Issue #493 -> PR #494
The AI failed correctly. The product waited anyway.
The transcript was already done. The delay came from placing an optional full-text AI rewrite between "transcription complete" and "show the user the result."
The actual insight
This was not an AI quality bug. It was an AI placement bug.
A summary compresses a transcript. A formatter must give the whole transcript back. That makes the formatter's output grow with the meeting. If that work fails after timeout and falls back to local cleanup, the product bought latency without buying a better result.
Workload shape
"Clean this transcript" is output-sized work.
System shape
Before: completion was downstream of optional AI.
The fix
PR #494 adds a contract before AI can spend time.
Is this transcript formatting, not dictation formatting?
Did the user leave Use for transcripts enabled?
Is the transcript <= 20,000 characters?
If not, skip AI and finish locally.
Proof, not vibes
The numbers tell the story.
Takeaway
AI enhancements need a placement contract.
If AI can improve the result, great. If it is optional, slow, output-sized, or likely to fall back, it needs to sit behind an explicit gate. The user should never pay timeout just to receive the result the app already had.