Running an Amazon brand solo means holding about thirty jobs at once. Before lunch you are a bid manager, an inventory planner, and a fee auditor. After lunch you are a copywriter, a photographer's art director, a customer-service rep, and a compliance officer. Most of those jobs are not hard. They are recurring - the same checks, the same reports, the same questions, every single week - and they compete for the one resource you cannot buy more of.
The existing ways to offload that work all have the same shape: trade money for hours. An agency takes the advertising slice for a monthly retainer plus a percentage of spend. A virtual assistant takes the repetitive slices for an hourly rate, plus the hours you spend training and checking them. Software takes the narrow slices it was built for, one subscription each. They can all be worth it. But stack enough of them to actually cover thirty jobs and the overhead starts to look like a payroll - for a business that might be one person in a spare room.
The observation that started this
AI can genuinely do a lot of those recurring jobs now - read the reports, spot the anomaly, draft the fix, explain its reasoning. The catch has always been the meter. Most AI-powered seller tools pay for intelligence by the token and pass that meter on to you, which means the more the AI works, the more you pay. "AI employees" priced that way are just a different agency.
Meanwhile, millions of people already pay a flat monthly fee for a frontier AI subscription - and use a fraction of what it can do. The compute is already bought. It sits idle overnight. It does not care whether it spends the morning summarizing emails or auditing FBA fees.
So we asked a simple question: what if the AI subscription you already pay for could staff your Amazon business? Not one chatbot you have to prompt - an organized workforce with departments, managers, and specialists, each with a defined job, running on a schedule, reporting back like employees do.
We run our own Amazon brands, so we did what we always do with ideas like this: built it and pointed it at our own money first. This post is what it is, how it works, and what it caught in its first week - published while it is still an internal tool, because building in public keeps us honest.
How an AI workforce actually runs
The architecture matters more than the org chart, so here is the honest version of it.
- The worker is your own AI subscription. A scheduled task in your Claude app checks in with our platform a few times a day. The platform hands it one bounded assignment at a time - one worker's job, one brand, one objective - and the AI does the work using live account data, then files a report. No API keys, no per-token meter, no software installed on your computer. Claude's scheduled tasks now run in the cloud, so the workforce clocks in even while your laptop is closed.
- The server is the manager. Which job runs, how often, in what order, and how much of your subscription a session may use is decided by the platform, deterministically - never improvised by the AI. There is a hard usage budget per check-in, sized to your plan, so the workforce can never eat the subscription you also use for everything else.
- Departments talk to each other through the platform. When the inventory seat sees a product running out of stock, it posts a warning. When the finance seat sees a product advertising past its break-even, it posts a warning. The advertising seats receive those notes before they touch anything, and treat the flagged products as reduce-only. Supply and profit gate the ads - the way a functioning company works.
- Nothing touches Amazon without you. Workers analyze freely, but any actual change - a bid, a negative keyword, a listing edit - is staged as a reviewable diff with the worker's reasoning attached, and executes only when the owner approves it in the dashboard. This is the same standard we wrote about in our AI Agent Policy piece: unsupervised AI does not get write access to the account that pays your bills.
Every seat also carries an honesty protocol: page through the full data before concluding, label every finding as confirmed or directional, say what was not checked, and treat "no change needed" as a valid day's work. A worker that manufactures findings to look busy is worse than no worker at all.
The org chart
As this publishes, the workforce running on our internal brands is 63 seats across seven departments, with more seats added weekly. Each seat is a specialist with a defined job, cadence, and reporting line. The seat-level playbooks are our secret sauce, but the departments deserve showing:
| Department | What it watches all day | From its first week on our brands |
|---|---|---|
| Advertising & Demand Capture | Bids, budgets, search terms, placements, creative performance, share of voice, conquest targets | Declined to cut four campaigns a simple ACoS rule would have killed - their computed break-evens said they were profitable. Roughly $1,800 of good spend protected by restraint. |
| Retail & Catalog Operations | Stranded and suppressed listings, silent Amazon-side page edits, buy box, organic rank, indexation, listing copy | Caught six listings that had silently lost buyable status while still taking paid clicks - about $1,100 of a fortnight's ad spend landing on pages nobody could purchase. |
| Creative Studio | Image stacks, video coverage, A+ content, brand voice, creative testing - as evidence-backed briefs written from real customer language | Turned review complaints about mismatched expectations into main-image briefs specific enough to hand a photographer. |
| Supply Chain & Inventory | Days of cover, stuck upstream replenishments, aged stock, demand forecasts, inbound reconciliation | Found hundreds of sellable units stranded upstream - counted as inventory, unable to sell - behind a dead offer nobody had noticed. |
| Finance & Analytics | Product-level profitability, fee audits, settlement reconciliation, reimbursement recovery, promo ROI, cash cadence | Flagged one product paying an oversized fulfillment fee tier worth roughly $900 a month, and caught a categorization error in our own profit pipeline that had overstated a month's profit by four figures. We fixed our platform the same day. That is what an auditor is for. |
| Customer Experience & Account Health | Account health trends, policy violations, review and return-reason mining, hijack signals | Split a month of returns into "the page set the wrong expectation" versus "the product has a problem" - two different fixes most reports blur into one refund number. |
| Workforce Operations | The workforce itself: data freshness, stuck queues, schedule tuning, a weekly change digest | Ranked a lagging dataset by which other seats were about to act on it - the AI equivalent of an ops manager catching a problem before the morning meeting. |
| Product Development and Brand & External Demand are on the chart with hiring closed - they open as their data connections ship. | ||
And this is what the work product looks like when a seat clocks out - a report from the fee auditor on one of our brands (identifiers redacted), exactly as it appears in the dashboard:
What it costs to run
Here is the number that makes the whole idea work. Measured on our own accounts, one worker assignment consumes roughly one percent of a Claude Max plan's five-hour session window. A full daily cycle for a mid-size brand - inventory checked, profit checked, ads reviewed, health monitored - ran six to eight percent of one window, out of the several windows a day the plan already includes. The platform's budget guard enforces this: you declare your plan tier and how heavily you use Claude for other work, and the server stops serving assignments before the workforce can crowd you out. Queued work is never lost; it waits for the next check-in. There is even a setting for weekends off - no runs, no notification emails, the queue resumes Monday.
What it does not do
Building in public means saying this part plainly.
- It does not execute anything on Amazon by itself, and that is permanent by design, not a temporary limitation.
- It cannot see data that has no connection yet: no DSP, no supplier or purchase-order data (that arrives with the next phase of our forecasting work), no buyer-message inbox. When a seat could not be backed by real data, we cut the seat rather than ship an empty suit - the org chart above is smaller than our ambitions on purpose.
- The creative department writes briefs and scripts from evidence; it does not generate images or video.
- It cannot read your Claude usage meter - no such API exists - so the usage budgets are deliberately conservative estimates, calibrated from our own measured runs.
Where this stands
The workforce is running on our own brands today, in private testing, finding real money and real problems weekly. It is built into AMZ Vault as the operational layer on top of the same data warehouse and approval system the rest of the platform uses. We are not opening it to customers yet: the approval flows, usage budgets, and worker playbooks get hardened on our own money first, and it ships when weeks of that have gone quietly. If you want to be early when it does, the rest of the platform - the P&L warehouse, the PPC tooling, and the AI connection all of this is built on - is live now, and existing users will see the workforce first.
Related: Amazon's AI Agent Policy and what it means for your tools · how AMZ Vault's AI connection works · try the platform with no signup