "we did not look" and "we looked and it is fine" are not the same sentence and get read as the same colour on every dashboard ever built.

And then you can ask it things, in the vocabulary you already use, with the map and the recognised rules behind the answer and narrowed to whichever areas you are looking at. You get the reasoning, not just the reply. A scenario that has been worked through in the hypothesis lab can be handed across from there as a proposed action, with its variables and the rules it was argued against travelling with it rather than arriving as a bare reference to a run somebody else did.

One limit on all of that, stated here rather than left for you to find. What the mapping reads is the files sitting in the instance's own data directory, and on a standard instance that directory is not read — so what you get back is the refusal, not a thin map. The classification rules are real and they are deterministic; the path that lands your live extracts in front of them is not finished. We would rather the page said which half is which than describe the whole thing in the present tense and let a pilot discover the seam.

Where an area of the picture turns into a line about to run short, the response is the next job along, in stock cover and supplier planning, and the signal underneath it is demand forecasting. Visibility is what makes those two arguable from the same set of facts instead of from three exports.

What arrives at a person is a short ranked list of proposed actions with the records attached, not a diagram to admire. A named person approves, edits or rejects each one, and that sign-off is kept. Approving is what sends it, and a decided item leaves the queue instead of coming back round next time somebody opens the board. Where a step behind the approval has no implementation yet, the response names that step as not executed rather than reporting the action as complete — so the board shows what was decided and separately what was actually carried out. It all runs on machines you own.

Who Owns This

Three desks, and what each of them is holding today.

• The supply chain director. Today the picture gets assembled for the monthly review, or once something has gone badly enough to be worth a morning. What changes is that the same set of facts sits behind the question and behind the answer, so a decision is argued from one place rather than from three exports.
• The operations manager. Today one ordinary question means four logins, four ways of naming the same site and four ideas of what a week is. What changes is that records are placed by what they are about, so the same kind of fact lands in the same place whether it came from an ERP, a warehouse system, a transport system or a spreadsheet somebody emails on Fridays.
• The analyst everybody relies on. Today the mapping between a site code in one system and a depot name in another is not written down anywhere. It is remembered, and while that person is on leave the question cannot be answered at all. What changes is that the map is derived from the structure of your own files by fixed rules, so it is something anyone can re-derive and check.