An AI tool that answers questions about sales or stock is only useful if people believe it. When the numbers don't match what they know, they go back to their own spreadsheets, and the system they stopped trusting stays on the server, still running, used by nobody.
Why a written list of meanings keeps coming back
The idea that fixed this in the 1990s never went away. Every major reporting tool since has built its own version, and Microsoft renamed Power BI's datasets "semantic models" in late 2023. The weakness was that each list lived inside one tool. A company that defined revenue in one product and then bought another had two definitions, and nobody noticed until the totals disagreed.
AI changes where the list has to live. A company's dashboards, spreadsheets, chat assistants and software agents all read the same data, so the definition has to sit underneath all of them. That is what Databricks' metric views do: a company defines a measure once, and reports, database queries and AI tools all use it. It is the 1991 idea moved out of one tool and down to the data itself.
The analyst who asked a question
Picture a regional building-supplies distributor. Finance counts revenue after returns and without sales tax. The sales director counts every order booked. Both are right for their purpose.
The owner types "what was revenue in the third quarter?" and the assistant picks one meaning. Say the sales director sees a figure 6% above the board pack. She does not investigate. She stops asking the assistant, and her team follows her.
Before chat, a human analyst stood between the question and the number, and the analyst's most useful habit was asking "which revenue do you mean?" Typing a question into an assistant removes that person. Some tools now ask the clarifying question themselves, but they can only ask a good one if the company has written down what the options are.
Agents make the definition a business rule
A person who misreads a number can ask a colleague. A software agent that adjusts prices or reorders stock reads the definition every time, on every product, with nobody to ask. If "available stock" includes goods already promised to customers, the agent reorders too little, and it does so across the whole catalogue before anyone looks. At that point the definition is no longer documentation. It is a rule the business runs on, so it needs an owner, a place where it is stored, and a change history.
A practical first step
Before you test any assistant on your company's data, list the five numbers your team argues about most, perhaps revenue, margin, active customer, on-time delivery and available stock. Write one sentence for each saying exactly what counts, and name one person who owns it. Then ask the assistant a question about each and check which meaning it used. Where the answer surprises you, you have found the gap that would have cost you your users' trust.