BrazilFinance and insurance2024 – ongoing

PagBank (PagSeguro)

Customer-service chatbot with image reading and generative fallback answers

PagBank said its AWS-based chatbot answered about 85% faster than human agents in a two-month pilot

WorkingCompany's own claim

Problem

Replacing a faulty payment machine was one of the most common reasons customers contacted PagBank. Each case needed an agent to identify the device and its error.

Approach

PagBank built a chatbot on AWS. The customer sends a photo of the machine screen, and Amazon Rekognition reads the serial number to identify the device and the error. Amazon Bedrock answers out-of-scope questions from the bank's FAQs. The bank planned to extend the approach to other contact reasons.

Results

85%average reduction in chatbot response time compared with human service, about two months after the pilot, as reported by PagBankCompany's own claim[1]

more than 90%accuracy of the generative fallback answers based on FAQs, reported without a named sourceCompany's own claim[1]

Takeaway

A narrow, frequent request, here machine replacement, gave the bank a clear first use for image reading and generated answers. The figures are the bank's own and cover a pilot of about two months.

Sources

PagBank inova no atendimento ao cliente com inteligência artificial da AWS
ABC da Comunicação28 Oct 2024News article

Source summaries

[1]ABC da Comunicação28 Oct 2024

Article on the PagBank and AWS chatbot for payment-machine replacements, using Rekognition for serial numbers and Bedrock for fallback answers. It gives an 85% shorter response time versus human service and fallback accuracy above 90%. It says recontacts fell significantly but gives no number.

Full details are in the original sources.

Added 1 Nov 2026 · Updated 1 Nov 2026