How do you handle dual-control for transactions above our threshold?
Thresholds are configured per legal entity, per workflow, and per channel. Any action that crosses the ceiling — provisional credit, fee waiver, wire confirmation, account opening — is queued for a named human authorizer in your existing escalation system (typically Salesforce FSC queues, or Pega cases). The agent never proceeds on its own; the customer is told the wait time, in plain language, and called back when the authorizer signs off.
Can the agent take action in our core banking system, or only read?
Both. Native read/write into Finacle, Flexcube, Fiserv DNA, FIS Profile, and TCS BaNCS — with the same SSO, the same RBAC, and the same audit posture your existing teller workstations use. Write actions are scoped per tool: the disputes agent can issue provisional credit up to the configured ceiling and cannot, for example, open an account. Scope changes flow through your existing change-management process, not ours.
What happens to a call recording under DPDP, GDPR, and GLBA?
Recordings are encrypted at rest under your KMS keys, pinned to the region you specify (EU, US, IN, GCC, etc.), and retained per your existing record-retention policy. PII is redacted pre-LLM; cardholder data is masked before it enters any model context, in line with PCI-DSS Requirement 3. Right-to-erasure requests under GDPR and DPDP are honored through the same workflow your CRM already supports.
How do you avoid hallucinated rates, fees, or product terms?
The agent does not generate rates. It retrieves them from your system of record — your product master, your pricing engine, your card platform — through scoped tools. If a customer asks something outside the retrieved scope (e.g. a competitor's rate, an unsupported product), the agent says it cannot answer and routes to a human. Compliance-critical fields (APR, fees, disclosures) are pulled by deterministic tool calls, not generated.
What does the audit trail look like for our second line of defense?
Every session emits a signed, immutable record: the channels touched, the tools called with arguments and return values, the policies evaluated, the disclosures read, the model trace, and the final disposition. Records are exportable to your existing GRC platform — MetricStream, Archer, ServiceNow IRM — on a schedule your team configures. We have walked a US OCC examiner and an Indian RBI inspection team through this exact log without remediation findings.
Can you deploy on-prem or inside our existing VPC?
Yes. Default deployment is single-tenant in your AWS, GCP, or Azure account, region-pinned, with bring-your-own KMS. We support fully on-prem deployments on customer-managed Kubernetes for banks under regulatory direction (RBI master direction on outsourcing, ECB DORA ICT third-party rules). Air-gapped builds are available for cases where outbound telemetry is prohibited.
Which LLM are you actually using — and can we bring our own?
Repliant is model-agnostic. Default deployment runs on whichever frontier model your tenant prefers (we currently have production deployments on Anthropic, OpenAI, and self-hosted Llama-derivatives). Bring-your-own-LLM is supported — your fine-tuned weights run inside your tenant, with the rest of the Repliant orchestration layer wrapping them. Model swaps do not require rewriting agents.
How do you handle vulnerable customers — and complaint detection?
We ship with a configurable vulnerability-detection layer mapped against the FCA's vulnerable-customer framework and the CFPB's consumer-complaint typology. Markers (hardship language, distress cues, accessibility flags, complaint terms) trigger immediate human handoff with full session context. The agent does not attempt to resolve a complaint; it logs, escalates, and apologizes — in that order.