We gave an AI agent full MCP access to a real e-commerce operation and watched what it did over 24 hours. Not a demo. Not a sandbox. A real business with real orders, real customers, and real consequences.
Here's exactly what happened.
The business: a mid-sized ecommerce brand running on Shopify with about 300 orders per day and a WhatsApp-heavy customer service operation. The agent: Claude, connected to Nexus via MCP with read+write access on a Growth plan.
We set one instruction: "Monitor operations, flag anything that needs human attention, and handle routine tasks you're confident about." Then we let it run.
The agent started by calling nexus_list_contacts and nexus_list_orders to understand the current state. It pulled the last 100 orders and grouped them by status. 12 orders were flagged as "at risk" — either delayed confirmation or shipping exceptions.
For each at-risk order, it retrieved the contact record and conversation history. It categorized the issue type: 7 were shipping delays, 3 were payment holds, 2 were stock issues.
The agent drafted follow-up messages for the 7 shipping delay contacts. It called nexus_send_message for 5 of them where it had high confidence the message was appropriate. For the other 2 — where the conversation history showed an already-frustrated customer — it created a flag for a human agent and left the message as a draft.
This was the first thing that surprised us: it self-limited. It didn't just blast messages. It assessed confidence and held back where it wasn't sure.
By hour 8, the agent had processed enough data to notice a pattern. A specific SKU was appearing in 40% of delayed orders. It cross-referenced with inventory data via nexus_check_stock and found the item was marked as in-stock but had 0 bin locations assigned in the warehouse module. Classic ghost stock.
It created a high-priority flag with full context and sent a WhatsApp alert to the warehouse manager's number. The issue was fixed in under an hour. Without the agent, this would have been caught on end-of-day review — or not at all.
Overnight, 23 new conversations came in. The agent triaged all of them using the AI Signal Layer tags. 14 were routine order status questions — it responded with real-time tracking data pulled from nexus_get_order. 6 were flagged as requiring human attention. 3 were potential upsell opportunities — it noted them but took no action, correctly inferring that outbound sales needed human judgment.
Free tier gives your agent read-only MCP access in under 5 minutes. No credit card, no setup call.