Agentic Commerce in 2026: AI Agents Can Shop, But They Still Can't Pay

Here's a number that would have sounded absurd two years ago: ChatGPT now handles roughly 50 million shopping queries a day. Not code, not homework—shopping. The way people find products has already flipped. The way they pay for them has not, and that gap is where the next payments war is being fought.

This piece closes a loop with the stablecoin article: there I argued programmable money was infrastructure looking for a killer workload. Autonomous agents are that workload—if anyone lets them touch the money.

The Flip Already Happened

The demand side is not speculative. It's measured.

FIG. 01 // THE FLIP

Consumers who prefer AI tools over traditional search

58%

PREFER AI TO SEARCH ENGINES, 2025

In two years the default starting point for buying something moved from the search box to the chat box. 73% of consumers now cite AI as their primary product-research channel; GenAI referral traffic to retail sites grew 4,700% year-over-year.

DATA: COMMERCETOOLS · METAROUTER · ADOBE · 2025-2026

58% of consumers now prefer AI tools over traditional search engines, up from 25% in 2023. 73% cite AI as their primary product-research channel, and GenAI referral traffic to retail sites grew 4,700% year-over-year. Google spent twenty years as the front door of commerce; the front door moved in about twenty-four months.

For anyone who owns retail or ad-tech exposure, this is not a future risk. It's a present repricing: product discovery is migrating to surfaces that have no auction, no ad unit, and no analytics standard yet.

The Cliff at Checkout

Now the part the press releases skip.

FIG. 02 // THE CLIFF

Where AI agents operate in the purchase journey

Agent usage collapses between comparison and checkout. The missing 39 points are not a demand problem — they are missing payment, identity and authorization infrastructure for autonomous transactions.

DATA: COMMERCETOOLS ENTERPRISE GUIDE · 2026

Agent usage is ~62% at product comparison, 23% at checkout, 19% post-purchase. Agents research brilliantly and then hand the human a link, because the infrastructure for a bot to transact—payment credentials, identity, authorization, liability—is still too immature for autonomous checkout.

That 39-point cliff is the most valuable unsolved problem in fintech right now. Visa, Mastercard, Stripe and PayPal all shipped "agentic payments" frameworks in the past year (tokenized agent credentials, scoped mandates, spend caps), and the card networks' urgency is easy to read: an agent that compares fifty merchants in a second has no loyalty to interchange rails. If agent-to-agent payments settle in stablecoins—instant, programmable, machine-native—the card networks lose the exact volume that's growing fastest.

Big, Fast, and Untrusted

FIG. 03 // SCALE VS FRICTION

Already too big to ignore, not yet safe to automate

AI-agent-influenced retail sales

$67B

≈ 20% of orders already touched by an agent

Shopping queries handled by ChatGPT

50M/day

a product-discovery channel with no ad market yet

Financial institutions expecting agent-driven fraud to spike

78%

87% call trust the #1 obstacle to agentic payments

McKinsey projects $3-5 trillion of agentic commerce by 2030. The constraint is not demand — it is that nobody has solved who is liable when a bot pays with your card.

DATA: MCKINSEY · NEVERMINED · PAZ.AI · 2026

The scale is already material: $67 billion in agent-influenced sales, ~20% of orders. McKinsey projects $3-5 trillion by 2030. And the friction is equally measurable: 78% of financial institutions expect fraud to spike from AI shopping agents, and 87% call trust the #1 obstacle.

The banks aren't wrong. An autonomous buyer is a prompt-injection surface with a credit card. "My agent got scammed" has no chargeback category, no case law, and no clear answer to the only question that matters in payments: who eats the loss? Until that's answered—by regulation or by a network bold enough to price the liability—the cliff stays.

The Quant Angles

Discovery moved; monetization didn't. Ad budgets still assume Google's funnel while 73% of research happens elsewhere. Watch the spread between ad-dependent retail intermediaries and merchants with direct agent integrations—that's a pair trade with a catalyst calendar (every OpenAI/Google commerce announcement).
The payments cliff is the option. Whoever owns agent authorization at scale owns the next interchange. The listed proxies are the networks themselves (defending) and the stablecoin infrastructure from the previous piece (attacking). The 23% checkout number is the KPI to track quarterly.
Agent traffic breaks e-commerce data. Conversion rates, bounce rates, A/B tests—all assume human sessions. A 4,700% growth in bot-mediated visits quietly poisons every dataset trained on "user behavior." If your models consume e-commerce alt-data, segment agent traffic now.
Fraud is the short-term trade, trust the long-term one. 78% of FIs bracing for agent fraud means fraud-prevention vendors get budget before agent-payment vendors do. Sequencing matters.

The honest assessment: agentic commerce is that rare thing—a hyped trend where the demand data is stronger than the narrative. But the last meter of the transaction is guarded by an unsolved liability question, and markets consistently underestimate how slowly payments infrastructure moves. The discovery layer repriced in two years; the money layer will take longer, and the spread between the two is where the returns live.

Building agent integrations, or defending a checkout flow from them? The adoption decks tell one story; the fraud queues tell another.