What It Takes to Let a Machine Spend Your Money
Shoppers expect AI-driven purchases within a year, but few trust an agent to pay. The mandates, guarantees, and machine-readable policies that turn hesitation into delegation.
On this page
Nobody loses sleep over letting an algorithm suggest a sweater. We have been trained on two decades of recommendation engines, and the worst outcome of a bad suggestion is a shrug. Handing that same algorithm the card is a different psychological event entirely. It converts a suggestion into a commitment, and commitments carry consequences: money leaves, packages arrive, mistakes must be unwound. That is the exact fault line running through agentic commerce right now, and the research is unusually consistent about where it sits.
Checkout.com's Agentic Commerce 2026 report found that a third of consumers (33%) expect at least 10% of their purchases to be AI-driven within a year. Demand is forming faster than most merchants assumed. The same study found that 72% of merchants believe consumers will adopt agent-led shopping faster than most sellers are prepared for. So the question is no longer whether people will delegate shopping to agents. The question is what, specifically, makes a hesitant person comfortable enough to let an agent complete the payment. The answer turns out to be less about model quality and more about artifacts: concrete, inspectable objects a shopper can point to and say, this is what protects me.
The gap between choosing and paying
Consumer psychology has a name for what happens at checkout: the moment of loss. Behavioral economists have documented for decades that paying activates something closer to pain than deliberation, which is why one-click checkout and stored cards were worth billions. Delegation to an agent inverts the problem. Instead of reducing the friction of a payment you chose, it asks you to pre-approve payments you have not seen yet. That is a genuinely new ask, and shoppers respond to it the way people respond to any open-ended liability: with suspicion.
The numbers reflect this. In the same Checkout.com research, 27% of consumers said they trust no organization at all to operate an AI shopping agent on their behalf, and 24% said they will never delegate purchases to AI. Read those figures carefully, though. They also mean roughly three quarters of consumers are persuadable. The hesitant middle is the entire market, and the hesitant middle does not want reassurance in the form of marketing copy. It wants mechanisms. Anyone who has studied ecommerce conversion statistics knows that trust objections compound at the payment step; agentic checkout simply moves that compounding earlier, to the moment of delegation.
The design decision: treat discovery permissions and payment permissions as separate products. Let shoppers grant an agent the right to search, compare, and shortlist long before you ask for the right to spend. Every merchant flow that bundles the two into one consent screen is leaving the persuadable middle on the table.
A mandate is a pre-nup for your agent
The most concrete trust artifact to emerge so far is the spending mandate: a signed, machine-verifiable statement of what the agent may buy, from whom, for how much, and until when. The Agent Payments Protocol (AP2), launched by Google Cloud and Coinbase in September 2025, structures every agent transaction as a chain of cryptographically signed mandates covering intent, cart, and payment, so any party can later prove the agent acted within scope. The output is a verifiable record that a payment was authorized, one that any settlement rail can check against.
Card networks are building the same idea into their own stacks. Visa's Intelligent Commerce program issues scoped tokens to agents with spending limits, approval workflows, and authentication requirements attached, while the shopper keeps the underlying card and its dispute rights. The psychology here matters more than the cryptography. A mandate works because it resembles instruments people already understand: a per-diem, an allowance, a standing order with a cap. It converts "the AI has my card" into "the AI has fifty dollars for razor blades this month," and those are entirely different sentences to a nervous delegator.
The design decision: surface the mandate in the interface, in human language, before and after every agent purchase. "Your agent can spend up to $75/month on household refills, at these three retailers, until March 1. Change this anytime." The control does not build trust if the shopper cannot see it.
Guarantee language a nervous delegator can repeat
Trust in agentic commerce is asymmetric in a way merchants underestimate: it accrues slowly and evaporates instantly. Checkout.com's report found that 25% of consumers would stop using an agent entirely if a purchase went wrong. One bad basket, one mischarged card, one wrong size, and a quarter of your delegated buyers revert to manual shopping and take their skepticism with them. This is the same dynamic retention teams see everywhere; the patterns in churn prediction work apply almost unchanged, except the churn event here is a collapse of delegation rather than a lapsed subscription.
Which is why guarantee language is not legal boilerplate in this context. It is the product. The guarantees that move hesitant shoppers share three properties: they are specific ("any agent-initiated purchase can be returned free for 30 days, no questions"), they are unconditional at the moment of doubt (no "may be eligible" hedging), and they are short enough for a customer to repeat to a spouse who asks why the AI just bought something. If your guarantee cannot survive being paraphrased at a kitchen table, it is not doing its psychological job.
The design decision: write a distinct agent-purchase guarantee, separate from your general returns policy, and put it in the agent flow itself. Make the first wrong purchase free to undo, loudly. You are not insuring transactions; you are insuring the relationship with delegation itself.
Return policies your agent can actually read
Here is the quietly decisive part: the shopper is no longer the only audience for your policies. When an agent decides between two merchants with similar prices, it weighs whatever it can parse, and a return policy written as 800 words of prose parses as uncertainty. Google's structured data documentation defines the MerchantReturnPolicy type precisely so that return windows, fees, and methods become queryable fields: 30 days, free returns, online process, no ambiguity. An agent comparing you against a competitor with clean structured policies will treat your prose as a risk factor, because from its vantage point, it is one.
This is also where the trust artifact does double duty. The same markup that makes an agent confident makes its explanation to the shopper confident: "I chose this merchant because returns are free for 30 days" is a sentence an agent can only generate if the data exists in machine-readable form. Merchants investing in AI search optimization tend to discover this early; the stores agents recommend are disproportionately the ones agents can read. If your policies live in PDFs and accordion menus, a schema generator is a faster fix than a redesign.
How far behind are sites on this? In September 2026 we audited the homepages of 54 well-known B2B SaaS vendors with our AI Visibility Checker: half carried two or more high-value schema types, and 15% had no structured data at all. Those are software companies' homepages rather than merchant product pages, so the audit says nothing about return-policy markup directly. It does show that even well-resourced teams leave structured data half-finished, which is the opening a merchant with clean MerchantReturnPolicy fields walks through.
The design decision: publish MerchantReturnPolicy and shipping markup on every product, and audit it quarterly the way you audit page speed. In agentic commerce, your returns policy is part of your product feed.
The merchants moving anyway
For all the consumer hesitancy, sellers are not waiting. Checkout.com's report puts merchants already testing agentic commerce at 42%, which means the trust artifacts described above are becoming competitive surface area right now, while most shoppers are still in the persuadable middle. And the direction of consumer expectation is unambiguous: Accenture's survey of more than 25,000 consumers found 71% expect generative AI to influence at least half of their spending decisions over the next twelve months. Influence precedes delegation the way browsing precedes buying.
There is a warm reading of this moment and a skeptical one, and both are correct. The warm reading: delegation is the most honest loyalty metric ever invented, because a shopper who hands your category to an agent with a standing mandate has committed in a way no points program ever measured. The skeptical reading: that loyalty is provisional, revocable in one tap, and conditional on you never making the agent look foolish. The merchants who win the next two years will be the ones who treat mandates, guarantees, and machine-readable policies as conversion assets with owners and budgets, the same way they treat personalization or their first-party data programs. Trust, it turns out, ships in artifacts. Build the artifacts and the delegation follows.
When mandates and guarantees fail to move the hesitant middle
Each artifact above has a failure mode, and most of them come from applying the artifact too literally.
- Mandates that ask for approval on every purchase. Per-purchase confirmation feels like control in the design review and like nagging by the second week; the shopper stops delegating, which was the whole point. The trade-off runs both ways: caps that are too tight generate constant re-authorization, caps that are too loose read as open card access. Start with a category plus a monthly ceiling, and adjust the ceiling to basket frequency. It depends on the category far more than on the protocol.
- Guarantees generous enough to be farmed. "Any agent purchase can be returned free, no questions" is the sentence a nervous delegator repeats, and also the rule an optimizing agent will exploit if it has no cap. Pair the promise with a per-account limit that is published in the same place, for the reasons laid out in what agents do with hidden versus published rules.
- Markup that contradicts the prose policy. A MerchantReturnPolicy block promising free returns beside a PDF that mentions a restocking fee is worse than no markup: the agent sees inconsistency and Google sees a mismatch. The first-timer's mistake is generating the schema from a template without reconciling it to the policy the legal team actually maintains.
- Splitting discovery and payment consent for trivial purchases. For a low-ticket replenishment, the second consent screen backfires, because the value of delegation is not having to think about razor blades. Counterintuitively, the more trivial the purchase, the more a bundled consent works; save the split for considered baskets.
- Legal hedging in the guarantee. "May be eligible for a refund" fails the kitchen-table test on contact. That said, the fix is editorial rather than structural: one unconditional sentence, then the exceptions below it.