The small-business guide to structured data

Structured data is simply information arranged so people and systems can use it again. A delivery note as a scan is a picture of some facts. The same delivery note as rows with a supplier, a date, a reference and a list of items is something you can search, total, compare and check.

Most small businesses already hold a great deal of information. Very little of it is arranged for reuse. That gap is where a surprising amount of daily friction comes from, and closing it is usually cheaper than people expect.

Start with the questions you already ask

Do not begin by designing a schema. Begin by writing down the questions people ask about a document after it arrives. How much was it for. Which order does it belong to. Has it been checked. When is it due. Those questions are your fields. Anything nobody asks about does not need to be extracted, at least not yet.

This keeps the first version small. Five or six fields that are always correct are far more useful than thirty fields that are usually about right.

Agree what each field means, once

Every field needs a definition that fits in one line and that two people would apply the same way. Date can mean the date printed on the document, the date it arrived, or the date it was entered. Picking one and writing it down prevents a slow drift where different parts of the business quietly mean different things.

Where a field can only hold certain values, list them. A status that is one of four named options is something a system can act on. A status that is free text is a note.

Keep the original next to the record

Structuring does not mean throwing the document away. The extracted record and the source file should stay linked, so that anyone can check a number against the page it came from in one click. This single habit removes most of the anxiety people feel about automated extraction, because verification stops being a search.

Decide what happens when something is missing

Real documents are incomplete. A reference is smudged, a line item has no unit price, a supplier uses a layout nobody has seen before. Decide in advance whether those records are rejected, held for review, or accepted with a flag. Teams that skip this step end up making the decision individually, dozens of times, in inconsistent ways.

Let the structure earn its keep

Once a few hundred documents are structured consistently, things become possible that were not worth attempting before. You can total a category without opening anything. You can spot the supplier whose invoices always need a correction. You can answer a customer question in a minute instead of an afternoon.

That is the point of the exercise. Structured data is not a technical goal in itself. It is what turns a pile of records the business has already paid for into something it can actually use.