Can an AI Store Manager Beat a Reordering Rule?

Four of five AI store managers finished behind a reordering rule in AutoVend’s 180-day retail simulation. In one DeepSeek run, a campus machine earned 5.14 yuan in gross profit between visits. The next day the agent sent someone back to add four units. The trip cost 35 yuan.

Online, the same decision might come down to a freight charge. An agent spots a likely stockout and brings a shipment forward. It saves the order. Will the sale pay for the rush, the stock left over and the time you spend sorting it out?

AutoVend charges for the trip from warehouse to shelf

Qiming Shi and colleagues gave each agent 40,000 yuan, one warehouse and 20 machines spread across five kinds of locations, including campuses, offices and residential sites. The agent chose what to sell, set prices and paid suppliers when it placed an order. Three days later, the goods reached the warehouse. A customer staring at an empty coil on campus still had nothing to buy.

The operator paid 35 yuan per machine on each day staff visited. Restocking six lines on one visit cost the same as restocking one. Come back tomorrow for a forgotten line and the meter starts again. Leave too many drinks in storage and the warehouse collects rent while the expiry date creeps up.

The researchers used real vending sales to model demand. Their experiment measures simulated physical retail. For an ecommerce trial, the part worth borrowing is the clock. Keep it running from the order decision through delivery and sale or expiry. Your customers and your fulfillment costs supply the evidence about your shop.

Which agent beat the inventory rule?

The rule restocked machines weekly. It set a target from the past 30 days of demand, allowed for swings in sales and checked what was already on order before buying more. It also adjusted the product mix and tried local price changes. On day 180 it held 210,970 yuan in cash and unexpired inventory at cost, counting goods on shelves, in storage and in transit.

GPT-6.1 Sol beat the rule in all three runs, averaging 309,350 yuan in ending assets. None of the other four models cleared the rule in any run. All 15 agent runs grew their starting capital. The researchers ran each model three times in the same simulated world. Switch the question from “Did it make money?” to “Did it beat the rule?” and only one model gets a yes. See Table 2 of the AutoVend preprint.

DeepSeek V4.1 Flash led on a different question. For each dollar of estimated model fees, how many yuan did the agent add to its starting assets? GPT spent about 80 times as much on inference as DeepSeek for about 2.9 times the asset growth. Dividing growth by the model bill rewards efficiency. If you want to know how much money you keep, put both amounts in one currency and subtract the bill.

What one extra sale earns after replenishment

Suppose your usual reorder rule would leave a specialty food box out of stock until Monday. Your agent brings in the supplier’s two-box minimum early and saves one $48 order. Each box costs $12. The buyer gets the first one. The second expires before anyone buys it.

Assume the early order adds $4 in inbound freight, packing and shipping the sale costs $7, and the model calls cost $1. A stock discrepancy takes a staff member 12 minutes to clear at $30 an hour. Compare only the extra revenue and costs you incur by ordering early.

The extra order contributes $6 toward overhead. Another 12 minutes on the discrepancy would use up the entire $6. Now suppose the second box still has a month left on the trial’s last day. Keep its $12 cost in inventory until it sells or expires. The supplier has your cash, but you still own the box.

The person clearing the discrepancy rarely gets a line on the sales report. In an October 21 account, an online merchant used agents to invoice customers, check payments and buy labels. Ask for a one-number invoice correction and the agent sometimes changed other fields. The owner kept the final say on amounts. In a later reply, the same owner said automation saved time in the store and freed up room to start another business.

Another merchant in the thread credited a warehouse kiosk with getting them out of daily packing. A commenter suggested timing corrections and reruns alongside ordinary orders. The shop can give its owner time back even if the owner never takes an afternoon off. Counting the invoice repairs gives you a fair measure of how much time came back.

A scorecard for an inventory-cycle trial

Run the agent and your current reorder rule against the same products, starting stock and demand over the same dates. A replay works if you have demand records for both paths. Sales alone hide the people who wanted an item while it was unavailable. Mark gaps in demand before treating a replay as a fair contest.

Pick the end date from the product. For a Friday reorder with a ten-day supplier lead time, a trial ending the following Friday misses the delivery. Follow the shipment through the next reorder and long enough to see what sells or expires. Value any remaining usable stock the same way on both sides.

At the same time each day, count one stockout for every product with zero sellable units. Count each late or unfilled order once under service failures. Include routine checks and repairs in human minutes on both sides, then price those minutes once in contribution. Fewer expired units look good until you see how many days the shelf sat empty.

A $48 order leaves $6 in the example, and one more stock check can spend the rest. Your own cycle report tells you whether the agent earns its keep. A .boss name at HeadlessDomains.com gives the service a public address where you can identify its operator and keep a contact route current for anyone chasing a late shipment.

Find a .boss name for your store-operations agent.