Use Cases

How Co-one helped Akbank improve its AI assistant

Akbank worked with Co-one to improve how its banking AI assistant recognizes customer intent and responds. The project combined knowledge base management, automated evaluation, and human review.

  • Türkiye
  • Banking AI
  • LLM evaluation
  • Intent classification
  • Entity annotation
  • Human review
Co-one and Akbank case study cover on evaluating a banking AI assistant.

What did Akbank want to improve?

Akbank’s AI assistant supports customers with loans, investments, and account transactions. When the collaboration began in March 2024, the bank wanted the assistant to identify requests more accurately and distinguish between similar banking products.

How did Co-one evaluate the assistant?

Co-one managed the knowledge base and combined automated evaluation with human review. The team used customer questions, support logs, and interaction records to understand where the assistant misclassified requests and where it needed additional examples.

  • Intent classification. The team initially organized customer queries into more than 300 classes. It refined the categories as new topics and seasonal patterns appeared.
  • Entity annotation. Annotators labeled more than 200,000 entities per month across thousands of categories to help the assistant interpret details in customer requests.
  • Evaluation and reporting. Co-one combined uncertainty estimates with human review to examine predicted intents and report misclassifications. Co-one reported 95% accuracy in its response validation.

How did new customer intents reach the live assistant?

Co-one delivered refreshed data each week. During peak campaign periods, the team identified recurring requests that the existing categories did not cover and recommended new intents. Akbank reviewed the recommendations and added approved intents to the live assistant.

What changed?

Co-one’s project summary reports that by early 2025 it was making about 15% fewer intent corrections than the 2024 average. The summary also reports a 24% increase in accuracy over one quarter.

The workflow gave the team a way to turn recurring misclassifications into changes to the knowledge base, then use automated evaluation and human review to check how the assistant handled them.

More in use cases

MANK and Co-one case study cover on evaluating a banking voice assistant.

Use Cases ·

How MANK evaluated its banking voice assistant

MANK worked with Co-one to automate call-flow testing, assess how its voice assistant handled adversarial conversations, and improve speech recognition and synthesized speech.

Talk to Co-one

Bring a workflow. We’ll start there.

Show us where your agent makes decisions or takes action. We’ll discuss what to simulate, what to validate, and where runtime controls belong.

Book a demo

Choose whether to allow optional analytics and marketing tags.