Catalog investigations cut from days to 15 minutes by an AI agent on Bedrock

The challenge

This distributor of promotional products and branded merchandise runs a catalog of hundreds of thousands of parts, fed from supplier systems on a five-minute sync. When one of those parts quietly failed to land, nothing flagged it. The first sign was usually a customer who could not find a size or color they expected.

Every report started a manual investigation. Customer service wrote a ticket, the ticket waited in a queue, and an engineer queried the catalog, the search index, and the supplier feed by hand. That took days, and the next refresh often cleared the symptom before anyone found the cause, so the same failure came back.

What we did

Trellist designed and built a catalog agent that customer service can ask in plain language. It runs inside the distributor's own AWS account, on the patterns its engineers already maintain.

• Plain-language intake. Customer service describes the problem where they already work. The agent parses the supplier and product from the text and decides which checks the description calls for.
• Six tools, five of them read-only. The agent can read what the catalog is serving, what the product system says is approved, how parts are grouped for display, and what the supplier offers live. Its one write tool rebuilds product groups, scoped to a single index.
• Claude on Amazon Bedrock. The agent reasons with Claude Sonnet 5 on Amazon Bedrock, behind API Gateway, Lambda, SQS, DynamoDB, and CloudWatch. The model can be swapped later and the tools stay in place.
• Answered in the same conversation. Every request is authenticated at the door and acknowledged right away, and the answer comes back in the same thread when the work is done.
• A full audit trail. Every diagnosis is logged with who reported it, which tools were called, the state before and after, and the verdict that came back.

How the agent decides

The agent classifies what it found before it changes anything. Only one of its four verdicts is allowed to write, and when it does, the agent re-reads the catalog to prove the fix worked before it calls the issue fixed.