Managerbot · supplier ordering
Managerbot supplier ordering
Managerbot is Square's AI agent for sellers. i designed and built the parts of supplier ordering where the seller signs in, checks what the agent did, and approves the order.
shipped · staging
restaurant owners spend hours on admin, including ordering from suppliers like US Foods and Sysco. Managerbot can turn a handwritten list into a supplier cart, but sellers have to trust it with their logins and their money.
i designed and built the steps where the agent hands control back to the seller: answering its questions, signing in, watching it work, and approving the order. i started with HTML prototypes, then built it in React and Go.
sign-in, the secure browser, and the US Foods commands are merged and shipping soon. the whole flow ran end to end on staging, and i ranked every issue it turned up for launch.
anything that spends money or uses a credential goes to the seller first.the rule i designed around
sign-in for a password supplier
the supplier and website rows are locked, so the seller can see where the login goes. the password never goes into the chat or the model.
sign-in for US Foods
i signed in to the real US Foods portal and changed the prototype to match it: the login is a user id, and a new one-time code arrives every time.
notes for engineering
each screen carries notes tied to the requirement it answers.
- 01a login belongs to the business and the person, is shared with the team by default, and has a toggle per login
- 02for one-time codes the seller picks a masked phone number like (***) ***-4416; Square stores the choice, not the number
- 03one question per card, so team sharing got its own card after sign-in
the goal was an alpha a US Foods seller could start using quickly, so hero sellers could tell us what to fix.
from a photo to a cart
each step is a point where the seller checks or approves something.
the proposed cart
the agent builds the order and stops there. nothing is submitted until the seller approves it.
running the whole flow turned up problems. i fixed what i could and wrote up the rest, ranked for launch.
AI did a lot of the legwork, and i checked its work at each step.
ordering through the agent shouldn't take longer than doing it yourself.
most of the design had precedent in patterns the team already used, so the interesting work was in the details and in how the agent behaves.
if the agent asks five follow-up questions, the seller could have picked the items in that time. i worked on cutting steps, sometimes down to one, and on what goes into the agent's instruction file so it only asks when the answer actually matters.
the hosted browser showed me how much small details affect trust. a cursor that doesn't change on hover makes a page feel fake, so i spent real time making the sign-in look like a real one.