PayPal offers one of the more grounded examples of what it takes to turn AI capability into real outcomes when the thing it moves is money. PayPal operates at a scale that makes the stakes stellar: roughly 400 million users and around 30 million merchant accounts, all inside one of the most heavily regulated corners of technology, financial services. When a company running a system like that describes what actually moves the needle, the details are worth slowing down for.
The three frictions inside every purchase
Anyone building commerce experiences cares about removing payment friction, which comes down to converting intent-to-buy into a sale with the least possible friction. Most checkout failures are really a failure in one of three lanes, and naming them separately shows where to spend the effort:
Friction 1: Validation. The system needs to know that this is a legitimate buyer and a legitimate transaction, and that the person is who they claim to be.
Friction 2: Information sharing. The buyer has to hand over the details a purchase requires, from shipping to payment credentials.
Friction 3: Affordability. Can the buyer actually pay for it, and can the merchant offer terms or financing that make yes possible.
For more than two decades, PayPal has been in the business of taking something complicated and making it feel simple enough that a small merchant will actually use it. PayPal’s answer is a single-click purchase that collapses all three. Because PayPal already knows the device and has authenticated the user, identity is validated the instant the buyer clicks. The wallet supplies balance and payment and shipping information, catalog and inventory systems confirm availability and pricing, and the order drops to the merchant while a security layer scans in the background for anything that does not look right. The output is network plus identity plus trust, delivered in one action.
That has always been the hard part, and it is easy to forget when the excitement shifts to model capabilities or innovative features. The same reasoning transfers to building AI capabilities.
This is the quiet truth of most enterprise AI right now: the capability is not the limiting factor, friction is. This includes the switching cost, the integration effort, the fear of breaking something that works. Small and medium businesses make up a huge share of the economy, and they do not adopt technology because it is clever. They adopt it because it removes a specific pain without introducing a new one. Framing AI as a friction-removal engine rather than a capability showcase is a more honest starting point, and it changes what a company chooses to build.
Maya
The most concrete example is an AI agent PayPal calls Maya, and it targets a pain that anyone who has integrated with a payments platform knows intimately. Every time the provider ships a new version of an interface, for the merchants it becomes technical debt: they need to stop, plan, and do the upgrade work. That work never makes it onto anyone’s roadmap willingly, yet skipping it means falling behind on security and new features.
Maya attacks that directly. It runs inside the customer’s own environment on their own compute, which matters enormously here, because it means proprietary code does not have to leave the building. It watches for new patterns, validates them across a few passes before it commits to anything, builds up a memory of what it has learned, and then automatically upgrades the integration to the latest PayPal version, with validation and certification handled as part of the loop. What used to be a months-long roadmap fight becomes a continuous, self-maintaining process. No more upgrades, as the pitch goes.
The automation is not what makes Maya work. The design decision to run locally and to validate before acting is. Those two choices are what make the tool adoptable inside a security-conscious organization. The hardest part of an agent is rarely the reasoning: it is earning the right to touch production.
The model is to let the agent do the first pass, reach very high accuracy, around 98 percent, and then have humans finish the final, consequential steps. As Srini Venkatesan of PayPal put it, “We still have humans in the loop,” and in a system moving trillions, that is critical.
This did not happen overnight. On the Stack Overflow developer survey, a striking share of developers, around 46 percent, report that they actively distrust AI. By Venkatesan’s account, PayPal’s engineers were deeply skeptical only a year ago, and any CTO watching an AI touch a live payment system has every reason to be nervous.
(Canvas from the 3 Benefits of AI)
From know-your-customer to know-your-agent
The industry has long lived by KYC, know-your-customer. PayPal extended that to KYB, know-your-business. Now agentic commerce forces a third layer: KYA, know-your-agent.
Twenty years ago, when internet commerce was young, it was PayPal’s buyer protection that gave hesitant merchants the confidence to sell online. Trust was the product. In an agentic world, every buyer might be an agent, and in the few milliseconds a transaction takes, both the identity and the legitimacy of a party that is not even human have to be verified. With 400 million customers and the data behind them, PayPal believes it can do that verification instantly, and it is extending the capability across open agent and commerce protocols rather than locking it to one standard. “We are not just reducing friction. We are removing friction completely,” Venkatesan said, and know-your-agent is how that promise survives contact with autonomous buyers.
The assumptions hiding inside all of this
There are two load-bearing assumptions underneath the whole vision that deserve daylight:
The first assumption is that trust can be centralized. The entire architecture assumes that one party, holding enough data, can act as the reliable arbiter of who is legitimate, and that the rest of the ecosystem will accept that arbitration. That works beautifully when the data moat is deep and the protocols are adopted. It becomes fragile the moment agents can be spoofed, delegated, or chained in ways the identity layer did not anticipate.
Verifying a human buyer is a problem with decades of tooling behind it. Verifying an autonomous agent acting on behalf of a human, possibly through several layers of delegation, is a genuinely new problem, and the standards for it are still being written.
The second assumption is that human-in-the-loop does not crack with automation. This sits inside that reassuring 98 percent. Letting the agent go first and having humans finish the remaining 2 percent only works if the humans stay genuinely engaged rather than rubber-stamping a system that is almost always right.
This is exactly the trap Lisanne Bainbridge named in her 1983 essay “Ironies of Automation”: the more reliable the machine, the less practiced and attentive the human supervising it becomes, so the moment their judgment is most needed is the moment they are least equipped to provide it. In a pipeline moving trillions, the tail is where the real risk lives, and the discipline that keeps humans meaningfully in the loop is a cultural achievement, not a technical guarantee.
None of this undermines the approach. It is the thing that has to be actively defended as the system scales.
Key takeaways
PayPal is using AI to sell trust, delivered with less friction, across whatever surface commerce happens to live on next. The one-click purchase at the center of it is deliberately headless. It can sit behind a button on a social feed, a prompt on a television ad, or an agent talking to another agent, and the underlying infrastructure does not change. That is the secret recipe: build the trust and identity layer once, then let it show up everywhere.
For anyone building in this space, the lesson is to stop asking what the AI can do and start asking which specific friction it removes, and whether the people who have to trust it will actually let it run.
- The constraint in enterprise AI is adoption, not capability. Design for friction removal, not capability display.
- Agents earn the right to production. Running locally, validating before acting, and keeping humans on the final steps are what make an agent adoptable.
- Every purchase contains three frictions: validation, information sharing, and affordability. Diagnose which one is costing you the sale.
- Identity is evolving from know-your-customer to know-your-business to know-your-agent, and agentic commerce makes the third layer unavoidable.
- Trust is the actual product. In an agentic world, the party that can verify legitimacy in milliseconds wins the transaction.
- Watch the hidden assumptions. Centralized trust and the human-in-the-loop 2 percent both work only under conditions that must be actively maintained, not assumed.