Same queues
Agents pull work from the same smart queues as your team — routed by the same priorities and SLAs, visible in the same command center.
AI Agent Workforce
AI that does the work — and can prove every step was allowed
LendEasy agents are not a chatbot bolted onto servicing. They are governed workers: same queues, same permissions, same audit trail as your team — with a compliance gate in front of every action and an evidence record behind it. Automation your regulator can examine, not just admire.
Collections call — AI Agent
Case #C-2417 · first-party · early delinquency
Live transcript
“…I can set that up. Your past-due amount today is $215.00 — would two payments, starting Friday, work for you?”
$215.00 · from the ledger — never the model
Compliance gate — cleared before the dial
- 7-in-7 budget · 2 calls remaining
- Borrower local time · 2:14 PM — inside window
- Consent · on file — voice, cell
Proposed action
Promise to pay — 2 × $107.50, starting Fri
Within guardrails: 30-day horizon · minimum amount · installments
Every figure from the ledger · every dial gated · every step hash-chained into the evidence graph
A worker, not a widget
Govern AI the way you govern employees
The safest place for an AI agent is inside the same machinery that already governs people. LendEasy gives agents a worker identity — nothing more, nothing less — so every control you trust for humans applies to AI automatically.
Same permissions
An agent holds a scoped worker identity on the deny-by-default API surface. It can do exactly what it is permitted to do, and nothing else.
Same audit trail
Every agent action lands in the same hash-chained evidence graph as human work — with the model and prompt versions that produced it pinned to the record.
Same approvals
Maker-checker policy applies to agents too, and the reviewing human is always distinct from the agent that proposed the action.
- 1
Propose
The agent drafts an action — a message, a promise, a payment — tied to the case it serves.
- 2
Compliance gate
Version-pinned, regulation-traceable rules evaluate it before anything executes. Missing facts never permit.
- 3
Approval ladder
Routed by the task type's autonomy level — a distinct human approves until policy says otherwise.
- 4
Execute
The action runs through the same governed interface humans use. Figures render from ledger facts only.
- 5
Evidence
Decision, model + prompt versions, approvals, and outcome land in a hash-chained record.
Learn & QA loopa quality program samples outcomes; findings tune prompts and policy — version-pinned, never the ledger
Graduated autonomy
Five levels, chosen per task type
Autonomy is not a switch you flip for the whole platform. Every task type starts in shadow, and case summaries can run autonomous while payment promises stay at approve — you promote or demote each task type independently, on your evidence.
Level 0
Shadow mode
Before a task type is activated, the agent runs in shadow: outputs are produced, recorded, and scored — never sent, never executed. Activation requires the evaluation gate to pass.
Level 1
Observe & summarize
The agent reads cases and produces summaries, timelines, and context briefs. It touches nothing.
Level 2
Assist
The agent drafts responses and recommends next actions. A human edits, owns, and sends everything.
Level 3
Approve
The agent prepares complete actions — messages, promises, payments — and a distinct human approves each one before it executes.
Level 4
Autonomous within policy
The agent executes on its own, inside hard guardrails — and still passes the compliance gate on every single action.
No task type skips shadow, and every level — including fully autonomous — still passes the compliance engine before each action and writes the same evidence record. Autonomy changes who clicks, never what is checked.
Structural grounding
Every number a borrower sees comes straight from the ledger
Most teams try to prompt hallucinations away. LendEasy removes the possibility: every financial figure in borrower-facing output is filled in by the platform from the ledger itself. The model writes the words; the ledger supplies the numbers.
What the agent composes
Hi Jordan — as of {today}, your remaining balance is {ledger.balance} and your next payment of {ledger.next_due_amount} is due {ledger.next_due_date}.
Highlighted slots are resolved by the platform from the ledger at send time — never generated by the model. A message with a free-typed figure cannot leave the system.
- Figures resolve at send time, so they reflect the ledger as of that moment — not as of model training or case open.
- The same mechanism covers balances, due dates, payoff amounts, and payment terms.
- If a required fact is missing or stale, the message is held — missing facts never permit.
Safety systems
Defense in depth around every agent
Grounding and the compliance gate are the front line. Behind them sits a full operational safety stack — built for the day something surprises you, because something always does.
Shadow mode before activation
A new agent type runs in shadow first: its outputs are recorded and scored, but never sent and never executed. It goes live only when the evaluation gate passes — not when someone feels confident.
Model registry, every version pinned
A model registry pins exactly which model, prompt, and context-packet version produced every output. Activating a new version is a governed, audited change — sessions cannot run on unregistered combinations.
Quality sampling & drift detection
The evaluation program is quantified — per-case-type sample minimums and pass thresholds — and re-runs on every model or prompt change and on a regular cadence. Behavioral drift is caught by evaluation, not by borrower complaints.
Fairness & consistency monitoring
Within the QA program, sampled outcomes are reviewed for consistency across borrower segments — and because every AI decision records its inputs, policy version, and model version, fair-lending review has the evidence it needs.
Instant kill switch
Suppress agents in one action, scoped to what is actually wrong: globally, per queue, per case type, or per channel. Human work continues uninterrupted.
PII scrubbed before model calls
Personally identifiable information is stripped before anything reaches a model — frontier or self-hosted. The model reasons over the case, not the identity.
Distinct-human review
The approver of an agent action is always a different party than the agent itself — structurally, not by convention.
Model-agnostic by design
Run any frontier model or a self-hosted one. The governance — gate, grounding, registry, evidence — lives in the platform, so swapping models never weakens it.
Day one
What agents do from the first week
No multi-quarter AI program. Agents start on the work that consumes your team today — with the phones leading — governed from the first action.
Work collections calls end to end — hardship conversations included
AI voice agents place and answer collections calls: they negotiate payment plans and promises to pay within your guardrails, recognize hardship language and route it to the hardship workflow, and hand the call to a human the moment the borrower asks or policy requires. Like every task type, voice graduates — start with reminder and early-delinquency calls, and expand as your evidence supports it.
- Every dial clears the compliance gate first — contact budgets like Reg F's 7-in-7, calling windows in the borrower's timezone, consent, and protected-borrower states.
- Every figure spoken comes straight from the ledger — the model speaks the words; the ledger supplies the numbers.
- Every call is recorded and transcribed into the case timeline — evidence from the first hello.
Summarize cases
Walk into any case with a grounded brief: who the borrower is, what has happened, what protections apply, and what is owed — as of right now.
Draft compliant responses
Replies drafted against the rules in force for that borrower, that product class, that jurisdiction — with every figure rendered from the ledger.
Recommend next actions
The agent proposes the next best step on a case — and shows the compliance evaluation alongside it, so reviewers see why it is allowed.
Negotiate within guardrails
Payment promises — on a call or in a thread — within limits you set: maximum horizon, minimum amount, maximum installments. Outside the envelope, the agent escalates instead of improvising.
Prepare payments for approval
Payment plans and transactions staged with full context and a maker-checker approval waiting — never an unsupervised money movement.
Your team stays in charge
Agents take the volume; humans take the judgment calls. The autonomy ladder decides where that line sits — per task type, on your schedule.
FAQ
The questions every risk committee asks
Yes. Governed AI voice agents handle collections calls end to end, hardship conversations included. Every dial clears the compliance gate first — contact budgets like Reg F's 7-in-7, calling windows in the borrower's timezone, consent, and protected-borrower states. On the call, every financial figure the agent speaks comes straight from the ledger, it negotiates promises to pay within your guardrails, and hardship language routes the case into the hardship workflow. The borrower can ask for a human at any point — and policy can require one — with a warm handoff mid-call. Every call is recorded and transcribed into the case's evidence trail, and voice graduates like every other task type: start with reminder and early-delinquency calls, expand on your evidence.
Put a governed AI agent on your hardest queue
Start assist-only with your own cases and your own policies. Watch every proposal clear the compliance gate, then decide how far up the autonomy ladder to go.