Most AI banking demos you see today are conversational layers. They answer questions, explain products, maybe fill in a form. When something actually needs to happen in a core system, a human takes over.
We wanted to see what changes when the agent has real access to real banking services. So we extended our own card management system with a set of MCP tools and put an agent on top of them.
Built on our platform, not on top of someone else’s
This is the part that matters most to us. The tools the agent uses are not generic connectors or middleware. They are MCP tools we developed directly on the APIs of our issuing CMS, the same platform that has been running card portfolios for banks across Central and Eastern Europe, Switzerland, Italy and the Middle East for 25 years.
Anyone can build an agent. Building one on a production grade issuing platform, using the same APIs that already serve live banking applications, is a different exercise. It means the agent inherits everything the platform already enforces: business rules, validations, entitlements, product configuration. Nothing had to be reimplemented for the AI layer, because there is no separate AI layer.
What the agent does
The agent covers issuing and dispute functionality. It can issue different card types and open service cases when a card is reported lost or stolen.
More interesting than single operations is what happens when several of them need to be chained together. Until now, a customer reporting a stolen card meant a conversation with a bank officer who knew the sequence: verify the customer, block the card, classify and open the correct service case, then initiate a replacement.
The agent runs that entire flow on its own. It identifies the customer, blocks the card, opens the right service case for the situation, and orders the replacement card. What used to be a call and a queue becomes a single conversation that ends with the problem solved.
One system, two interfaces
The agent is integrated into a demo internet banking application. Because its MCP tools call the same issuing CMS APIs as the application itself, every change the agent makes is immediately visible in the standard user interface.
Block a card through the agent and it shows as blocked in the app. Order a replacement and it appears in the card list. There is no synchronisation, no reconciliation, no eventual consistency to manage, because there is nothing to synchronise. The agent is not a parallel system running alongside the bank. It is an equal user of the same banking services, with the same permissions and the same constraints as any other channel.
That is also what makes this approach realistic for production. A bank does not have to open a new integration surface, duplicate business logic, or maintain a second version of the truth.
Capability is the easy part. Control is not.
In banking, an agent that can do things is only half of the requirement. The other half is being able to prove exactly what it did, when, on whose instruction, and under what authority.
So every critical operation, meaning anything that changes data or starts a business process, is presented to the customer before it executes and requires explicit confirmation. The customer sees what is about to happen in plain language and approves it. The agent does not act on inference or assumption when real money and real cards are involved.
Those confirmations, together with every action the agent executes, are logged in detail. Not as conversation transcripts, but as a structured record of operations tied to the underlying system events. This gives banks what they actually need: traceability for regulatory requirements, evidence for security reviews, and a usable trail for forensic analysis after the fact.
Why we built it this way
AI in banking will not be won by whoever ships the most fluent chatbot. It will be won by whoever can connect language models to core systems without giving up control, auditability or compliance.
That is a platform problem before it is an AI problem. And it is easier to solve when the platform is yours.
Watch the demo below to see the agent handle a full lost card flow, from report to replacement.