Vendor Inbound: the A&K Repeat
What went wrong when a vendor tried to schedule a delivery, why it was a repeat of a July incident, what we shipped same-day, and the longer-term opportunity areas. Long-term section: NOT a short-term priority
What happened (all times CT, Aug 5)
- 10:40 Maddie (A&K Appliance Service) calls asking for Loretta. Transferred to a human. Fine.
- 1:52 She calls back: "Scheduling… Friday, August 7th." Clara: "Got it, I'll get that to the property manager." Nothing is recorded anywhere — no note, no task, no work order, no notification. Call marked resolved.
- 2:20 Third call: "I wanna talk to Loretta, damn it… did they already do that?" Clara can't answer — there is nothing to read back. Transferred.
- 2:45 Fourth call (Donnie, A&K) reaches Joanna directly on a transfer. On the recorded leg Joanna says: "This AI system is getting on my last nerves." The Friday delivery plus a fridge/stove for unit 422 (PO 695) get confirmed human-to-human.
Resolution: a human called Maddie back same day and confirmed Friday's delivery and the unit 422 appliances. Customer-side, closed.
Why it happened
Contributing color: everything shipped since July went to the outbound vendor lane (robo-vendor battery, Willows go-live Aug 3). Inbound-cold-call capture was never anyone's sprint.
What we shipped same-day (PR #5469)
The chosen arc (Fede, Aug 5): shadow → scribe → jump back in slowly
1. Shadow mode now. Vendors always get transferred to the office; Clara listens but doesn't try to help yet. (= the dial-down above, on for Camellia.)
2. Shadow scribe next — the first pass at adding value without getting in the way. Listen to the transferred human leg and put agreed jobs/days on the property calendar (marked unconfirmed). Not built yet, but ~70% of the parts exist: post-transfer transcription works, the outcome extractor works, the calendar projection works (jobs Clara herself books already land on the calendar — PRs #4622/#4380/#4920). The missing wire: run the extractor on the post-transfer transcript and let it write a calendar entry without requiring a bound work order.
3. With more data and examples, Clara re-enters lane by lane — capture+notify, then handling — each step gated on the gauntlet proving it (the eval suite now grades the transfer-first posture directly).
Long-term opportunity areas not prioritized short-term
Recorded so the vision isn't lost; none of this is scheduled. Roughly in impact order.
A. Vendor capture — the durable fix (needs a decision)
Give Clara a durable record for what vendors ask for: a visit/commitment entry keyed to property + unit + job, plus a capture-and-notify path so "Friday the 7th" becomes a stored item the PM sees and Clara can read back on the next call. This is ADR-0113's Phase 1; the dial-down's own rollback trigger points at it. Decision needed:
Ship ADR-0113 Phase 1 as scoped (commitment record + capture-and-notify on the existing lanes).
Smallest path to "vendors' words survive." Reuses the existing note-delivery leg. Exit criteria for the dial-down are already written against it.
Days not weeks, but notes aren't structured — no read-back on the next call, no calendar tie-in, and we likely rebuild it later.
Zero eng cost now; accepts that the office owns vendor scheduling and Clara never learns the lane. The corpus keeps growing either way.
B. Promises must be mechanically backed
Principle worth adopting lane-wide: Clara may not say she'll do a thing unless a tool call records it. Concretely: un-whitelist the pass-along phrase in the say–do guard and make "promise with no write" page someone. Cheap, catches whole classes of future incidents, not just vendors.
C. Trust the timeline
Two small integrity bugs made the PM's view actively misleading: the call summary inverted who called whom, and a fully-identified vendor still displays as "Unknown caller." Small fixes; disproportionate trust damage when the PM reads the log while a vendor is chasing them.
D. Vendor contact data is the root constraint
78 of 814 vendors have a phone on file; 27 have email; A&K — 211 bills at Camellia — has neither. Any future where Clara coordinates vendors proactively (confirmations, reminders, reschedules) needs a contact-coverage push, PMS-side or ours.
E. Graduated autonomy as the operating model
The dial-down switch generalizes: new properties start transfer-first, and Clara earns lanes back per property as the eval suite proves them (the gauntlet now grades the transfer-first posture directly). "Onboard cautious, dial up on evidence" could become the standard rollout posture for every capability, not just vendor inbound.
Status
| Item | State |
|---|---|
| Maddie / Friday delivery + unit 422 (PO 695) | Confirmed by a human, Aug 5 |
| Transfer-first dial-down (PR #5469) | In review pipeline; flipping on for Camellia at merge |
| Shadow scribe (post-transfer → calendar) | Chosen as the next value step; not yet built or scheduled |
| A — vendor capture decision | Waiting on Fede (options above) |
| B, C — guard + timeline fixes | Small PRs, unscheduled |
| D, E — data coverage, graduated autonomy | Vision, unscheduled |
Sources: prod conversation records for the four Aug 5 calls; the five-agent investigation of the same day; ADR-0113; the rated Camellia vendor corpus. Quotes are verbatim from call transcripts.