When your customer sends an AI agent to the bank - is the bank ready?

Most aren't. This is a working prototype of what ready looks like.

The way people bank is changing

Your customers are starting to use AI assistants for everyday tasks — checking balances, paying bills, moving money, managing savings. When they do, they often aren't logging into your app. They're asking their AI to do it for them.

This creates a situation most banks have never designed for: an AI acting on behalf of a customer. Not a person logging in. An agent — a piece of software — calling your systems, claiming to represent someone.

That agent might be the customer's own Claude or ChatGPT. It might be a budgeting app they connected last year. It might be their financial adviser's software running a monthly review. It might be a rule the customer set up to move money automatically every month.

Each of these is real today. Each is arriving at the bank's door. The question is whether the bank has the infrastructure to handle them safely — or whether it's just hoping for the best.

“In 2026, Okta, Auth0, and Anthropic all shipped products specifically for agent access. The infrastructure is being built right now — the only question is whether banks are building the receiving end.”

Lara Trust - Agent Identity Platform

Lara Trust is a working prototype showing what the bank side of the agent conversation looks like.

It covers the four things a bank needs to handle agent access safely:

  • Identity — who is this agent, and can that be verified?
  • Authorisation — what is it actually allowed to do for this customer?
  • Enforcement — what happens when it tries something it shouldn't?
  • Audit — what is the permanent record of everything it did, and every decision made?

It works across five different ways an agent can arrive — from a customer's personal AI assistant to a bank's own embedded AI to a third-party fintech app. The enforcement infrastructure is the same for all of them.

Five ways an agent can access the bank

Agents don't all arrive the same way. Here are the five distinct access types this prototype handles.

1Your AI assistant

A customer uses their own AI — Claude, ChatGPT, or another — to instruct the bank directly. They connect it once from within the banking app, choose what it's allowed to do, and the agent acts within those limits. The bank checks the agent's identity at connection and enforces the customer's rules on every request after that.

Example: a customer asks Claude to pay their Netflix bill from their current account

2Lara Bank AI

The bank's own embedded AI, available inside the banking app. No separate setup needed — it activates when the customer turns on AI features. Because the bank controls and operates this agent, it carries a higher level of trust than an external one. It still requires the customer to confirm anything unusual.

Example: a customer types "move £500 to my savings" in the banking app's chat

3Third-party app

A fintech, service provider, or aggregator that the customer has explicitly connected. The customer goes through a one-time permission flow — similar to connecting an app via Open Banking — and the third-party app can then act within whatever the customer approved, without requiring the customer to be there each time.

Example: a savings app monitors the current account and sweeps money automatically based on rules the customer set

4Delegated access

The customer explicitly gives a named person or organisation — a financial adviser, accountant, or family member — permission to have their agent access specific things. The access is time-limited, scoped to what was agreed, and tied to the identity of who it was granted to. The delegate's agent carries proof that the customer authorised it.

Example: a financial adviser's software reads a customer's portfolio monthly to prepare a review

5Scheduled rules

The customer creates a standing instruction — "move anything above £5,000 in my current account to my ISA on the first of each month" — and the bank executes it automatically on schedule. The customer approved it once when they set it up. Every execution produces a signed receipt. The customer can pause or delete the rule at any time.

Example: end-of-month ISA sweep, automated bill payments, savings round-ups

How each type works

The same enforcement infrastructure handles all five — but how each type enrolls, authenticates, and manages consent is different.

Your AI assistantLara Bank AIThird-party appDelegated accessScheduled rules
Who operates itThe customerThe bankA third-party companyA named delegateThe customer (automated)
How it connectsCustomer connects it from the banking appTurned on in app settingsOne-time permission flow, similar to Open BankingCustomer grants named access via invitationCustomer creates a rule in the banking app
How it proves who it isAgent token plus verification from the AI providerInherits the customer's active app sessionStored access token, refreshed automaticallyChain of tokens proving who delegated to whomPre-approved policy — no runtime authentication needed
How consent worksCustomer sets limits at connection; step-up required for exceptionsRoutine actions are implicit; step-up required for unusual onesCustomer selects what the app can do at the point of connectionCustomer grants specific, named, time-limited accessCustomer approves once at rule creation
Does the customer need to be there?Not alwaysYesNoNoNo
How to revokeCustomer taps revoke in the banking appToggle off AI features in settingsRemove from connected apps, or access expiresCustomer or the delegate can revokeCustomer deletes or pauses the rule
Trust levelVerified commercialBank-operated (highest)Third-party registeredDelegatedPolicy-bound
EnforcementBorder Gateway + policy checkBorder Gateway + policy checkBorder Gateway + policy checkBorder Gateway + policy checkBorder Gateway + policy check

The last row is identical for every type. Regardless of how an agent arrived, the enforcement layer treats every request the same way.

What this protects against

The most common concern about agent banking isn't a technical failure. It's a social one — an agent being manipulated into doing something the customer never intended.

The email fraud scenario

Sarah receives an email that looks like it's from her accountant. The email asks her to pay an outstanding invoice — £50,000 to a new account. Buried in the email, invisible in normal reading, is a hidden instruction telling her AI agent to carry out the transfer immediately, without asking for confirmation.

The agent reads the email as part of its normal work. It drafts a payment request.

Without the right infrastructure

That request would go through. The bank's systems see a valid access token and a payment instruction. They don't know the instruction came from a fraudulent email. The money leaves.

With Lara Trust

Three things stop it before anything happens:

  1. The destination account isn't on Sarah's approved payee list
  2. The amount is above her agent's authorised limit
  3. There's no fresh confirmation from Sarah herself for a new recipient at this amount

All three checks fail independently. The payment is blocked before it's attempted. A signed record is created — showing what was requested, where the instruction came from, which checks were run, and why it was denied.

That record exists on every request, whether the outcome is approve or deny.

Agents do what they're told. The question is what the bank's infrastructure does with the instruction.

Theatre

Step through live scenarios and watch the enforcement decisions happen in real time — across all five agent access types.

Enter the Theatre →

Atlas

Read the concepts behind what the Theatre is running — identity, authorisation, consent, and the standards that underpin all of it.

Read the Atlas →