The closer you are to the source of information, the better software you can deliver. In the AI era, I think this matters even more.
When AI makes implementation faster, waiting for clarification and rebuilding misunderstood requirements can eat up the time you saved. If we want to accelerate the release of good software, the path from a real problem to an understood problem needs to become shorter too.
A user explains something to support. Support summarizes it for product. Product writes a ticket. A developer turns that ticket into a prompt. Each step can keep the requested feature while losing some of the reason for it.
Imagine a ticket that says: "Add CSV export." The agent builds a button that exports every invoice. The file opens correctly. The feature works according to a reasonable interpretation of the ticket.
Now watch the person who requested it. Every Friday, they filter the invoice list to overdue payments for a particular customer. They need those results for reconciliation. Exporting everything leaves them selecting the same invoices again in a spreadsheet.
The useful information was in the work behind the request. DORA recommends letting developers observe user research directly for this reason.
Give the agent the clarified requirement, relevant business rules, current code, and an example of the expected result. Check that the export contains exactly the matching invoices, then let the user try it in their workflow.
The sources can still be wrong or outdated. Judgment and testing remain necessary. But keeping the original evidence close gives you somewhere to check an assumption and someone to ask when the next question arrives.