Meet the agents

LendEasy servicing runs on an AI workforce. Agents take the calls, answer the messages, read the documents, and carry cases through their workflows to completion. People are not the throughput of this system — they are its judgement. A matter reaches a person when it needs a decision only a person should make, and it arrives with the work already done and the evidence already attached.

end to end, no person in the loop

The workforce closes

• Inbound and outbound calls, through to a recorded outcome
• Conversations over chat, SMS and email
• Payments, arrangements and callbacks agreed on the call
• Documents classified, extracted and screened
• Cases advanced stage by stage as their conditions are met

where judgement is owed

People decide

• Terms outside the lender’s configured bounds
• A borrower in distress, or a protected status to confirm
• A settlement, a write-off, or anything that forgives money
• Anything the workforce declines rather than guesses at

The roster

Agent	What it does
Nova	Takes the phones, in both directions. Handles inbound borrower calls and places outbound first-party collections calls, verifies identity, explains the account, negotiates hardship arrangements and cures within the lender’s bounds, takes payment, and closes the call to a recorded outcome.
Arya	Handles borrower conversations across chat, SMS and email. Resolves who is writing and about what, answers within its authority, extracts commitments, schedules the next step, and sends the reply under the same compliance check as any other outbound contact.
Docgest	Reads every document that arrives — hardship proof, paystubs, identity documents, court filings, certificates — scanning, classifying and routing each one. Where a document type has an extraction profile, it also quotes the page field by field, and the platform re-derives every value from that quote before the case can act on it.
Cookie	Works beside the specialists on the cases that do reach a person. Answers from the record with its sources cited, flags what needs attention, and prepares the next action for confirmation.

Each agent is a registered worker with its own eligibility, tool grants, autonomy ceiling and evidence trail, working the same queues under the same permissions and the same compliance gate as the people beside it. Hiring, bounding, reviewing and stopping them runs through one contract — how agents are governed.

Each agent has its own page under this one. Everywhere else in these guides they appear as the voice agent, the communications agent, the document agent and the case copilot, because what matters operationally is the role and its authority rather than the badge on it.

A case that never reaches a person

A delinquency signal opens a collections case. The communications agent answers the borrower’s email, recognises hardship, and links it to the case. The voice agent calls, verifies the borrower, hears what they can actually pay, and agrees a two-part arrangement inside the lender’s configured bounds. The document agent reads the paystub the borrower uploads and confirms the income figure the hardship policy needs. The case advances through its stages and closes.

Nobody was asked anything. The evidence for every step — the call, the arrangement, the document, the policy check — is on the case for whoever asks later.

The same case, one stage further

The borrower asks for six months rather than three, which is outside the bounds. Now a person is involved, and the shape of that involvement is the point:

What reaching a person actually looks like

1. The agent recognises the request exceeds its authorityautomatic
2. It does not agree, and does not refusethe borrower is told it is going for review, not turned downautomatic
3. The work is packagedcall recording, the arrangement proposed, the income evidence, the policy resultautomatic
4. A person decidesone judgement, on a case that is already fully workeda person
5. The workforce carries out the decisionterms applied, borrower informed, case advancedautomatic

Autonomy is configured, not assumed

Every agent runs at a level the lender sets per action, and the ceiling can be lowered without redeploying anything. An agent may be allowed to answer only, to draft for review, to prepare an action somebody confirms, or to execute inside policy and report afterwards. Raising a ceiling is a decision the lender makes deliberately and can reverse instantly — the kill switch narrows or stops AI work by queue, worker, channel, task type, case type or tenant, and takes effect on work already in flight.

That is what makes autopilot a safe default rather than a leap: the workforce runs at whatever authority the lender is comfortable with today, and the same evidence trail supports moving that line in either direction.

Next on the roster

• Interaction QA reviewer — scores every human and AI interaction against the regulation-traceable rule library, producing the per-case-type quality measurements the autonomy ladder promotes on.
• Reconciliation investigator — works the drift queue: investigates intent-versus-ledger mismatches, proposes the classification, and drafts the evidence note.
• Complaints agent — intake, classification, and the regulated response clock, with the specialist reserved for the finding itself.