Most people have met AI through a chatbot, and inside a bank the reaction is that it looks interesting. What goes unexamined is that putting one in front of clients is a bigger thing than serving information on a bot. It is mostly regulatory: transparency, and claims against bad financial advising. The capacities have to be increased before going out.
The European AI Act applies in full from 2 August 2026. A system that interacts with a person has to disclose that it is a machine, offer a route to a human, and leave an audit trail complete enough to reconstruct a complaint afterwards. Chapter 9 alignment makes the Act the planning baseline in Albania well before accession formally binds anyone to it. The financial-literacy baseline here is low enough that a customer may not register what they are talking to, which gives the disclosure obligation more weight, and the complaint the bank eventually has to answer is about advice a machine gave on money.
An internal assistant carries none of that. A model answering an employee's question about the bank's own procedures sits outside the Act's high-risk tier and outside its transparency article. A new person getting a faster answer about a procedure is the whole of the return, and internal chatbots can solve many problems with lower risk of being wrong, in a period when information, capacities and understanding are all still increasing.
Fraud detection and AML facilitation have shown direct and measured benefits of AI in the bank. Banks running AI on transaction monitoring and sanctions screening are effectively operating in the Act's high-risk tier, a heavier classification than the customer chatbot's Article 52 transparency duties. On the AML desk an analyst reads every alert the model scores before the bank acts on it.
The customer-facing move comes later. AI belongs on the AML desk first, and earns the chat window by doing that job well.