Agentic Commerce and PCI DSS: Are AI Agents in Scope?
Payment Card Assessments | Artificial Intelligence | PCI DSS
Published: Nov 20, 2026
A few weeks ago, I bought a new pair of running shoes with the help of an AI chat bot. I usually pay between $160 and $200 for that kind of shoe. This time, I only paid $95.
The catch was, the 15 to 20 minutes I spent going back and forth with the assistant, because some of the links it found didn't work, and in the end, I still had to click through and complete the payment myself.
That friction is exactly what AI companies want to remove with agentic commerce. This raises a question the payments industry has not yet fully answered: when an AI bot or agent takes your card data, who is responsible for protecting it?
Why AI Companies Want to be Your Shopping Agent
Much of the revenue in AI today is derived from enterprise contracts, with large organizations buying tokens or subscriptions to make their workforce faster and more productive.
Selling directly to consumers is another big opportunity. A bot that can find a product, compare prices, and complete the purchase is one of the most compelling consumer features possible.
This is agentic commerce: an AI agent that finds, decides, and pays on your behalf.
The Grok Moment in Agentic Commerce
In mid-October 2026, Elon Musk posted on X that to buy things, you could take a picture of your credit card, drop it in the chat, and Grok Bot will find the best deal and order it.
The reaction was immediate. A community note pointed out that AI warns users not to include personal information in prompts, yet credit card photos may contain that data.
The terms of service also state that the company is not responsible for User Content or Agentic Actions. Yet, Musk later said the company would “make you whole” if the bot makes a purchase mistake or leaks a card.
Some of the reaction was shaped by who posted it, and I suspect the response would have looked very different if another provider had made the same announcement. Multiple AI providers are building agent-driven purchasing features, so this is not unique to one company. What is unique is that the company behind Grok is the one talking about it openly.
I'm not endorsing the approach, but that openness is a good thing. It makes the conversation public, where security professionals, the payments industry, and consumers can weigh in, which is far better than these capabilities arriving quietly.
One more point on fairness: the product's own changelog describes a more controlled flow than the post suggests, one where the bot shows an amount for the user to Allow or Deny, then a link to a payment processor completes the payment. That difference matters, because how card data actually reaches the AI agent determines the PCI scope.
How Agentic Commerce Actually Works
The details vary by provider, but most AI agent purchases follow the same basic path. Understanding this path helps demonstrate where card data can be exposed.
Steps one through four are about intent and consent. Step five is where PCI DSS considerations begin, because this is where card data either stays protected or lands somewhere it shouldn't.
Two Ways an AI Agent Can Pay
- Direct Card Interaction: The consumer types or photographs the card, and the AI provider sees the full PAN, expiration date, and possibly the security code. That is Cardholder Data (CHD) in a chat window.
- Tokenized or delegated credentials: The agent never holds the real card number. It uses a token tied to a specific agent, purpose, or spending limit, and the consumer authenticates through their bank or wallet.
The first model is easier for a consumer to grasp, but the second is much better for security.
Why a Photo of a Card Matters
A picture containing a payment card’s information is not a harmless attachment. The front typically shows the PAN, the cardholder name, and the expiration date. The back typically shows the security code, which PCI DSS treats as sensitive authentication data. Issuers are the only entities permitted to retain them after authorization.
Once that image is in a chat, several questions follow:
- Is the image being stored, and if so, for how long?
- Is it written to logs, caches, or backups that other systems and staff can access?
- Can the PAN or security code be extracted from the image and reused?
- Is chat content used for anything beyond the purchase, such as troubleshooting, human review, or model improvement?
I'm not suggesting any specific provider does these things. The point is that a card image is CHD, and the answers to these questions are what determine whether the data stays protected.
What Agentic Commerce Means for PCI DSS Scope
PCI DSS applies to any entity that stores, processes, or transmits CHD, or that can affect the security of the cardholder data environment. The label on the entity doesn't change that.
Here is how the common agentic scenarios look through a scoping lens. These are illustrations, not formal scope determinations, because real answers depend on the specific data flows and technologies involved.
- The AI provider receives a raw PAN. Whether through a photo or typed text, the provider is transmitting, and likely storing or processing, CHD. This is the clearest case for PCI DSS coverage, and the requirements for storage, transmission, access control, logging, and incident response all apply.
- The AI provider receives only a token or approval signal. The provider's exposure to CHD can shrink considerably, but scope doesn't vanish automatically. How the token is created, where credentials are held, and who can influence the payment flow all still matter.
- The merchant receives an agent-initiated transaction. The merchant's checkout still handles payment, so its existing obligations remain. What changes is the need to identify that an agent is acting and to understand how that traffic reaches the payment page.
- A third party sits between the agent and the merchant. Wallets, payment facilitators, and orchestration providers may each touch credentials, and service provider management mandates (such as those in Requirement 12.8) become relevant.
In every case, the question is the same: who in the chain actually touches CHD, and are they protecting it?
What the PCI Standards Say Today
In September 2026, the PCI Security Standards Council published Security Considerations for AI Systems, providing guidance, not mandatory requirements.
Its core scoping position is that AI systems should be treated like any other technology when deciding which PCI requirements apply. Its recommendations include a “least agency” approach that limits each AI system to only have the access it needs for the task, and that access should be limited by separate controls, not by prompts to the model. It also says some actions should require human approval. For agentic AI, the main advice is to keep access narrow and use separate controls.
The document also says plainly that it does not address the use of AI to perform payments, such as agentic commerce. It covers securing the data given to AI systems and defending against the malicious use of AI, but not the AI shopping agent itself.
What Card Networks Think About Agentic Commerce
The payment networks have not been idle in this conversation.
Several have released frameworks that register and verify agents, issue tokens scoped to a single agent, and capture a record of what the consumer intended. In its published material on agentic commerce, one network says it is collaborating with issuers, acquirers, PSPs, merchants, and AI platforms to establish suitable liability frameworks.
That notion is easy to skim past, but it is important. The industry is telling us that the liability rules around agentic commerce are still being written.
Even in the clearest scenario, where the AI provider is handling raw CHD directly, there is still no established mechanism that mandates that provider to comply with PCI DSS at this point. It is typically not the merchant of record for the purchase, and it is not acting as a service provider to the merchant either. It is working on behalf of a consumer, one transaction at a time, which puts it in a gray area outside the categories PCI DSS applicability was built around.
How to Approach Agentic Commerce Now
- Consumers: Avoid sending card photos or card numbers to an AI bot unless the provider clearly explains how that data is handled. Prefer flows that use your bank's or wallet's own approval step, use virtual card numbers or spending limits if your issuer offers them, and regularly review your statements closely.
- AI builders: Keep raw card data out of the conversation wherever possible. Use tokens and processor-hosted payment steps, enforce approvals and spending limits outside the model, avoid logging or retaining payment data, and talk to a QSA early to understand scope.
- Merchants: Ask how agent-initiated traffic reaches your checkout, update your data flow diagrams, and confirm expectations with your acquirer or processor.
- QSAs and assessors: Ask who in the chain handles CHD, document any agent in the scope and data flow review, and don't assume a token means the question is answered.
What to Expect Next with Agentic Commerce
The frameworks released so far assume a registered agent and a tokenized credential. They don't clearly address an AI provider that receives raw card data in a chat window, and they don't say who bears the responsibility when something goes wrong.
In Part 2, we'll look at the accountability gap, and why “who makes them comply?” may be the most important question nobody has answered yet.
About Matt Crane
Matt Crane is a Director at Schellman, where he excels in project management and client relations while overseeing assessments against various PCI Standards. With a primary focus on PCI DSS Compliance for organizations spanning diverse industries, Matt leverages over a decade of expertise in information security services. In addition to leading PCI engagements, Matt conducts multiple PCI workshops each year and frequently speaks on PCI-related topics at industry events. He is well-versed in SWIFT CSCF and regularly collaborates with Fortune 500 companies to navigate complex compliance challenges. Prior to joining Schellman in 2017, Matt held key positions in both the private and public sectors, specializing in PCI and NIST assessments, as well as intelligence analysis. His extensive background includes performing risk assessments and providing consulting services for merchants and service providers across multiple industry verticals. Matt holds a BBA in Information Security and Assurance as well as several industry certifications, including CISSP, CISA, CRISC, and QSA. With an exceptional track record and a profound understanding of the industry, he ensures clients receive unparalleled guidance in achieving their compliance goals.