Bank of Georgia’s GenAI banking assistant now resolves 78% of customer queries with no human involved, at 91% satisfaction.

Bank of Georgia is a retail bank serving customers across Georgia. It has embedded a GenAI banking assistant directly into its mobile app. The assistant now resolves more than 78% of customer queries without a human agent. Satisfaction with those answers sits at 91%. Meanwhile, 53% of users open the app every day.

What the GenAI banking assistant replaced

Most banking apps route customer questions to human agents. Costs then climb as query volume grows, and customers wait longer. Bank of Georgia targeted two high-volume moments instead. First, the assistant handles routine questions inside the app. Second, the bank automated its unsecured lending end to end. As a result, 95% of those loans are approved and paid out in under 15 seconds.

Where the GenAI banking assistant stops

Chief Digital Officer Giorgi Vakhtangishvili draws a firm line on what runs unsupervised. Routine, reversible actions sit with the assistant. However, anything irreversible still needs a human tap. A large transfer qualifies, and so does a new credit line. He expects GenAI to go further and start generating the interface itself, assembling around what a customer needs in the moment.

Why the results drew attention

The approach won the bank nine categories in Global Finance’s 2026 World’s Best Digital Banks awards. The pairing of high resolution rates with high satisfaction is what stands out. Many deployments achieve one or the other. Other consumer businesses have found similar gains by concentrating GenAI on a few frequent interactions, as GoKwik’s messaging results showed.

Why It Matters

  • The reversibility line is the design decision worth copying here. Decide which actions a customer can undo, then automate only those.
  • Daily app opens tell you more than transaction counts. That metric shows whether an AI interface became a habit or stayed a novelty.
  • High resolution rates paired with high satisfaction is the harder claim to earn. Test which of your own query types are safe to automate that fully.
  • The value here clustered around a few very frequent actions. Spreading GenAI thinly across many small features rarely produces numbers like these.