Ask an AI assistant how many regular customers you have, then ask again tomorrow. There is a fair chance you will get two different numbers. Most people assume the tool is unreliable. More often, the tool simply does not know what you mean by “regular”, so it makes a reasonable guess, and a different reasonable guess the next time.
A data glossary is the fix. It is a short document that lists the words your business uses and says exactly what each one means.
Why AI guesses
Anthropic, the company behind Claude, advises treating the model like a brilliant new employee who does not yet know how your business works. A new starter who is told to count the regulars would come back and ask what counts as regular. An AI tool usually does not ask. It fills the gap with whatever seems most likely and carries on.
The words that cause trouble are rarely technical. They are the everyday terms everyone in the office thinks they agree on: active customer, regular, overdue, margin, completed job, new client. Ask three people in the same business what “active customer” means and you may well get three answers. One counts anyone who bought this year. Another counts anyone with an open account. A third counts anyone who has not asked to be removed from the mailing list.
People cope with that ambiguity because they share context. AI tools do not have that context unless you give it to them.
What a glossary looks like
It does not need to be clever. A plain list is enough, with one line per term:
- Regular customer: visited four or more times in the last 30 days.
- Active account: placed at least one order in the last 12 months.
- Overdue invoice: unpaid more than 30 days after the invoice date.
- Gross margin: sale price minus product cost, excluding delivery.
The test of a good definition is whether two people could apply it to the same data and reach the same number. “Customers who come in a lot” fails that test. “Four or more visits in 30 days” passes.
Start with the ten or so words that come up most in your reports, your management meetings and the questions you would like to ask an AI tool.
Agree the meanings first
This is the part most businesses skip, and it matters more than the tool. If the sales team and the accounts team define “active customer” differently, writing one version down does not settle the disagreement. It only hides it.
So the glossary is a business decision before it is a technical one. Get the people who own each number in a room, agree the definition, and write it down. That conversation is often useful in its own right, because it shows where reports have quietly been measuring different things for years.
Putting it to work in Claude
Once you have agreed definitions, you can give them to the AI tool. In Claude, the simplest route is a Project:
- Open Claude and go to Projects.
- Click New Project.
- Choose Set project instructions.
- Paste in your glossary.
- Save.
Take Example Café Limited, an invented business. Asked in an ordinary chat how many regulars it has, Claude gave different answers on different days because it was choosing its own meaning of “regular” each time. Inside a Project with the glossary saved, the same question uses the same definition every time, so the answers are consistent and can be checked.
Microsoft Copilot and other assistants have similar places to put standing instructions. The principle is the same whichever tool you use.
The limits
The glossary only applies inside the Project where you saved it. Start a normal chat and the tool is back to guessing. It is also only as good as your definitions. A vague definition gives you consistent vagueness, which is not much of an improvement.
Keep one master copy somewhere your team can see it, and update the AI tool when a definition changes. Otherwise you end up with two glossaries that drift apart, which is the problem you set out to solve.
Why this comes first
A glossary is one of the cheapest pieces of data groundwork you can do, and almost every AI project depends on it. Reporting, dashboards and assistants that answer questions from your own data all need the same thing: agreed meanings for the words you use. Sort that out first and the AI has something solid to work from.
If you would like help working out which definitions your business needs and where they disagree today, that is part of what our Data Foundations Audit looks at.