Banks, lenders, and servicers

AI worker teams for lending and servicing operations

Lending operations rarely fail on capability. They fail on evidence — the inability to prove, months later, that a contact was permitted, an identity was checked, and an offer was inside policy. That is the problem worth solving first.

BANKS, LENDERS, AND SERVICERSCONSTRAINT ENVELOPE4 sector constraints designed to firstHUMAN AUTHORITY — NEVER AUTOMATED· Hardship and forbearance· Dispute adjudication· Anything affecting credit sta…ACTS WITH APPROVALThresholds and gates on anything with cost or risk attachedACTS WITHIN POLICY2 blueprints start hereAutonomy widens outward only on evidence from the recordearned autonomy
The authority boundary we design to in Financial services — autonomy widens outward from the centre, and only on evidence.

Our view

Where we think the real opportunity is

In most sectors the question is whether AI can do the work. In financial services the work is comparatively simple and the burden of proof is the hard part. A collections conversation is not difficult to hold; what is difficult is guaranteeing that ten thousand of them respected consent state, permitted hours, identity thresholds, and the approved offer matrix — and being able to demonstrate it to an examiner from a record you did not assemble by hand. That inverts the usual design brief. We do not start from what the worker can say; we start from what the workflow will refuse to let it do, and treat the audit trail as the primary output rather than a by-product.

Operating reality

How demand actually behaves here

Volume has a shape, and the shape is what breaks teams. This is what we assume about your operation before we design anything.

  • Demand is calendar-shaped, not random

    Statement cycles, due dates, and month-end drive volume, which means the work is forecastable and the contact strategy is a scheduling problem before it is a conversation problem.

  • Right-party contact rates are low

    Most attempts do not reach the account holder. Human agents therefore spend the majority of their day on activity that produces nothing, and the value only appears in the minority of conversations that connect.

  • The same customer is contacted by several teams

    Servicing, collections, retention, and marketing often operate from separate systems with separate suppression logic, which is how a customer receives three contacts in a week and a complaint follows.

  • Channel preference is generational and stark

    Some segments will never answer an unknown number but will reply to a message within minutes. Contact strategy that ignores this reads as diligence and performs as harassment.

Where the hours go

The cost we would try to move first

Almost every operation spends its most expensive time on its least valuable work. Naming that precisely is what makes a pilot measurable rather than impressive.

  • Verification before value

    A trained agent spends the opening minutes of every call establishing who they are speaking to, before any conversation that could resolve the balance begins.

  • After-call wrap and disposition

    Outcome coding and note-writing happen after the customer has hung up, which is both unpaid conversation time and the point at which the system of record starts to diverge from what was agreed.

  • QA that can only sample

    Manual review covers a small fraction of conversations, so compliance confidence is statistical rather than actual, and a systematic script problem can run for weeks before sampling finds it.

  • Rework from unlogged promises

    An arrangement agreed on a call but recorded incorrectly generates a broken promise, a further contact, and often a complaint — all downstream of a data-entry failure.

The trust envelope

What constrains the design, and what never gets automated

We design to the boundary first. A control added after launch is a control nobody believes — and in this sector the boundary is not negotiable.

Sector constraints

  • Consent state and revocation

    Permission to contact is not static. It changes mid-relationship, per channel, and a revocation must suppress future attempts everywhere immediately rather than at the next batch refresh.

  • Permitted hours and contact frequency

    Timing rules vary by jurisdiction and product. These belong in the workflow that decides whether to place an attempt, not in guidance an agent is expected to recall under pressure.

  • Identity confidence before disclosure

    Balance and account detail cannot be discussed until identity is established to a defined threshold, and the threshold itself has to be a configured gate rather than a judgement call.

  • Duty of care to customers in difficulty

    Vulnerability and hardship signals carry an affirmative obligation to change approach. Recognising them reliably matters more than resolving the balance.

Stays human, permanently

These are structural boundaries, not trust levels waiting to be relaxed.

  • Hardship and forbearance

    Deciding to vary terms for a customer in difficulty weighs circumstances a transcript does not contain. It belongs to a trained specialist, and the worker's job is to reach one quickly.

  • Dispute adjudication

    A disputed balance is a factual disagreement about the record. Arguing it is how complaints are manufactured, so the worker records the dispute and stops.

  • Anything affecting credit standing

    Decisions with consequences for a customer's credit file need a person accountable for them, by name, in the record.

  • Complaint handling

    Once a customer frames a contact as a complaint, the regulated clock starts and the process is a human one from that point forward.

Where to start

Sequencing matters more than scope

The first blueprint should be the one whose success criteria your team already agrees on. Each step below links to the blueprint that implements it.

  1. Start

    Early-stage recovery, suggest-only

    One portfolio segment where the outcome is unambiguous and the compliance surface is smallest. Workers propose, your agents send. The value of this stage is the scored baseline it produces, not the collections lift.

    Read the blueprint
  2. Then

    Widen autonomy inside the offer matrix

    Once the record shows consent and identity gating holding, let workers own contact, verification, and in-matrix arrangements. Every exception still routes to a supervisor with the conversation attached.

    Read the blueprint
  3. Next

    Reuse the governance for onboarding

    New-product and new-client intake needs the same identity gating and data minimisation you have already had reviewed, which makes the second blueprint a configuration exercise rather than a fresh approval cycle.

    Read the blueprint

Systems we would expect to read and write

  • Core banking or loan servicing platform
  • Collections and recovery system
  • Payment gateway and mandate management
  • CRM and complaint case management
  • ViciDial or Twilio dialler infrastructure

Blueprints that apply

2 operating blueprints for Financial services

Each one carries the full team, process, and control detail — the mechanics this page deliberately does not repeat.

Push back on this

Objections we actually hear in Financial services

Answered as we would answer them in the room, including the cases where the right answer is that this is not for you.

Our regulator will not accept AI contacting customers about debt.
The position we would take into that conversation is not "trust the model" — it is that consent, permitted hours, identity thresholds, and the offer matrix are enforced by the workflow before a worker speaks, and that every check and offer is recorded immutably. That is a stronger control environment than a human agent working from training and memory, and it is auditable at full volume rather than by sample. If your regulator has ruled on this specifically, that ruling governs, and the blueprint still applies with workers in suggest-only mode behind your agents.
Our QA team cannot review the volume this would produce.
They should not have to. Quality scoring runs on every conversation rather than a sample, so QA moves from listening to calls to reviewing exceptions and tuning thresholds. The realistic change is that your QA function stops being a sampling exercise and starts being a policy function.
Customers hate automated calls about money.
They hate being chased ineffectively — repeat calls at bad times, repeated verification, no memory of the last conversation. The blueprint attacks exactly that: contact on the channel the customer actually answers, context carried across attempts, and immediate routing to a person the moment hardship or dispute appears. Where the customer wants a human, the fastest path to one is the design goal.

Other sectors

How the argument changes elsewhere

Next step

Argue with this in a working session.

Bring your own numbers. We will map your version of the operating reality, agree what must stay human, and scope the first blueprint against a measurable win — or tell you it is not worth doing yet.