Use cases
The work that cannot ship without evidence.
Plenty of teams can put a language model in front of a customer. The hard part in a regulated industry is everything after that: proving what the agent was allowed to do, what it actually did, who reviewed it, and that the version you shipped passed its worst cohort before release. These pages describe the use cases where that burden decides the architecture, and how the same set of controls covers each one.
Every use case below runs on the same primitives: agents declared as YAML plus Markdown, your own APIs registered as typed Python tools, guardrails enforced as config on every step, a hash-chained audit trail, human handover with a resume path, and eval cohorts that gate the release. Nothing leaves your network.
Banking & lending
High interaction volume, hard conduct rules, and a supervisor who will eventually ask for the transcript.
Banking customer support agents
Balance queries, card controls, statement questions, and payment troubleshooting, answered from your core systems rather than from model memory, with third-party disclosure hard-blocked and a human one step away.
tools · guardrails · handover
Loan collections and repayment assistance
Contact windows, per-customer contact caps that count across channels, hardship detection that escalates instead of pressing, and a record of every message for the day someone alleges harassment.
contact caps · windows · audit
Fraud alerts and dispute intake
The customer taps "I did not make this" and the agent verifies the transaction, gathers a clean statement of facts, and files the case, never deciding the outcome and never revealing another customer's data.
tools · never-rules · handover
KYC and onboarding operations
Document chasing, mismatch explanation, and case preparation for the analyst who makes the actual decision, with PII masked before any model call and every step attributable after the fact.
redaction · oversight · audit
Beyond banking
The same evidence problem shows up wherever an automated decision touches a person's money, health, or service.
Insurance claims and policy service
First notice of loss, status chasing, and renewal questions handled end to end, with coverage decisions left to a human and a per-claim record of what informed each step.
claims · oversight · evals
Healthcare patient and billing support
Appointment logistics, billing and claim status, and coverage questions, with clinical advice refused by policy, identifiers masked before any provider call, and the whole exchange retained inside your own perimeter.
redaction · refuse_topics · self-hosted
Telecom customer operations
Plan changes, billing disputes, port-out retention, and outage updates at call-centre volume, with retention offers bounded by policy rather than by a model's enthusiasm.
tools · guardrails · scorecard