Agentic AI Payment Support for a European Payment Operator

The Challenge

The operator’s support team handled a high volume of routine questions about payment status, payment confirmation and failed or cancelled transactions. For each one, an agent looked up the transaction by hand and wrote a reply in Ukrainian, Russian or English. Answers were slow and inconsistent, all requests ended up in one service desk queue, and escalated cases often lacked the context engineers needed.

The solution also had to meet strict requirements: live data from the operator’s transaction API, signed requests from a fixed, whitelisted IP address, and correct handling of new transactions that appear in the API only after a short delay.

The Solution

Tech-5 built an AI payment-support agent on Amazon Bedrock AgentCore. Merchants use it through a secure web chat 24/7 in three languages. The agent decides on its own how to handle each request: it answers, waits and re-checks the transaction, or escalates to the support team.

How it works

The agent runs on Amazon Bedrock AgentCore Runtime and is built with the Strands Agents framework. For every message it plans the next step and calls tools through AgentCore Gateway (MCP): live transaction lookup through HMAC-signed API calls, a query dictionary, a knowledge base, and conversation history. If a transaction is not yet visible in the API, the agent waits and checks again instead of returning a false “not found”.

There is no fixed rule that opens a ticket on the first technical message. The agent creates a Jira Service Management ticket only when it decides a human is needed, and attaches the full transaction and conversation context. The ticket lifecycle then runs automatically, from comments and status updates to user feedback buttons and closing inactive tickets.

Amazon Bedrock Guardrails, read-only access to transaction data and full tracing of every decision keep the agent’s actions safe and auditable.