Letting AI touch money — carefully

Ask anyone who runs servicing at a lender what keeps them up at night, and it’s rarely volume. It’s the fear of getting one regulated action wrong — at scale. Freezing a card, setting up a payment plan, quoting a payoff: each is routine, and each has rules that don’t bend. “You can’t just point a chatbot at a core banking system and hope,” is roughly how their Head of Servicing put it.

Which is why this project didn’t begin with a demo. It began with a list.

Before a single customer conversation was automated, the team wrote down what could not be compromised. Everything else was negotiable; these were not:

• Verification first — no action of any kind before confirming who’s asking.
• Consent and contact rules — honoured in the moment, not reconciled afterwards.
• A human on anything sensitive — hardship, disputes, and anything outside policy.
• An audit trail on every action — replayable, with no exceptions.
• Data in region — residency was not up for discussion.

Only once those were wired in — and enforced by the system, not by a training deck — did the first live conversation go through.

Trust was earned in stages, not granted. For the first few weeks Wiztra ran in shadow mode, drafting each action for a human to approve. Then approvals were relaxed for the lowest-risk requests, with a compliance officer reviewing a sample of the audit logs every week. The reviews got shorter. The sample stayed clean. That, more than any accuracy score, was what moved the needle internally.

Everyday servicing is resolved in the moment now — compliant by construction — while officers spend their time on hardship and exceptions, the cases that actually need judgement. The rules are enforced as the agent works, and every step is on the record. A compliance burden became a controlled, auditable workflow, which is a sentence the servicing team never expected to say about automation.