Stripe’s 2025 annual letter lays out five levels of agentic commerce, and what makes the framework stick is that each level isn’t defined by what the AI starts doing. It’s defined by what the human stops doing.

Level 1: Eliminating web forms. You still research and decide. The agent just handles the typing, the checkout fields, the payment details. It’s a clerk, not a buyer.

Level 2: Descriptive search. You stop searching for specific products and start describing situations. “I need back-to-school supplies for a third grader in Chicago, nothing too itchy or tight, he likes KPop and tennis.” The system reasons across weather, materials, sizes, taste, reviews, and delivery timelines.

Level 3: Persistence. You stop reintroducing yourself. The system already knows your preferences from previous interactions. You’re still choosing, but from a set that already reflects your taste and budget.

Level 4: Delegation. You stop choosing altogether. “Get the back-to-school shopping done. Keep it under $400.” The agent handles search, evaluation, and purchase. You set the constraint and walk away.

Level 5: Anticipation. There is no prompt. The system knows the school calendar, your son’s preferences, and your typical budget. You get a notification that the shopping is done. The things you need show up before you ask for them.

What’s Interesting

The jump from Level 3 to Level 4 is the real trust cliff. Levels 1 through 3 are all variations of “help me decide faster.” Level 4 is “decide for me.” That’s not an incremental shift. That’s a fundamentally different relationship between a buyer and a system.

And Level 5 is stranger than it sounds. If the agent is anticipating your needs and executing purchases without a prompt, who is the customer? The human who never asked, or the agent that decided? Marketing, brand, product discovery, the entire funnel that commerce is built on, all of it assumes a human is paying attention at some point. Level 5 removes that assumption.

This connects to the bifurcation data in the same letter: the sorting machine is already accelerating, with winners pulling away across every industry. Agentic commerce could be the mechanism that makes that sorting even faster. If agents are doing the buying, they’ll optimize ruthlessly on price, quality, and reliability. The companies that are already winning on those metrics get more volume. The ones that were surviving on inertia or brand familiarity lose the one thing propping them up: a human who recognized the name.

The Auction Question

Auctions are interesting here because they already sit at Level 4 for some buyers. Proxy bidding, max-bid systems, auto-bid: these are delegation tools. The bidder sets a ceiling and the system executes. We’ve been doing this for years without calling it “agentic.”

But Level 5 in auctions would mean an agent scanning upcoming sales, recognizing items that match a collector’s profile, bidding within pre-set parameters, and winning lots the buyer never knew were listed. That’s not science fiction. The infrastructure for it mostly exists already.

The question is what that does to the auction experience. Auctions thrive on competition, emotion, the feeling of winning. If agents are bidding against agents, the emotional layer disappears. What’s left is pure market efficiency, which is great for price discovery but terrible for the theater that makes auctions worth attending.

What’s Open

  • At what level does brand stop mattering? If agents are choosing, do they care about brand at all, or just ratings, price, and specs?
  • What does “trust” look like at Level 4? Right now, people barely trust recommendation algorithms. Delegation requires a different order of trust entirely.
  • Is there a Level 6 where the agent sells on your behalf too? Not just buying but liquidating, listing, pricing, negotiating, the full cycle?