An old FAQ says raw photo files come with every package. The current package guide says they cost extra. A customer’s confirmed order includes them as an agreed addition.
All three documents belong to the business. Only one directly answers whether that customer should receive the files.
In this example, a photography studio does not need more documents. More files can give an AI more conflicting material, while making its answer look well supported.
Make it clear which source answers each question, including what you agreed with that customer.
Decide which source governs which question
The studio’s general package guide describes what a new order includes. A confirmed order describes what this customer bought. An old FAQ may explain past practice, but it should not settle a current question about one customer.
“Use the newest document” would not solve this. A package guide updated yesterday could be newer than the customer’s order without changing what the studio agreed to deliver. Recency and relevance are different tests.
Create a small map of your recurring questions. For the studio, it could look like this:
| Question | Source that should settle it | What must be checked |
|---|---|---|
| What does the standard package include today? | Approved current package guide | Start date and package name |
| What did this customer order? | Confirmed order, including approved changes | Right customer, right job and latest approved change |
| Has the customer’s gallery been delivered? | Delivery record for that job | Actual delivery status, not the planned date |
This map is a business decision. The person responsible for the service should settle it before anyone tries to solve the conflict through a clever prompt.
Make a source’s status visible
A filename such as “Packages final updated” tells a reader very little. The source itself should show what it covers and whether it is still approved.
For the package guide, a short status block could say:
Applies to: New portrait bookings from November 1.
Approved by: Studio manager.
Replaces: Package guide used for new bookings through October 31.
Exceptions: Check the customer’s confirmed order and any approved changes.
Review trigger: Any approved change to package contents or delivery terms.
These details prevent a common mistake: treating a change in general policy as a change to every existing job. They also help employees give the same answer, with or without AI.
Take replaced guides out of the documents used for current answers, or, if your system allows it, limit them to questions about the past. Keep the archive where the business needs it. Keeping a record and treating it as current guidance are separate choices.
Decide how current each answer needs to be
Some answers change when a manager approves a new policy. Others change each time someone completes a job. A monthly document review cannot tell you whether a gallery was delivered ten minutes ago.
For information that rarely changes, name the event that means it needs review: a new package, a changed service or a revised delivery process. For live status, check the job record and show when it was checked. If no one can check it, the answer should say the status is not confirmed.
A good answer might be, “The job record shows the gallery was sent at 10:15 this morning.” If the workflow has only yesterday’s export, it should not turn that into a claim about today. It can report what the export says and ask for a current check.
Test conflicts, not just easy questions
Once the studio has chosen its sources, it should test the question that exposed the problem. Does the answer about raw files follow the customer’s confirmed order? Does a question about a new package follow the current guide, and a question about an old package stay with the archive?
Then edit one thing in a test copy: remove the approved change to the order, add an unresolved conflict or make the delivery status unavailable. Check that the answer changes too. A useful system must recognize when the material in front of it does not settle the question.
Ask for the source behind each answer during these checks. A citation to a real document is only the beginning: it must support the exact statement and fit that customer, service and time. If your tool cannot show its source, the reviewer needs another way to check.
Good sources make wrong answers less likely, not impossible. If one still reaches a customer, know how you will correct it.
Begin with five questions your employees often bring to the owner. Name the source that answers each one, settle the conflicts and decide what happens when the answer is missing. That small exercise can make the next AI answer much more useful than uploading another hundred files.
