Not a replacement for the floor. A coworker on it.
The FormantAI agent takes the bucket-1 reminders and the after-hours queue. Your people take the escalations, the legal threats and the judgement calls — and when a call moves between them, the context moves with it.
The worst moment on any floor is "so, what's this about?" — a customer repeating an angry story to a second person. We built the coworker so that moment never happens.
Hands the call. Whispers. Does the paperwork.
With a full brief.
Who the customer is, why the call is moving, what has been promised, the sentiment, and which frozen lines have already been read. The human never asks "what's this about?"
On the human's screen, not the customer's ear.
Mid-call: the UTR lookup, the policy check, the next line in Hindi. The agent suggests; the human decides what to say.
Notes, disposition, CRM.
The agent drafts the wrap-up for every human call. The human accepts or edits in one click — and each edit is a correction the weekly cycle learns from.
Humans can step in. Humans grade.
A supervisor can take over any live agent call. Every week, your QA reviewers grade a blind sample — the same rubric the evaluation model uses.
What the human sees the second it lands.
A warm transfer is only warm if the context arrives first. The brief is built from the call's own trace — the verified identity, the stage the agent reached, the promise-to-pay state from the system of record, the sentiment curve, and the compliance lines already read. Nothing is inferred; every line has a timestamp.
Routing is a governance rule, not a model's mood: a legal-threat phrase, a deceased-borrower mention, a fee-waiver request or a customer who asks for a person goes to a named human queue, inside calling hours, with the transcript and audio attached.
identity verified 00:14 (DOB) · consent captured · calling-hours check ✓
Frozen lines verified verbatim — do not repeat the disclosure.
Illustrative event log. replace with a live handover trace
The agent stays on the call as the quiet one.
After the transfer, the agent keeps listening — for the human, not the customer. It runs the UTR lookup when the customer says "already paid", flags the Fair Practices Code line when tempers rise, and drafts the dispute summary in the customer's language. Every suggestion is a suggestion: the human speaks, the human clicks "OK".
It works the other way too. A supervisor can take over any live agent call from the floor console. The agent hands over the same brief about itself, steps back to whisper, and the call is flagged for the weekly QA sample — because a takeover is the most useful correction there is.
How corrections become next week's releaseWho does what, across a shift.
The agent takes the volume: bucket-1 reminders, callbacks, after-hours service. The humans take the judgement: escalations, legal threats, waivers, grief. The line between them is a policy your ops lead owns — not a setting we hide.
Reviewed weekly, like anyone on the floor.
Every week your QA reviewers grade a blind, stratified sample of the agent's calls — by language, by cluster, and every call a human took over. The grades go two places: to the evaluation model, which recalibrates against them, and to the weekly cycle, which turns the corrections into the next candidate release. A named human on your side signs that release.
The floor is not competing with the agent. The floor is training it — the same way a good team lead trains a new joiner: one corrected call at a time, with the frozen lines never up for debate.
Volume.
Bucket-1 reminders · callbacks · "already paid" confirmations · after-hours service and renewals · every after-call wrap-up draft.
Judgement.
Legal threats · fee waivers · deceased-borrower and hardship cases · anything the customer asks a person for · the weekly grade and the release signature.