Eight stages
- DiscoverResults the shopper can correct
- DecideA pick checked against the ask
- AuthorizeMoney moves only with a clear ask
- Check outA clear path to pay
- TrackThe order, not only the receipt
- Return & resolveA claim the shopper can approve
- Re-order & watchA watch that is not a buy
- Review & revokeOne place to turn spend off
Payment is covered. Fit and aftercare are not.
You can already approve a charge, set a cap, name the seller, and look up spend. What still breaks is the pick, and everything after the money moves. The agent buys the wrong thing. Then there is no pattern for tracking, a return, or turning the agent off.
Design principle
Walk the purchase in order. One goal and one decision per stage.
Eight stages
Each stage has a design question, what to build, and the patterns that cover it.
Discover
Results the shopper can correct
- Question
- Can I see why these products, and change the constraints?
- Principle
- Show why a result matched. Let the shopper edit price, size, or brand without starting over.
- Breaks as
- A product grid with no reason attached, or filters that wipe the whole ask.
- Use when
- The agent is turning an ask into a set of products.
- Build
- Search by meaning, with a reason on each match
- Editable constraint chips
- A way back to text when a photo search is weak
Decide
A pick checked against the ask
- Question
- Does this item match what I asked for?
- Principle
- Comparison and reviews help. The buy still has to match what the shopper asked for.
- Breaks as
- A buy button on a result that fails the constraints.
- Use when
- The shopper is choosing, or about to let the agent recommend one item.
- Build
- Side-by-side differences that matter
- Review themes with sources
- A fit check against the ask before anyone can buy
Authorize
Money moves only with a clear ask
- Question
- Approve this charge, or let it buy inside rules I already set?
- Principle
- One charge needs an approval. A repeat buy needs a limit set in advance. Keep the seller, the amount, and the reason on the request.
- Breaks as
- A charge that shows the amount and forgets the ask. Or a phone prompt that times out in silence while the charge still goes through.
- Use when
- The agent is ready to spend.
- Build
- One-charge approval with merchant, items, and reason
- A spending limit with cap, merchants, and expiry
- A one-time card scoped to one merchant and amount
- A record of the ask bound to the approval
Check out
A clear path to pay
- Question
- Who is the seller, and do I finish here or on their site?
- Principle
- Some buys finish inside the agent. Others hand the cart to the merchant. Say which path this is before anyone pays.
- Breaks as
- An in-chat pay button that hides a redirect. Or a redirect that drops the cart.
- Use when
- The shopper is finishing the purchase.
- Build
- Instant buy with totals and consent still visible
- One cart when items span merchants
- A handoff when the merchant stays the seller
- The seller named on the way out
Track
The order, not only the receipt
- Question
- Where is the order, and what can I still change?
- Principle
- After the charge, the shopper needs shipment status and a path back to the merchant order. A spend history is not tracking.
- Breaks as
- A spend row with no link to the order it created.
- Use when
- The shopper comes back after the agent has already paid.
- Build
- A log of what the agent bought
- Order status from placed to delivered
- One card per purchase with merchant, items, and status
Other related patterns
Return & resolve
A claim the shopper can approve
- Question
- Something is wrong. Who files the claim, and who pays me back?
- Principle
- A bad delivery needs a claim, a decision, and a refund path. A merchant return-risk score is not that flow.
- Breaks as
- The agent files a return the shopper never saw. Or sends them to start over on the merchant site with no record.
- Use when
- The item is wrong, damaged, or not what was asked for.
- Build
- Support that already knows the order
- A handoff to a person with the order context intact
- An agent-filed claim the shopper approves first
Re-order & watch
A watch that is not a buy
- Question
- Watch this, then buy only when my rule hits?
- Principle
- A price alert is not permission to buy. The rule that turns a watch into an order is its own decision.
- Breaks as
- A price drop notification that places the order on its own.
- Use when
- The shopper wants the item later, or the same buy again inside a cap.
- Build
- A price alert that stops at the notification
- A watch-then-buy rule with target price, merchant, and cap
- A recurring mandate the shopper can see and end
Other related patterns
Review & revoke
One place to turn spend off
- Question
- Which agents can still spend, and how do I stop one?
- Principle
- List every agent that can still move money. Give one control that stops the next charge.
- Breaks as
- A wallet connection you can only remove by deleting the account.
- Use when
- More than one agent can pay, or the shopper is done with one of them.
- Build
- A list of connected agents and what each can still do
- Revoke that kills the next charge
- Spend history so the shopper knows which agent to stop
Design workflow
Walk one purchase with this framework before you design the screens.
- 1
Start at discovery
Mark the stage you are designing. Checkout is late in the journey.
- 2
Skip merchant tools
Dynamic pricing, inventory forecasts, and return scores belong to the store. The person who paid never sees those screens.
- 3
Pair the frameworks
Agentic UX for how much authority the agent has. Trust Scaffolding for how the decision is shown.
Cheatsheet
Checks before you ship an agent that can spend.
- Can the shopper see why these products matched?
- Does a bad fit block the buy?
- Does the approval show seller, amount, items, and reason?
- Is the seller named before pay, on both in-agent and handoff paths?
- After the charge, can the shopper find the order?
- Is there a claim path the shopper can approve?
- Does a price alert stop before it buys?
- Can the shopper list agents that spend, and stop one?
Other frameworks
- Agentic UXUse when the product can take actions (send, spend, edit, delete) and you need to decide how much authority is safe.
- Chat UXUse when the product is a conversation: composing turns, streaming answers, remembering context, and recovering when a turn fails.
- Trust ScaffoldingUse when people must decide whether to believe or rely on the system, from first run through evidence, control, and long-term use.




