You typed one line into a chat box — "reorder the coffee, but only if it is under twelve dollars and ships by Thursday" — and walked away. Twenty minutes later a confirmation email lands. Something bought something for you, with your card, while you made lunch. That small moment is the entire fight over agentic commerce in miniature: the software is ready to spend your money, and most of us are not yet ready to look away while it does.
Why the money question is different
Letting an assistant find a product is low stakes. If it surfaces the wrong pair of boots, you scroll past. Letting it complete the purchase is a different category of trust, because a mistake now costs real money, ships to your door, and drags a return through your week. That gap between "help me look" and "go ahead and buy" is the line almost every shopper is quietly drawing right now.
And the hesitation is not vague nerves. In a 2026 checkout.com study, 27% of consumers said they trust no organization at all to run a buying agent, and 24% said they will never delegate a purchase to one. Read those two numbers together and a picture forms: a large slice of the market is not waiting for a better price or a smoother screen. They are waiting to feel safe about the moment money leaves their account.
Money is also flowing in the other direction, and fast. AI-referred retail traffic converts far better than it used to, and product recommendations from an agent close sales at rates a plain search page cannot match. According to a 2026 McKinsey outlook, agentic commerce could move between three and five trillion dollars globally by 2030, and Adobe Analytics clocked a sharp year-over-year surge in AI-referred shopping traffic in early 2026. The tools are not a curiosity. They are becoming a checkout lane.
~6 min
to build a multi-store cart
$3–5T
projected volume by 2030
300M
users on Amazon's Rufus
393%
YoY AI-referred traffic, Q1
The ~6 minute figure is the one worth sitting with. A person hunting the same deal across four stores burns half an hour and gives up cranky; an agent does the legwork before your coffee cools. That speed is exactly why delegation is tempting, and exactly why a wrong call can slip past you before you notice.
Not every agent shops the same way
"AI can shop for you" hides a wide spread in how these systems actually behave at the register. Some reason slowly and flag uncertainty; others move fast and rarely show their work. Using 2026 platform benchmarks compiled by commercetools, here is how the major assistants line up on the things that decide whether you hand over the card.
| Dimension | Claude | ChatGPT | Perplexity | Gemini |
|---|---|---|---|---|
| Checkout conversion rate | 16.8% | 15.9% | 10.5% | 3.0% |
| Multi-store price hunt | Strong | Strong | Moderate | Weak |
| Shows its sources | Yes | Partial | Yes | Rarely |
| Tone on risky buys | Cautious | Eager | Source-led | Minimal |
| Oversight recommended | High | High | Medium | High |
| Best Suited For | Cautious big-ticket buys | Everyday high-volume orders | Bargain hunting | Quick Google-linked picks |
The conversion spread tells you something the marketing never will: an agent that closes fewer sales is often the one being careful on your behalf, not the one failing. Match the tool to the job. A cautious reasoner for the expensive, irreversible buy; a fast one for restocking the pantry.
It also helps to picture how far you are actually letting go, because delegation is a ladder, not a switch. Most people are comfortable a rung or two up and get uneasy near the top.
The delegation ladder above is the safe way to adopt these tools: start where the agent only proposes, move up only as it earns your confidence on small, cheap, reversible orders
Where this quietly goes wrong
Speed and reliability are not the same thing, and shopping agents are still shaky exactly when the task gets interesting. A model that lands a simple job on the first try can stumble badly once the request stacks up steps, comparisons, and edge cases. That is fine when you are watching. It is a problem when you have handed over the card and closed the tab.
- Reliability drops off a cliff on hard tasks. In 2025–2026 agent benchmarks (WebMall and DeepShop), systems that succeed roughly 60% of the time on one attempt fall to about 25% across eight consecutive runs, and top agents finished under 65% of genuinely hard jobs like locating the cheapest option across several shops.
- Fraud follows the money. Roughly 78% of financial institutions, in 2026 industry polling, expect AI-driven shopping to push fraud higher — automated buyers are a fresh, fast-moving target for scams and spoofed storefronts.
- Confident wrong answers cost real cash here. When a chatbot invents a fact you catch it; when a buying agent picks the wrong variant, size, or seller, the mistake arrives in a box with your name on it.
There is a genuine grey area worth admitting: nobody has a clean answer on who eats the cost when an autonomous agent buys the wrong thing. Is it your mistake for delegating, the retailer's for a confusing listing, or the model maker's for a bad decision? Refund policies were written for humans clicking buttons, not software acting on a loose instruction, and that unsettled question is a real reason to keep purchases on a short leash for now.
Let the agent do the hunting, the comparing, and the boring tab-juggling — that is where it genuinely saves you time and often finds a better price. Keep the final tap on the buy button yours until the trust is earned in small, cheap orders you can afford to get wrong. The technology is ready to spend. You get to decide, purchase by purchase, whether it has actually earned the wallet.
