Seven-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3, 4 and 7 need a person. Step 3 is waiting for approval right now. Click it to open the full walkthrough.
Estimates are directional and based on stated assumptions. All names, organizations, and identifying details have been anonymized in accordance with our confidentiality agreements.
How a Growth-Stage E-commerce Company Built a Storefront Where Each Week's Sales Change How It Runs
Every product in the store, from the morning its supplier data lands to the Friday its sales record changes how the store runs. The agents keep the catalog complete, the listings honest, the prices inside the band the owner set, the shelves stocked, and the stalled carts chased. A person is asked three things: a price that wants to leave the band, a purchase order before money leaves, and the weekly changes to how the store itself works. Seven stations, run as a loop: what the week close learns about what sold feeds the listings station and the pricing band the following Monday.
What Was Broken
What We Built
Seven stations and 14 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Catalog Intake
Nobody is asked anything hereAgents collect every new and changed product off the supplier feeds and the warehouse system, match each row to the catalog record it belongs to, and hold anything incomplete or contradictory. Nothing waits on a person here. A held product is never published half-described and never overwrites a record.
Listing Writing
Nobody is asked anything hereAgents write the title, description, bullets and search terms for every product off its own attribute sheet, and cut any claim the sheet does not carry. Nothing waits on a person here. A claim with no attribute behind it never reaches the page.
Price Setting
A person answers hereAgents collect both rivals' prices every morning, move each product's price inside the band the owner set, and hold every price that would leave the band. This is the first place a person works. The band is theirs, and a price that wants to leave it waits.
Stock Reordering
A person answers hereAgents project six weeks of demand for every product, draft the reorders with cost and ship date, and put last time's stockout on screen beside each one. Money leaves only over a signature. This is where a person signs. An order the forecast cannot support is withheld, not padded.
Shopper Matching
Nobody is asked anything hereAgents rank every page for the shopper looking at it, pair products bought together, match promotions to the shoppers whose history says they will use them, and pull a sold-out product off every shelf. Nothing waits on a person here. A sold-out product is off the shelves within a minute, and a promotion never reaches a shopper it would teach to wait.
Cart Recovery
Nobody is asked anything hereAgents trace every visit from landing to payment, group the drop-offs by where they stopped, and chase each stalled cart with the offers already approved. A message stops the moment its shopper buys. Nothing waits on a person here. An offer bigger than the approved one is a price decision, and it is flagged, never sent.
Week Close
A person answers hereAgents close the week's ledger, trace what the store settled on its own against what reached a person, and propose the listing rewrites, band moves and rules the record argues for. A person accepts or rejects each one. This is where a person decides. How the store prices, describes and promotes is never changed quietly.
Where a Person Decides
Operating Model
This changes how work flows through the team.
What Transfers, What Must Be True
Failure Modes
What breaks this pattern:
The description writer puts a claim on the page that no attribute supports. Shoppers buy on that claim, and the returns and the complaints come back to you.
A rival price seen once can be a scrape error or a one-hour sale. An agent that matches it cuts your margin against a price that no longer exists.
When a product's sales break from their own history, the forecast behind the reorder is wrong. The agent spends real money on stock for demand that already changed.
A promotion sent to a shopper who paid full price this week refunds a sale you already closed. Every shopper who sees it learns to wait for the next one.
Directional Outcomes
What the agent counts, and the station that counts it.
These counts are the tallies from one monitored run of the agents. They are not monthly or annual totals.
What Runs Where
Every step names the subagent that does the work, the record it writes, the thing that raises a question for a person, and what it is allowed to touch. This is drawn from the source, not from a diagram somebody kept in sync by hand.
catalog-intakecatalog/intake/<sku>.jsonhold-contradictory-recordcatalog/intake/**, catalog/held/**, the supplier feeds read-onlylisting-writercatalog/listings/<sku>.mdattributes-too-thincatalog/listings/**, catalog/intake/** read-onlyprice-deskpricing/<sku>.jsonprice-outside-bandpricing/**, catalog/** read-only, the shop platform's price listreorder-deskinventory/orders/<order-id>.jsonawait-your-signatureinventory/**, the sale record read-onlyshopper-matchshop/placements/<visit-id>.jsonpromotion-breaks-the-floorshop/placements/**, catalog, pricing and stock read-onlycart-recoveryjourneys/recovery/<cart-id>.jsonoffer-exceeds-approvedjourneys/**, shop/** read-only, the mail senderweek-closerules/proposed/<rule-id>.json, reports/week-<date>.jsonaccept-changerules/proposed/**, reports/**, the history read-onlyStack
Every system this agent reads or writes.