When the AI proposes a task is complete
Not every reply that satisfies a task is obvious. When the classifier is reasonably but not-fully confident, it flags the task as "AI-proposed complete" and waits for your call.
How this works
Every inbound reply on a deal runs through a classifier that scores it against every open task. Each task gets a confidence number 0-1:
- ≥ 0.85 (high) — task auto-completes. The task drops into the Plan view's Done group with "AI auto-completed" and the quote as evidence.
- 0.60 to 0.84 (medium) — task moves to
proposed_complete. A card surfaces under the Approvals filter. - < 0.60 — no match. Task stays where it was; the reply is still stored as a message on the thread.
The Approvals card
When you open the Approvals filter on a deal (or land on the Today page), proposed-complete cards sit alongside outbound drafts awaiting your approval. Each one shows:
- The task that might be complete
- The quote from the reply that triggered the proposal
- The confidence percentage
- Who sent the reply and when
Two buttons:
Confirm — task flips to complete. Completion source is recorded as ai_proposed (distinct from ai_auto or manual) so the audit trail shows you explicitly approved.
Reject — task returns to pending. The classifier is welcome to try again on a future inbound that satisfies the criteria more clearly.
When to confirm vs. reject
Confirm when the quote really does say the task is done, even if the phrasing is unusual. Example: inspection task, quote "we got a clean report yesterday, all clear." That's complete.
Reject when the quote is ambiguous, references something else, or is a partial confirmation. Example: EMD task, quote "I'll get that wired tomorrow." That's intent, not confirmation. Reject → the actual wire confirmation will proposed-complete later.
Why we don't auto-complete in the 0.60-0.84 range
Because the cost of a wrong auto-complete is high. If BuyerFlo auto-closes a task that isn't actually done, the engine stops following up — and by the time you notice, the party has moved on and the deadline is a week past. Waiting 10 seconds for you to click Confirm is cheap insurance.
Cross-party proposals
A reply from the listing agent can propose completion of a task originally assigned to the appraiser. This is intentional — the coordination engine cares about the outcome, not who reported it. The card calls out "Satisfied by [name]'s reply (not assignee)" when that happens, so you can see at a glance whether the source matches the task.
If you reject everything
Some deals just generate unclear replies. If you're rejecting most proposals, it's worth checking whether the task's completion criteria are set correctly. Rare edge case, but email us if you see a pattern.