The problem
An Australian e-commerce business runs procurement, customer service, and factory operations across multiple sales channels and hundreds of SKUs. The system of record is a third-party SaaS where every meaningful answer requires someone to search through tables and stitch the result together by hand.
Monday morning, the procurement lead spends two to three hours on the weekly reorder. By Wednesday, the customer service lead is chasing complaint cases that have drifted past their resolution target without anyone noticing. By Friday, a factory delay surfaces, three days too late to do anything about.
The owner is the bottleneck on every meaningful decision because nothing in the system is trustworthy enough to act on without him verifying it first. He can feel the operation slowing down even as he hires more people.
The turn
The conventional answer is another CRUD tool: better forms, better reports, faster search. That just makes searching a bit quicker.
The real problem is the team is driving the workflow when AI agents could be driving it and the team could be supervising. Reframe the platform from "procurement workspace" to "agent supervision interface". A control tower where AI handles the routine 95% and humans review the 5% that actually need judgment.
The system
A team of specialist agents running on AWS Bedrock and Anthropic's AgentCore Runtime. One agent forecasts demand. Another plans reorders. Another watches for quality complaint patterns. Another chases overdue factory POs. Each is single-purpose; together they cover the operational surface.
The agents read continuously from the company's existing system of record, so nothing has to change for the operations team's data layer. The work appears in a single inbox: what each agent did, what needs the human's eye, what was proposed and waits for sign-off. Every recommendation comes with the reasoning shown, so the operator can argue with it rather than rubber-stamp it.
Nothing writes back to the system of record without the operator approving it. There's a kill switch the owner can hit during incidents that pauses all writes while reads continue. The UI is bilingual English and Chinese from day one because the team operates in both.
Status
Active build. Multi-agent procurement and quality team going live across the e-commerce client's ERP. Weekly reorder collapsing from 2-3 hours to 15 minutes.