27

Blog · November 5, 2026 · 8 min read

Agentic Checkout UX 2027: Design the Spending Permission Before the Buy Button

For twenty years, checkout UX meant one screen and one button. A person looked at a cart, checked a total, and pressed Pay. In 2027 a growing share of purchases will be assembled by software agents acting for a person or a company, and the moment of review moves. Sometimes it happens one second before the charge. Sometimes it happened three weeks earlier, when someone typed "keep our design tools renewed" into an assistant and clicked Allow.

The payments industry is already building the plumbing for this. Google's Agent Payments Protocol, AP2, separates two cases: a human present flow, where the person sees the final checkout and approves it, and a human not present flow, where the person approves a set of constraints and the agent completes purchases inside them. Card networks and commerce platforms are shipping their own agent payment programs on top of the same idea.

The protocols handle cryptography and liability. They leave the interface to you. Firmly, one of the agentic commerce providers, says so directly in its docs: its checkout APIs do not enforce a review step, a disclosure of the total, or consent capture, and the surface that talks to the user owns all of that. So in 2027 the product teams building assistants, wallets, procurement tools, and marketplaces will be designing the trust layer themselves. Most of them will design it as a single Allow button. We think that is the mistake that defines this category.

The permission screen is the new checkout

When an agent buys on your behalf, the permission you granted is the real checkout. The purchase that follows is just execution. That means the permission screen has to carry everything the old checkout carried, plus the things a person never had to think about when they were pressing the button themselves.

A good permission answers these questions in plain language before anyone agrees:

  1. What can it buy? A category, a merchant allowlist, or one specific item.
  2. How much, and how often? A per purchase cap and a total budget are different limits, and users confuse them constantly. Show both, with currency, and say whether tax and shipping count.
  3. Until when? Every permission needs an end date. "Forever" should require a deliberate extra step.
  4. What happens when something changes? Price goes up, item is out of stock, a substitute exists. The user picks the rule up front: stop and ask, substitute within limits, or skip.
  5. Which money? The exact card or account, shown by its last four digits.
  6. Who gets told? The person, an admin, a finance channel, or all three.

"Allow this assistant to make purchases" answers none of these. It feels lightweight in a demo and it creates the dispute queue of 2027.

Our opinion: write the permission as a sentence the user could read back to a friend. "Renew Figma and Notion each month, up to $400 total, charged to the card ending 4421, until March 31. Ask me if either price changes." If your team cannot write that sentence for a given flow, the agent is not ready to spend money in that flow.

Human present flows: one confirm, frozen

When the person is around, the right pattern is a single confirmation on a frozen proposal. The agent can browse, compare, and negotiate as long as it likes. Then it stops, validates inventory and totals, and shows one summary:

  • The item, variant, and quantity.
  • The real merchant, named the way it will appear on the statement.
  • An all in total, with anything charged later shown separately.
  • The refund posture in one label: refundable, credit only, or final sale.
  • The constraints the agent honored, such as "under budget" or "ships before Friday."
  • One execution sentence: "When you confirm, we will charge $312.40 to the card ending 4421."

Then freeze it. If anything on that summary changes between the preview and the charge, the agent has to come back and ask again. A total that drifts after approval is the fastest way to lose a user's trust in an agent permanently, and it is also how a confirmation record stops being useful in a chargeback.

Keep two exits visible next to the confirm button: edit, which returns to the proposal with context intact, and cancel, which ends the task with no side effects. Hiding edit pushes people back to doing the purchase by hand, and then the agent has no reason to exist.

Human not present flows: the receipt does the work

When the agent buys while nobody is watching, the receipt replaces the review screen. That makes the receipt one of the most important screens in the product, and most teams treat it as a notification.

A receipt for an agent purchase should show what was bought and from whom, what it cost and how much budget remains, which permission allowed it, and why the agent decided to buy now. For B2B and fintech products, the "why" should point to evidence: the quote, the policy rule, the budget line, the expiring subscription. "Task completed" with a dollar figure is a support ticket waiting to happen.

Every agent purchase should also land in an activity feed that sits next to the person's normal order or transaction history. Mixing agent actions into the same list without a marker hides them. Splitting them into a separate tab buried in settings hides them too. We put them in the main history with a clear agent label and a filter.

Revoke has to be one tap from the receipt

The control people need most in 2027 is the one most products hide: stop. Pause and revoke belong on the receipt, on the activity feed, and on the permission itself. Three levels deep in account settings is too far.

Two details matter more than teams expect. First, say what revoking does and does not do. Stopping future purchases does not cancel an order already placed, and users assume it does. Second, give a dispute path from the same place: report this purchase, mark the agent's reasoning as wrong, or send it to an admin. That feedback is also the best training signal the product will ever get.

Where this hits regulated products hardest

Fintech products carry this first because they hold the money and the disputes. Wallets, neobanks, and card issuers will need permission objects that map cleanly to the statement line, the dispute form, and the fraud model. An agent purchase the fraud model cannot tell apart from a stolen card is a bad outcome for everyone.

B2B SaaS and procurement tools carry it next. Here the person approving the permission is often an admin, and the person affected is someone else. Separate the two roles in the interface, and show the affected user what the agent is allowed to do on their behalf.

Marketplaces carry the merchant side. Sellers will receive orders from agents, and they need to know it, because return behavior, substitutions, and support expectations change when the buyer never saw the listing.

Health and insurance products will meet this through payments for care, prescriptions, and premiums. Anything that touches eligibility or coverage should default to a human present flow until the rules settle.

What we would reject

  • A single "allow purchases" toggle with no amount, merchant, or end date.
  • Permissions that never expire by default.
  • Totals that can change after the person approved them.
  • Merchant names that differ from what appears on the statement.
  • Receipts that say "done" without the reason or the permission that allowed it.
  • Agent purchases hidden in a separate history nobody opens.
  • Revoke controls that live only in account settings.
  • Silent substitutions when an item is unavailable.
  • Admin approved permissions that the affected user cannot see.
  • Usability testing that only measures how fast someone grants permission.

That last one deserves a note. Speed to Allow is the metric that makes these flows look great in a sprint review and fail in production. The tests that matter ask a person to explain, in their own words, what they authorized, to spot an exception, and to stop the agent. If they cannot do all three, the flow is not ready.

How we approach an agentic checkout engagement

Our team is senior only, and we design and build in working software from the first week. For agentic checkout, that means we map every place your agent can spend money, write the permission sentence for each one, and build the permission screen, the frozen confirm, the receipt, and the revoke path as real components your engineers can merge. AI handles the scanning, the edge case enumeration, and the first pass at states. People on our team make the calls about risk, consent language, and what a user sees before money moves.

If you want a fast read on where your current checkout loses people today, scan your own page with our free UX Scan. It takes a few minutes and flags the friction before an agent ever touches it. When you are ready to design delegated spending properly, book a discovery call or email us at ash@uxsignalstudio.com.

Let’s talk about your next step

Have a question or a project in mind?

Tell us where you are and what you want to figure out. We’ll review your message and reply within two business days.

Contact UX Signal Studio

We’ll use your details to respond to your inquiry. This form does not subscribe you to a mailing list. Privacy policy.

Prefer to talk? Book a discovery call

Get the next essay in your inbox

One practical piece on UX and product design, every week or two. No noise.