Financial Services · In discussion
An AI virtual assistant that an auditor can read
Front-line servicing conversations handled by AI avatars and virtual receptionists, with controlled workflows, human escalation on anything sensitive, and a decision log behind every interaction.
- Confidential
- under NDA
- In discussion
- engagement stage
What we did
- AI virtual assistant
- AI avatars and virtual receptionists
- Auditable decision logging
Built with
- Controlled workflow engine
- Human-in-the-loop escalation
The challenge
Branch and contact-centre teams were absorbing routine servicing questions by hand, with every automated answer needing to be explainable to an auditor.
What we built
An AI virtual assistant fronted by AI avatars and virtual receptionists, handling front-line servicing conversations with controlled workflows, human escalation on anything sensitive and a full decision log behind each interaction. Engagement in discussion; details confidential.
This engagement is in discussion and covered by confidentiality. The system is described here; the client is not claimed as a reference and no performance figures are published.
Banking support has a shape that resists ordinary automation. Most of the volume is routine — balances, statements, card status, branch hours — but the answers sit next to questions that must never be handled by a machine alone, and every answer has to survive being read back by a compliance reviewer months later.
The constraint that drove the design
The interesting requirement was not accuracy. It was explainability. For any interaction, somebody has to be able to ask: what did the customer say, what did the assistant decide, on what basis, and what would have happened had it decided otherwise.
That rules out an architecture where a model free-associates its way to a response. It calls for a bounded set of workflows the assistant is allowed to run, with the model handling language and intent rather than policy.
What the system does
- Fronts the conversation. AI avatars and virtual receptionists take the first contact across voice and chat, so routine servicing never queues behind a human.
- Runs controlled workflows. Each supported request maps to a defined procedure with defined inputs, not to an open-ended generation.
- Escalates on anything sensitive. Disputes, hardship, fraud signals and anything outside the supported set hand off to a person with the full context attached, rather than restarting the conversation.
- Logs the decision, not just the transcript. Every interaction leaves a record of the path taken and the inputs behind it.
Why it is structured this way
An assistant in a regulated environment is judged on its worst interaction, not its average one. Narrowing what the system is permitted to do — and making the boundary explicit rather than implicit in a prompt — is what makes the rest of it deployable at all.