AI customer assistant for banking and financial services: what is built today, and how account questions would work
Honest scope for banks: a public branch and ATM locator exists today; balance and statement questions would use the same identity-first mechanism.
Banks get two kinds of contacts. Public questions: where is the nearest branch, when is it open. And account questions: my balance, that transaction, my statement. They need very different safeguards, and the platform treats them differently.
What exists today, and what does not. Built and running: banking.atmsByCity, a public ATM and branch locator over open data. Not built: a connector to any bank's core system. The account examples below describe how the existing mechanism would be configured; they are not a deployed product, and no bank is named.
What a customer asks, and how it would be answered
- "Where is the nearest ATM in my city?" Answered today, from open data, with no personal data involved.
- "What is my balance?" Identity is resolved on the server from the session and is never a parameter the model or the client supplies. A balance would need a sign-in strong enough for that data, as a tax balance needs today (assurance levels explained, a tax balance without the call centre).
- "Stop this card" or "move this payment." A change never runs on the model's say-so: the assistant shows exactly what it will do and carries out only what the customer confirmed (confirmed before it runs).
- A person, when it matters. Every staff view of customer data is an audit record with an actor and a purpose.
Where the money moves
The bold rows are the levers a bank leans on most. Rows describe the mechanism, not a live bank deployment.
| Lever | What Kav does here | What you measure |
|---|---|---|
| Fewer repeat contacts | Branch, hours and ATM questions are answered from open data today; account lookups would follow the same pattern once a connector exists. | Contacts that are pure locate or opening-hours lookups. |
| Faster when a person steps in | A handover carries the conversation, so the agent does not start from zero. | Handling time on transferred contacts. |
| Out of hours, every channel | Public questions are answered when branches and phone lines are closed. | Contacts outside opening hours. |
| Fewer errors and rework | A figure comes from the system of record through a template; the model never writes one. | Corrections to quoted balances or dates. |
| Revenue kept | Approved product information can be answered from your own published content, with a person one tap away. | Product enquiries that reach a person already qualified. |
| Compliance and audit | Server-side identity, step-up by assurance level, per-operation permissions and an audit record for every staff view. | Effort to answer an auditor's or regulator's access-log request. |
Work out your own number
- Count public locate and hours contacts in your own log for four weeks; those are the ones a locator answers today.
- Cost per contact = agent hours × loaded hourly cost ÷ contacts handled.
- Value = those contacts answered without a person × cost per contact. Treat any account-question saving as a later phase that depends on a connector to your systems.
Work out your break-even from the published prices, or talk to us about what a connection to your systems would involve.