What should an AI do when one request asks for three things?
A feedback pipeline lost explanations when it completed fixes. I used labeled requests to design ordered intents and separate clarification and completion drafts.
The initial problem
“Explain why this happened, check one detail, and fix the rest” is a normal customer request. The existing pipeline chose one category for the entire request, so it could complete a fix and overwrite the explanation in the customer-facing draft.
In a labeled set of 35 requests, 19 contained multiple intents. Ten combined explanation and a fix. A successful implementation could therefore still leave part of the customer's request unanswered.
How I approached it
I used the labeled requests to examine which obligations could coexist and which depended on missing information. Explanation, clarification, investigation, and implementation needed an order; they could not be reduced to one exclusive label.
For a request containing both a clarification and a scoped fix, I distinguished dependent work from independent work. A change that depends on the answer must wait. Work that is sufficiently specified can proceed, while the customer still receives the question first.
I also checked the existing recovery behavior before changing time limits. The stages could be canceled and recovered individually. Adding a timeout for the entire request risked skipping work that could still be salvaged after a stage failed.
How I solved it
I changed the design to an ordered intent plan and preserved separate communications. The clarification draft arrives first, followed by a separate completion draft. That separation makes it easier for the customer to distinguish a pending question from an action already completed.
I kept limits at the stage level to preserve recovery behavior. Where the question content was already known, I assembled it deterministically instead of adding another model call, with the goal of getting the clarification to the customer sooner.
An explanation also needed support. Internal root-cause claims required evidence from the system and change history. If investigation could not establish the cause, the process escalated instead of inventing an explanation.
What the evidence supports
The documented changes established ordered intents and preserved separate drafts. The labeled set was small and specific to these operating requests; it is not a general estimate of all customer feedback.
This analysis does not establish a quantified adoption or resolution-time improvement. The next useful test is whether reviewers and customers can see every obligation fulfilled: the question answered, the cause supported, the change reviewed, and the final message accurate. Generation success alone cannot establish that.