I built an AI agent that handles customer operations — refunds, order lookups, and support tickets — with a human approval gate built into the workflow. The agent proposes the action, pauses, and waits for a human to approve before anything executes. The LLM never makes the final call. Role-based access and a full audit trail are enforced in code, not prompts. Deployed live on Azure Container Apps.

The pause for human sign off keeps refunds safe while still offloading the lookup work to the agent.

The client's inventory wasn't organized. Products weren't consistently recorded, there was no SKU structure, and the system's stock figures didn't match what was physically on the shelves. That made it hard to know what they really had, what it was worth, and what their cost of sales should be.

What I did

Chose the right tool. I recommended QuickBooks Plus, which handles inventory tracking properly, rather than patching up spreadsheets.

Built the product list. I added every product to the system and allocated a clear SKU to each one.

Sorted the inventory. I organized the stock records so each item could be identified and tracked.

Reconciled to a physical count. I compared the system quantities with the actual stock on hand and resolved the differences.