A Sydney pet brand now has a team of AI agents running their Amazon business

The problem

Running an Amazon FBA business well requires pulling data from several places — ad performance, inventory reports, settlement data — and doing the analysis yourself to understand what it means. Is this campaign performing or bleeding? Is that product about to run out of stock? Did last week's settlement match what we expected?

None of those are hard questions. But answering any one of them from scratch takes time and context-switching that adds up across a week. Before you factor in the operational work underneath — sending review request emails after each order, handling the bookkeeping, tracking competitor moves.

For a small brand, the choice is usually between hiring people to do this work or not doing it consistently. Hiring costs money directly. Not doing it means missing things that compound quietly over weeks: a campaign running on the wrong bid, a SKU drifting toward a stockout, a margin problem nobody caught.

What we built

A team of thirteen specialist agents that each own a specific area of the business. They run automatically every morning, monitor for issues throughout the day, and are available on demand to answer questions or run analysis.

The agents cover everything the business needs watched: ad performance and bid optimisation, inventory levels and reorder timing, profit margins per product, competitor pricing and keyword rankings, customer review trends, and bookkeeping. Routine operational tasks — sending review request emails, logging expenses — run automatically without any input required.

Every morning at 9am AEDT, the system pulls current data from all connected sources, runs each agent's analysis, and delivers a single briefing to Slack. The owner reads it over coffee. Anything that needs attention is flagged clearly, with the context already attached.

What we had to figure out

• Amazon's data lives in too many places and updates on different schedules Ad data, inventory figures, settlement reports, and review feeds all update at different rates. The agents needed to know when their data was fresh enough to act on and when to fetch again. We built a memory layer that tracks data age for each source and decides whether to use cached data or make a new call. This cut external calls significantly and made the system faster.
• The morning briefing sounded like a report, not an advisor Early versions listed metrics accurately but didn't interpret them. It told the owner what happened, not what it meant or what to watch. We spent significant time on the synthesis step — getting it to reason about the data, note what changed, and flag what warrants attention. The current version reads like something a good analyst would write.
• The first version escalated too many things Early on, the owner was getting Slack notifications for routine fluctuations that didn't need human attention. We worked through where the real thresholds were — what the owner actually needed to decide versus what the system should just handle. Fewer interruptions, higher signal on the ones that come through.
• Receipts and invoices don't arrive through a clean data feed Expense documents come as files in Slack — PDFs, images, spreadsheets, varied formats. The bookkeeping agent needed to handle all of it, extract the relevant figures, and log them correctly. Building reliable document reading that could handle the range of formats the business actually produces took more iteration than the more structured data work.