Six questions your data must answer before an AI uses it
An AI agent that reads a price list, a contract or a customer file will act on it without hesitation. It does not know that the price list expired last month, that the contract forbids sharing, or that the customer asked to be deleted. People know these things from experience. A machine only knows what is written down. So before an agent touches a piece of data, the data has to answer six questions about itself.
The six questions
1. What is it? The exact type and format: a supplier invoice, version 3 of the price grid, a signed delivery note. Not "a file in the shared drive".
2. Who says so? Where it came from and who vouches for it. An invoice entered by accounting and checked against the order is not the same as a PDF forwarded by email.
3. How may it be used? Read only, used as the basis for a decision, used to trigger an action, shared outside the company, used to train a model. Each is a separate permission.
4. What are the legal terms? The contract or licence it falls under, whether it contains personal data, where it must be stored and for how long.
5. What happens if it is wrong? Who notices, who corrects it, and who bears the cost. A wrong address on a delivery is an annoyance; a wrong rate on a thousand invoices is a loss.
6. How is a dispute settled? If a customer contests a figure the agent used, what is the path: correction, reversal, escalation to a named person.
Why this matters more with AI
Most companies answer these questions informally, in people's heads and in the habits of the team. That works while people do the work. When an agent does it, the informal knowledge is missing, and the agent fills the gap with confidence. The errors that follow are not dramatic. They are quiet: a discount applied from an old grid, a document sent to the wrong party, a decision based on a draft.
Writing the answers down and attaching them to the data itself fixes this. The answers travel with the record. Any agent, any system and any auditor reading the record can see what it is and what it may be used for.
Where to start
You do not need to document everything. List the ten kinds of records an agent will read or write in your first automated workflow. For each, write the six answers and name one person who owns them. If a question has no answer, that is the finding: the agent does not get that record until it does.
Six short answers per record is less work than one bad decision taken at machine speed.