Your next reply
{{ c.headline }}
{{ c.sub }}
{{ c.inlineFlag }}
Grounded in 1 source
{{ c.sourceWarn }}
Tier 3 · autonomous send withheld
{{ gateSummary }}
Spec notes
{{ emptyTitle }}
{{ emptyBody }}
Escalated from Resolve
{{ e.headline }}
Everything you need to take this forward.
Suggested next step
Never request a PIN or OTP.
Grounded in 1 source
Your reply
Written by you — Resolve withheld a draft: no authorised payment record.
Spec notes
Quality over activity
Did we actually help?
Three numbers that keep resolution honest.
Approve, edit or hand off a draft in the inbox — every action writes here with the approving user.
Draft quality on held-out conversations, before interface work counts as progress.
SAMPLE DATA — Illustrative, separate cohorts. Full 7-day follow-up. Not pilot results.
Manager queue
Decisions only a person can make.
Everything Resolve blocked or an agent escalated, with the rule that stopped it and the evidence still missing.
You are signed in as an agent. Approving a refund above the limit is a manager action, checked per action and category — not a screen you can simply open.
Blocked drafts and agent escalations land here with their rule and evidence. Hand off a case in the inbox to see one arrive.
POST /v1/cases/:id/approvals · requires limit.approve (manager)
approval is a record, not a prompt flag · draft re-checked after override
audit: limit.approved{by, amount, rule} · limit.declined{by, reason}
Pilot workspace · demo
Two channels, one queue.
Secure login with agent and manager roles. Permission to draft a reply is separate from permission to send it or change a customer record.
Onboarding, configuration and invoicing stay manual in v1 — no self-service signup.
POST /v1/auth/session → {account_id, user_id, role}
role ∈ agent | manager · checked per action + category
account + role checks on reads, searches, exports, jobs, sends
Workspace setup · step 2 of 2
Connect the two channels.
A customer-facing launch is conditional on approval and working end-to-end delivery. Offline evaluation continues while approvals are pending.
POST /v1/channels/:id/verify · webhook dedupe on (channel_id, message_id)
retry with backoff · delivery failures surface on the case, not in logs only
concurrent agent actions guarded by case version (optimistic lock)