Reflections on my reading: Stripe’s vision of the five levels of agentic commerce
Stripe’s 2025 annual letter from its co-founders included a new five-level model for agentic AI that isn’t very agentic and should probably only be two levels, if we’re being honest.
Standard Disclaimer: The “Reflections on my reading” series of this Substack are me using writing as a tool to capture thoughts and mold ideas as I read reports, articles, etc. Writing is a great way to cement what you’ve read and that’s the goal with these. The pieces published in this section are not refined or even really edited. You’ve been warned!
In late February, Stripe co-founders Patrick and John Collison released their annual letter recapping 2025. I finally got around to reading it. A lot of it is corporate back-patting, but there were many parts that jumped out at me surrounding things like Stripe’s ubiquity, the brothers’ view of how the economy is sorting winners and losers, stablecoin progress and, most of all, how they see agentic commerce playing out.
In this piece, I want to focus on Stripe’s “five levels of agentic commerce.” In this section of the letter, the Collisons attempt what seems like a reset in expectations for the current status and ongoing rollout of agent-driven shopping and payments. But, they do it in the most predictably AI-serving way, laying out a five-step hierarchy of agentic commerce that renders the term even more meaningless than it’s already become. I think that’s telling, and worth looking at.
Stripe’s vision of agentic commerce is not particularly agentic
Personal note of annoyance: Substack’s spell checker puts a red line under the word “agentic,” which, at this point, c’mon. But, to be fair, it also puts one under the word “Substack,” so…
On page 9 of Stripe’s letter, the Collison brothers lay out what they see as the “five levels of agentic commerce.” To their credit, they’re trying to be more realistic about the current state of the technology. To their not-so-credit, they’re still very much agent-washing much more basic AI assistance in an attempt to place us further along on the continuum of progress than we really are. But, in any case, I think it’s interesting that we’re seeing a massive company like Stripe trying to pump the brakes on all this agentic commerce talk.
Interesting, but not surprising, especially in light of the recent reporting from The Information that OpenAI Instant Checkout, which Stripe was a high-profile partner in, has been all-but-scrapped because shoppers wouldn’t use it. That system was a key proof-of-concept on the way to agentic commerce that validated whether or not consumers would adopt even the most basic AI-integrated payments. And it failed. Surely Stripe knew this was coming, and it may have influenced the hedging angle to this letter.
In any case, I think we’re in for a long train of these high-profile “well no, not like that” walkbacks from major AI boosting companies. But, anyways, on to the five levels.
Level one: eliminating web forms
Stripe describes level one as follows:
“You research and decide what to buy. But filling out web forms is no one’s favorite way to spend a few minutes. It would be handy if you could simply send the URL to your agent and have it fill out your payment and shipping details, coming back to you with the confirmation.
The system isn’t making any decisions; it’s just typing and clicking “buy” on your behalf.”
What?
First and foremost, there is absolutely nothing agentic about this. It requires a manual prompt and completes a single, simple and well-defined task. This is a perfect example of how just because something involves even slight automation, that doesn’t make it agentic.
But, more importantly…what?
Consider what is being offered here. You still research and decide what to buy. But then, instead of spending two clicks and maybe three seconds checking out using the auto-fill feature provided by Google Wallet or Apple Wallet or Amazon or whatever you‘re running on your device or computer, you instead hand that task off to an AI system, which…clicks buy for you?
Why? What is the value proposition here? It’s true people don’t like filling out forms. Luckily, we’ve had multiple solutions for that, for many years, all of which work exceptionally well and don’t involve AI. Here, Stripe is suggesting level one AI agents as a fix for a problem we haven’t had since like 2015.
But, it’s worth noting that, while involving AI in Stripe’s first level of agentic commerce adds absolutely nothing of value and saves arguably no time, it absolutely does add risk, because AI becomes an unreliable middleman. This is like saying “I already have a perfectly functional system to automatically pay my bills, but I feel like mixing things up, so I’m going to have my mistake-prone child do it instead.”
It makes no sense.
Alright, moving on.
Level two: descriptive search
Stripe describes level two of agentic commerce as follows:
“You stop searching for products or specific attributes and start describing situations.
I need back-to-school supplies for a third grader in Chicago, including clothes (nothing too itchy or tight!), pencils, notebooks, and a lunch box. My son likes KPop Demon Hunters and tennis. School starts in late August.
The system reasons across weather, materials, sizes, durability, taste, reviews, and delivery timelines. Specialized and long-tail products become easier to find. Annoyingly blunt keyword search is no longer a thing.”
This is pure agent-washing. This is not agentic at all and there isn’t even any commerce involved. Presumably, Stripe imagines that after getting the product recommendations, you then use the head-scratching checkout from level one. At least in this case you can argue that it allows you to buy from within the LLM (although, we now know consumers don’t want to do that). But, still, this is just describing LLM search. Which is very useful, mind you. But there is no autonomy here. There is no recursion here. If this is level two of agentic commerce, then telling ChatGPT what ingredients you have in your cupboard and getting back a list of potential recipes is level two of “agentic cooking.”
Even in a letter that aims to dial back expectations for agentic commerce, these big companies can’t help but try to launder it to make it seem like it’s closer than it is.
Level three: persistence
You stop reintroducing yourself.
Find me options for back-to-school clothes for Bobby.
The system already knows your preferences and remembers any requirements, inferred from your previous conversations and purchases. You’re still deciding what to buy, but you are choosing from a set of options that already reflects your taste and budget .
Still requires a manual prompt to begin each task, still requires manual selection, still requires a manual payment authorization. All we’ve added at this level is better memory. This is still not agentic and the value proposition here is still just the same LLM-enabled search we’ve all been using for years now.
Please note that we are now at level three of the five stages of agentic commerce and we’ve yet to see anything notably…agentic. It invokes Ian Malcolm leaning into the CCTV camera and saying “now you do eventually intend to have some dinosaurs on your dinosaur tour, right?”
Level four: delegation
“You stop choosing altogether.
Get the back-to-school shopping done. Keep it under $ 400.
The system handles the search, the evaluation process, and the purchases on your behalf. You trust it will weigh trade- offs as you would and choose things your son will like. All you do is determine the budget. (This is what most people mean today when they talk about agentic commerce.”
At level four, we’ve arrived at a point where the agent actually performs an autonomous task — in this case, making the product selection on its own and authorizing the purchase without requiring direct user sign-off. As Stripe correctly points out, this is starting to look more like the promise of agentic commerce as a technology that actually adds real value.
Level five: anticipation
“There is no prompt.
The system already knows the school calendar, your son’s preferences, and your typical budget. All you do is receive a notification: here’s the back-to-school list of everything that’s been purchased. This is the most futuristic vision, where the things you need show up right before you need them, without you having to ask.”
This level is truly autonomous, truly agentic commerce. Finally.
So, having now taken the journey all the way to the ultimate promise of the technology as something truly useful, what can we take away from Stripe’s five level model?
First, it spends the first three levels attributing basic LLM functions like conversational search to agents. It also fails to introduce any kind of autonomy into the equation until level four. And autonomy is sort of the whole point of agents. The company also aptly notes that, “today, the industry is hovering on the edge of levels 1 and 2.” I.e., we are nowhere near agentic commerce in any meaningful form that creates real value for consumers and actually has a chance of being adopted.
Levels four and five on Stripe’s hierarchy introduce so much risk, at so many levels, for so many stakeholders, and offers so little incentive for merchants, that far from being right around the corner, level four is probably five to ten years away at a minimum, if it ever happens at all, and level five is probably a fantasy.
Stripe calls level five “futuristic,” which is an apt term considering Merriam-Webster uses space elevators, space rangers and the action of Bladerunner as examples under the definition. All of those things also could happen. I’d love to hear where we are on the five levels of Bladerunner.
In all seriousness though, by tacitly admitting that the industry is currently between two levels of near-zero agentic impact, the company is giving away the game. They try to soften the blow by setting the bar for the first three levels of agentic commerce so low that they can be pointed to as progress already behind us, but it’s just more hype and agent-washing in a letter that goes out of its way to say that agentic commerce has been overhyped.
But, the fact Stripe felt the need to publish this attempted re-set of agentic commerce expectations is telling in and of itself. Stripe has invested heavily in this and so it can’t just say “whoops, it turns out it was mostly all bullshit.” So this kind of gentle reframing is probably about as strong a message as we’re going to get. Combine Stripe’s newly moderated outlook with OpenAI’s recent admission that nobody wanted to checkout through its LLM, and it’s becoming clear that AI-powered checkout and payment is not the slam dunk a lot of people seemed to think it would be.
And why? Simple. Trust. We all know this intuitively. It’s common sense.
Everyone is happy to use AI of all types when it’s for low-to-no-risk tasks. Making meme images, writing the first drafts of emails, and yes, even purchase research, all require little to no trust. But AI-driven purchases, and especially autonomous agentic commerce, require a mountain of trust. And it just isn’t there. And it isn’t just around the corner either. It may as well be light years away because LLMs will never stop hallucinating, which means they may never be reliable enough for consumers to trust with their hard-earned money. And if consumers don’t want to use agentic AI, there is no incentive for merchants to adopt it. And if merchants aren’t incentivized to adopt agentic AI, there is no incentive for consumers to use it. It will be a technology without a user base, and bajillions of dollars will have been wasted.