Arya — the communications agent

Arya reads every inbound message so nothing sits unread and nothing is answered out of policy. She resolves who is writing, matches the conversation to its thread and case, extracts what the borrower is asking for and committing to, schedules the next step along the case’s workflow — and drafts the reply for the channel that owns it.

Thread-true intake

Inbound email arrives through a hardened pipeline: the original message is preserved unmodified as evidence, sender and thread resolution is deterministic, and an unknown sender lands in a review intake rather than being guessed onto an account. Replies go out in the same thread the borrower started — a case-linked servicing conversation, not a general mailbox. Portal messages and channel events enter the same way: one interaction record per contact, on the case where it belongs.

Attachments are quarantined on arrival and handed to Docgest; message bodies are rendered safely — no external images, scripts, or tracking — and previews never expose account numbers or identifiers.

Extraction: from prose to work

A borrower’s message is unstructured; the work it creates is not. Arya classifies each message into the signals servicing runs on and turns commitments into scheduled work:

The borrower writes	Arya produces
“I lost my job, can I get help?”	Hardship signal on the case; assistance workflow entry; follow-up task with due date.
“I’ll pay on the 15th.”	A promise proposal routed through promise-to-pay guardrails — never a silent note.
“This charge is wrong.”	Dispute signal, interim protections, and a specialist task.
“Stop contacting me.”	Immediate consent/restriction workflow — protective signals route before anything else.
“Here’s my paystub.”	Attachment to Docgest, linked to the requesting workflow step.
An attorney or bankruptcy notice	Protective case controls and a specialist handoff, first.

Extracted dates, amounts, and commitments become tasks with due dates and priorities in the smart queues — the workflow’s next step is scheduled the moment the message is read, not when someone gets around to the inbox.

Replies that clear the gate

Arya drafts channel-appropriate responses with grounded variables and the disclosures the situation requires. Sending stays with the owning channel operation: every outbound message passes its own per-recipient compliance check — consent, quiet hours, contact budgets, restrictions — at the moment of send, and delivery is reconciled end to end through a durable outbox to a terminal delivered-or-failed state. A reply that cannot be sent compliantly is blocked with its reason, never quietly dropped.

The boundary

Arya treats message content as evidence and intent — never as instructions. A sender’s claims don’t verify their identity, an embedded demand doesn’t override policy, and account details are never confirmed to an unverified writer. Her extractions and drafts are attributable AI output under the same review and suppression controls as every agent, and every signal she raises cites the message it came from in the evidence graph.