BBVA cuts average handling time by more than 15% in Italy and Germany with generative AI copilots

BBVA customer service agents in Italy and Germany answer nearly 100,000 customer enquiries a month by telephone, email and the 'My Conversations' messaging channel on the app and website. To do so they previously had to consult multiple internal documents and product and service manuals specific to each market, a time-consuming task that meant searching through multiple sources for each case.

Teams managing the contact centres of BBVA's two fully digital banks in Italy and Germany built two generative AI assistants on BBVA Assistants, the Group's corporate generative AI platform, which combines OpenAI's models with access control, knowledge management and supervision mechanisms. Agents submit a customer's request in natural language, and the copilot turns the bank's internal knowledge into a structured response proposal in the customer's language within seconds. The agent reviews and adapts the response before sending it, and the tool has no access to customers' personal or sensitive data. Before going into production the assistants were assessed under BBVA's AI Assistant Governance Framework, and agents received training on writing prompts in natural language.

The two assistants reduced the average handling time for the most common customer enquiries about products, services and general procedures by more than 15%. More than 260 customer service agents across both countries use them daily, and in June alone agents exchanged around 40,000 messages with the copilots, making them among the most widely used assistants on BBVA Assistants. Routine enquiries are resolved faster, leaving agents more time for complex cases that need deeper analysis and personalised support.

• Start from problems the front-line teams raise themselves; here the need came directly from the people handling customers.
• Keep agents as the final reviewer of every AI-drafted response and keep personal customer data out of the tool.
• Put assistants through a governance review before launch and train agents to write good prompts, which supports accuracy and adoption.