The booking bot that gets read every Monday
A founder-owned home-services marketing agency of 20 to 30 people had its pest-control client's booking bot, built by another firm, give wrong answers, block access to a person, go unreviewed after launch, and get pulled.
Booking bot gave wrong answers and nobody reviewed its conversations
A customer-facing booking bot answers only from a client-owned approved-facts file, discloses it is automated, and offers a person on every turn. Each Monday a review strips personal data, groups failed chats into clusters and proposes changes. The client's customer-service lead approves or rejects each one, and a replay on held-out chats gates v2.
- on 60 held-out conversations the review never read, 33 ended in a failure signal on v1 and 15 on v2 (conversations we scripted, failures we planted, a keyword-rule bot standing in for a model)
- the customer-service lead approved 4 changes and rejected 2, and the run checked v2 for both rejected changes: 0 found
- Booking chatbot
- Approved-facts file
- Weekly conversation review
- Personal data redaction
- Failure-signal rules
- Human approval gate
- Held-out replay testing
- One-command rollback
- Node.js
Same problem?Send it over.
Tell me what your version looks like and which tools it runs in. I reply with how I would build it.
-
Pharma marketing
The pharma creative that stopped dying at legal review
creative rejection went from 75% to 10%, and the remaining 10% is rejected over creative differences, not compliance
-
Performance marketing
The retainer, on the record
all 6 approved spend changes carry a named approver, and the gate sent 1 approval back for a missing reason
-
B2B marketing agency
The launch page that asks before it learns
the 220 simulated visits that declined or ignored the banner left 0 stored events and 0 storage keys, and a real browser that declined held 0 keys and 0 cookies