Luna.AI in Production: Lessons From a Live Banking Assistant
Shipping a banking assistant into a demo is easy. Running one that talks to real customers, about real money, every day, is a different discipline. Six months in, here is what actually mattered.
Know what you do not know
The single most valuable behaviour was calibrated uncertainty — the assistant reliably recognising when it was out of its depth and handing off cleanly. Customers forgive "let me get a specialist" far faster than they forgive a confident wrong answer.
Escalation is a feature, not a failure
We stopped measuring success as "contained without a human" and started measuring resolution and trust. A smooth hand-off that carries full context to an agent often produces a better outcome than a fully automated but mediocre one.
Grounding beats fluency
Every answer that touched a balance, a rate or a policy was grounded in the bank’s own systems and documents, with citations. Fluency without grounding is where hallucinations — and complaints — come from.
- Calibrated uncertainty earns more trust than confidence
- Measure resolution and trust, not just containment
- Ground every factual answer in the bank’s own systems
See it in your own environment
Request a tailored demo of the products behind this research.
