Manual dialing, all day
Agents spend their shift working down a list — re-reading notes, re-checking balances, re-dialing numbers — instead of resolving the accounts that actually need a person.
Solutions · First-party collections
First-party collections that start when a payment is missed
A missed payment can open and route a case with current account context. AI agents call, draft, offer payment plans, and follow up within approved limits. Every contact passes a policy check before it executes. A kept promise can close the case automatically, with each step recorded.
The signature flow
System
A missed payment opens a case
Compliance Gate
Call-frequency policy checked — 2 attempts remaining
AI Agent
AI calls within approved limits
System
A kept promise closes the case
Every step checked and recorded — full walkthrough below
The status quo
Why collections work fragments
Many first-party teams work across a dialer, policy documents, spreadsheets, and manual reviews. That fragmentation makes prioritization harder and leaves compliance evidence scattered.
Compliance after the fact
Call-frequency policies, calling windows, consent, and protected states are often spread across spreadsheets, policy documents, and systems that are reconciled only after outreach occurs.
Agent burnout
High-volume, low-judgment work drives turnover, and turnover drains the experience your hardest cases need. The team you keep spends its time on the work that needs it least.
A complete case flow
From missed payment to resolved case
Follow one account through case creation, policy checks, AI outreach, human approval, a payment plan, settlement, and closure.
- System
A missed payment opens the workflow
The lending core sends the delinquency event when the account becomes past due, without waiting for a nightly worklist.
Evidence: Delinquency event with the ledger facts that triggered it.
- System
A case opens with current borrower context
The case includes balances, due dates, payment history, prior contacts, consent, channel preferences, protections, and a saved view of the opening facts.
Evidence: Context snapshot pinned to the case record.
- Compliance GateAllowed
Contact and channel rules checked before outreach
A configured call-frequency threshold is computed per debt in real time, including Regulation F presumptions where applicable. Borrower-timezone calling windows, consent, state rules, and protected states are evaluated too.
Evidence: Budget evaluation with the exact rule versions consulted.
- AI Agent
AI agent drafts the outreach
The model drafts the language while balances, dates, and payment figures come from the ledger. The channel can also be an AI voice call.
Evidence: Draft, grounding facts, and model and prompt versions.
- Compliance GateAllowed
Rules are checked again before send or dial
The result is allowed, warning, human approval required, blocked, or missing facts. If required information is missing or stale, the action waits.
Evidence: Decision record listing every rule consulted and its result.
- Human
Human approves the borrower-facing message
Early AI workflows require a person to approve the message. Later, you can allow selected task types to run within policy without changing the pre-action checks.
Evidence: Approval, approver identity, and the autonomy policy in force.
- AI Agent
AI offers a payment plan within approved limits
Set the maximum term, minimum amount, and number of installments. Hardship language opens the hardship workflow, the borrower can ask for a person, and requests outside the limits route to a supervisor.
Evidence: Call recording and negotiation transcript plus every guardrail check.
- System
Promise recorded, follow-up scheduled automatically
The agreed terms are recorded and tracked against the ledger, with the next follow-up scheduled automatically.
Evidence: Promise terms and the scheduled follow-up.
- Ledger
Payment in flight — authorization recorded
Authorization and settlement are tracked separately. Borrower status, fee treatment, and collections suppression follow the lender's configured value-dating policy while ACH remains in flight. The promise stays in pending evaluation until funds clear.
Evidence: Value-dated entry and the suppression record it triggered.
- System
A kept promise closes the case
The platform evaluates the promise against ledger facts as kept, partially kept, or broken. A kept promise on a current account can close the case automatically under your configured rules.
Evidence: Closure decision and the complete, exportable case history.
Division of labor
What the AI does, and what your team does
AI agents work inside the same queues and audit trail as your team, with separately scoped access. You decide which tasks require human judgment and where each autonomy boundary sits.
The AI agent
- Places and answers collections calls — AI voice agents negotiate promises to pay by phone, inside policy bounds: horizon, minimum amount, installments
- Recognizes hardship language mid-call and routes it to a hardship case — with a warm handoff to a human whenever the borrower asks or policy requires
- Drafts every outreach with figures rendered from ledger facts — spoken figures included
- Reads replies, classifies intent, and recommends the next action
- Schedules follow-ups and keeps the case timeline moving — call recordings and transcripts land there too
Your team
- Approve borrower-facing content in early autonomy levels
- Take the conversations that need a person — hardship, disputes, complexity
- Set policy: guardrail bounds, autonomy levels, escalation paths
- Review the manager view: queues, SLA health, escalations
The guardrails
Set clear limits on AI collections work
Configure payment-plan limits, call-frequency policies, human approval, and the actions that must always route to a person.
Promise-to-pay bounds
- Maximum horizon: 30 days out, no further
- Minimum amount as a percentage of what is owed
- Maximum number of installments
Promises then evaluate themselves against ledger facts: kept, partially kept, or broken — pending-evaluation while a payment is in flight.
Example contact budget · trailing 7 days
2 of 7 calls remaining
Attempt countedRemainingVoicemails & limited-content messages count under this policy
This configured budget is checked per debt before each attempt by a person or AI. This is an illustrative view; calling windows, consent, and recorded protections are checked separately.
Check before execution
Every send and every dial is re-checked at the moment of execution. Five outcomes, and only one of them lets the action through:
AllowedWarningHuman approval requiredBlockedMissing facts — action held
If required information is missing or stale, the action waits for updated facts.
The manager view
See the whole collections operation
Managers see queue depth, service levels, escalations, and approvals for people and AI in one command center.
SLA health at a glance
Which cases are aging, which queues are backing up, which promises come due today — without pulling a report.
Smart queues
Work routes by skill, capacity, and case state. People and AI agents draw from the same queues and audit trail, with access scoped separately for each role.
Exceptions needing attention
Gate warnings and cases that need human judgment land in one place — with the full context snapshot attached, not a case number to go look up.
What changes
The outcomes this is built for
These are the outcomes the architecture is built to produce — and what we measure with every design partner.
Faster case starts
A delinquency event can open and route the case as soon as the core reports it, without waiting for a manual worklist.
Policy checks before contact
Each call or message is checked against current contact limits, timing, consent, and recorded protections before execution.
People focused on exceptions
The repetitive middle of collections is automated under guardrails, so your team's judgment goes where judgment is needed.
FAQ
First-party collections, answered
Yes. AI voice agents can place and answer collections calls, including hardship conversations. Each dial passes configured contact-frequency, calling-window, consent, and borrower-protection checks first. Financial figures come from the ledger, payment-plan terms stay inside your approved limits, and the call hands to a person when the borrower asks or policy requires. Calls can be recorded and transcribed where permitted and configured. Voice starts with scoped use cases and expands as evaluation results support it.
Walk one delinquency from missed payment to resolution
Bring a representative scenario from your portfolio. We will trace the case lifecycle, policy checks, human handoffs, and evidence with the founding team.