LISTENConnectors, not uploads
RingCentral / Office@Hand with JWT refresh, detailed call legs and recording fetch; CallRail; iovox CSV; MP3 upload. Every inbound call lands in a tenant phone log; recordings stay in the tenant’s bucket; ElevenLabs Scribe v2 transcribes diarized on demand. Nothing crosses tenants.
LEARNThe miner
Claude Haiku labels outcomes and extracts signals from each call (questions, objections, the winning line); Claude Opus synthesizes the playbook; the human booking baseline is measured from the log before the AI takes a call. Industry packs hold aggregate patterns only, never transcripts. The first pack is direct-response events, learned from the pilot.
LAUNCHOne answering rule
Voice and web chat from one agent definition on ElevenLabs Agents, Claude reasoning inside. Rollout is supervised by default: after-hours → overflow → full line, flipped by a phone-system rule. Recording-consent line chosen by caller area code, strictest form when unknown; AI disclosure on every greeting. The onboarding wizard goes site → draft → questions → build → “Talk to your desk”.
GATESLint, contract, regression, drift
A renderer turns pack + playbook into the system prompt, tools.json and expected agent config. The contract gate fails the build if any reason the API can return lacks a prompt branch. Regression runs simulated conversations seeded from mined objections. Nightly drift diffs the live agent against the repo; publish() refuses a version that failed.
GUARDThe live-agent guard
A database cron probes every live agent each minute through pg_net, so the API can stay asleep. On a block it alerts, then restores the last version proven by a real conversation, once per incident. 1,440 of 1,440 checks passed in its first day. A version is “proven” only when a caller actually talked to it.
PROVEA receipt-true board
A call counts as booked only when a booking row carries its conversation id. The owner sees booking rate against the human baseline, a stage timeline (Listen / Learn / Launch / Prove), playbook evidence counts and version history with gate results. The Call Activity screen gives call-tracking parity: tiles, by-hour charts, filters, export, recording playback, tags, notes.
SCANThe call scan
Quick and deep scans over a business’s phone log: a switchboard jack grid lights in the real answered / missed proportions while findings drop in with their evidence counts. Paced playback of real scan events only, never a canned animation. Also ships as a CLI for an offline first look.
TENANCYOne database, RLS everywhere
tenant_id and row-level security on every table. Dashboard sessions are read-only; every write goes through an admin route under the service role, so no one can flip a gate result. Capacity lives in SQL functions, never in the prompt. Agent-facing routes never return 4xx/5xx to the agent’s own mistake: always 200 with a reason the prompt knows how to handle. No timers in the process; schedules live in pg_cron.
VERIFIED38 live checks after every backend change
One script exercises the live demo line end to end: agent tool dry-runs, consent by area code, signed post-call ingest with audio and outcome labels, every job, cron picking up a queued scan on its own, admin routes with 401/403/RLS, billing, then deletes everything it created. 38 of 38 pass.