Collections & payment recovery

Recover overdue balances without losing the customer

Collections punishes inconsistency. The same customer gets contacted twice one week and missed the next, verification is repeated from memory, and outcomes reach the system of record hours later. This blueprint runs contact, verification, and resolution as one supervised process where consent state and approval thresholds are enforced by the workflow, not recalled by an agent under time pressure.

AI WORKERSORCHESTRATED PROCESSHUMAN AUTHORITY3 roles, trust-scoped6 stages, each with an owner2 decision ownersOutreach workerVerification workerResolution worker1Select & checkCollections super…2ReachOutreach worker3VerifyVerification work…4Understand inte…Resolution worker5Resolve or esca…Resolution worker6Write backResolution workerHardship & disput…Collections super…CONTROLS AROUND EVERY STAGE4 guardrails · 4 decision rules · 4 channels · tamper-evident audit
The blueprint at a glance — worker roles feed the orchestrated stages, and the dashed paths are the escalations into human authority.
The team
3 AI worker roles escalating into 2 human decision owners.
The process
6 orchestrated stages, each with a named decision owner.
The channels
Voice, SMS, WhatsApp, Email on one shared context.

The shift

What breaks today, and what changes

The challenge

  • Agents spend the day dialling and repeating identity checks instead of resolving the accounts that can be resolved.
  • Consent, DNC, and contact-frequency rules live in training and memory, so compliance depends on the individual agent having a good day.
  • Promises to pay and dispute notes are written up after the call, so the system of record lags behind what was actually agreed.

What changes

  • Every contact attempt is checked against consent, DNC, and frequency rules before the call is placed.
  • Customers who can self-resolve get an approved plan in the first conversation, on the channel they answer.
  • Hardship, dispute, and vulnerability signals stop the collections path immediately and reach a trained specialist with the full history.

The team

Every seat has one job and a defined level of autonomy

This is what gets configured in Team Canvas: role-scoped AI workers, the humans they escalate into, and a trust level on each connection that sets what a worker may do on its own. Trust starts tight and widens on evidence.

AI workers

  • Outreach worker

    Opens the conversation at a permitted time on the channel the customer actually answers.

    Owns
    Contact attempt and consent state
    Trust level
    Acts within policy — blocked entirely outside consent and permitted-hours rules
  • Verification worker

    Establishes identity confidence before any balance or account detail is discussed.

    Owns
    Identity confidence threshold
    Trust level
    Acts within policy — cannot disclose account data below the configured threshold
  • Resolution worker

    Explains the balance, presents approved arrangements, and captures the commitment.

    Owns
    Offers inside the approved matrix
    Trust level
    Acts with approval — anything outside the plan matrix needs a supervisor

Human decision owners

  • Hardship & disputes specialist

    Takes vulnerability, hardship, and disputed-balance conversations with full context attached.

    Owns: Forbearance and dispute decisions

  • Collections supervisor

    Reviews scored conversations, approves exceptions, and tunes the policy boundaries.

    Owns: Approval thresholds and quality standard

The process

The conversation starts the work. The workflow finishes it.

Each stage records the decision that was made and the role accountable for it — which is what makes the run reviewable afterwards rather than a black box. Where the platform executes a stage rather than a person or agent performing it, that is named too, and accountability still sits with the role.

  1. 1

    Select & check

    Is this account eligible to contact right now under consent, DNC, and frequency rules?

    Collections supervisor

    Run by workflow

  2. 2

    Reach

    Which channel and time has this customer actually responded to before?

    Outreach worker

  3. 3

    Verify

    Is identity confidence high enough to discuss the account at all?

    Verification worker

  4. 4

    Understand intent

    Pay now, arrange a plan, dispute the balance, or declare hardship?

    Resolution worker

  5. 5

    Resolve or escalate

    Is the requested arrangement inside the approved matrix, or does it need a person?

    Resolution worker

    Run by workflow

  6. 6

    Write back

    Which record is created — arrangement, dispute case, or do-not-contact update?

    Resolution worker

    Run by integration

Decision rules

The non-negotiables encoded in the workflow, not left to a prompt.

  • Any vulnerability or hardship signal stops the collections path on the turn it appears and routes to a specialist.
  • Disputed balances are never argued: the worker records the dispute and hands it to a human handler.
  • Arrangements outside the configured plan matrix require supervisor approval before they are offered, not after.
  • A revoked-consent or DNC signal is written back immediately and suppresses every future attempt across all channels.

Controls & guardrails

What makes this safe to run in production, and provable afterwards.

  • Consent state and DNC lists enforced before dial, not checked after.
  • Identity confidence thresholds gate every disclosure of account detail.
  • Approved script and offer boundaries, with supervisor approval required to exceed them.
  • Tamper-evident record of every attempt, disclosure, offer, and escalation.

Platform capabilities

What this blueprint runs on

Nothing here is bespoke. Each blueprint is a configuration of the same platform, which is why the second one you launch reuses the governance you already reviewed.

Channels

  • Voice
  • SMS
  • WhatsApp
  • Email

Systems it reads and writes

  • Collections & loan servicing platform
  • CRM
  • Payment gateway
  • ViciDial or Twilio telephony

Measurement

What we instrument from day one

These are the measurements the pilot puts in place, not benchmark results. You set the targets against your own baseline — and the same instrumentation is what decides whether a worker's trust level widens.

Contact-to-resolution

Instrumented per attempt

Attempts, reaches, and resolutions are tracked per account so you can see which contact patterns actually resolve rather than just dial.
Promise integrity

Agreed vs kept

Every arrangement is written back with its terms, so kept-rate is measured against what the worker actually offered.
Every action

Audit coverage

Consent checks, identity outcomes, offers made, and escalations are hash-chained for regulator-ready review.

Rollout

How this one goes live

A deliberately narrow start, supervised, with autonomy widened per role once the record supports it.

  1. Weeks 1–2

    One portfolio segment, suggest-only

    Point the blueprint at a single early-stage segment. Workers draft and propose; your agents send and confirm. The scored transcripts become the baseline.

  2. Weeks 3–4

    Act within the approved matrix

    Workers take contact, verification, and in-matrix arrangements end to end. Every exception still routes to a supervisor with the conversation attached.

  3. Week 5+

    Widen the portfolio

    Extend to further segments and channels once quality scores and audit review hold. Trust levels are relaxed per role on evidence, not on schedule.

Questions we get asked

Collections & payment recovery, in practice

How does the AI team stay compliant with collections regulations?
Consent state, DNC lists, and permitted contact hours are enforced by the workflow before an attempt is placed, so a non-compliant contact cannot be initiated. Approved language boundaries constrain what the worker can say, arrangements outside the configured matrix require human approval, and every check, disclosure, and offer is written to a tamper-evident audit trail you can export for review.
What happens when a customer says they cannot pay?
Hardship and vulnerability signals stop the collections path on the turn they appear. The worker does not negotiate, argue, or continue the script — it hands the conversation to a trained human specialist with the full transcript, account context, and the reason for escalation already attached.
Can this run on our existing dialler?
Yes. Voice Operations runs natively on ViciDial and Twilio, so outbound contact uses the telephony and campaign structure you already operate rather than requiring a parallel phone stack.

What changes in Financial services

Consent state, identity confidence, and approval thresholds become part of the workflow rather than a script the agent is asked to remember.

Read the full Financial services analysis

Related blueprints

Next step

Build this team around one measurable win.

We map your version of this process, compose the worker roles, connect your systems, and agree the escalation points with the people who own them — then run it supervised.