Should Your AI Agent Ever See the Card Number? The Case for Credential Wallets
An agentic credential wallet lets an AI agent pay without holding your card details. Here is what the research shows, who is building it, and what to look for.
An agentic credential wallet is a payment layer that lets an AI agent pay using a scoped, revocable stand-in for your card or bank details, so the agent can complete a purchase without ever holding the underlying credentials. If you are building an agent that needs to transact, this is the design choice that decides how much trust, liability and compliance work lands on you.
Agent Pay Safe is building one, so we have a stake in the answer. Wherever we cite research, we name the source and say who funded it.
Why not just give the agent a card?
It is the fastest way to get an agent paying. It is also the pattern now causing friction. Three problems come up.
- Credential exposure. Anything the agent can read, it can log, leak or send to the wrong place. Amazon has reportedly blocked Meta's Muse agent from shopping on its site over concerns that the agent appeared to capture and store customer credentials, and JPMorgan Chase and other large banks have reportedly announced plans to block AI agents more broadly (PaymentsJournal, 21 October 2026). These are trade-press reports of announcements and discussions, not confirmed implementations, but the direction is clear: credentials in an agent's hands are becoming a reason for merchants and banks to say no.
- Compliance scope. Kernel, an agent-browser infrastructure company, says it designed its wallet support so the agent, the application and the browser never handle raw card numbers or CVCs, which it says keeps developers out of PCI DSS scope (Kernel, September 2026). That is Kernel's description of its own design, but the logic is simple: code that never touches card data has less to secure and less to audit.
- No built-in limits. A card number says nothing about "up to £50, at these merchants, for this task." Limits have to be enforced somewhere, and the agent's own judgment is a weak place to do it. In ACI Worldwide and YouGov's June 2026 survey of 2,080 UK adults, 43% trusted AI to follow spending limits, but only 15–18% trusted it with broader judgments such as financial best interest, payment data security or problem resolution. People trust AI with narrow, bounded tasks far more than with open-ended financial ones.
What do users say they want?
Several independent studies point the same way: let the agent do the work, keep the keys away from it, and keep a human checkpoint.
- Global Payments/Worldpay, The Agentic Commerce Report 2026 (8,027 consumers, seven markets, May 2026): 83% say it matters that no one, including the AI agent itself, ever sees their actual card or bank account number, and 57% call this extremely important. Among consumers open to AI, only 21% would let an agent complete a purchase fully autonomously, while 59% would allow it only with approval on each purchase. A confirmation step before every purchase was the top trust-builder, at 34%.
- PYMNTS Intelligence/PayPal (2,179 US adults, June 2026; sponsored by PayPal): 43% say they would likely link a digital wallet to an AI agent within two years. The controls that would make them willing, in order: approval and visibility (41%), spending limits (just over 30%), then purchase-scope and payment-source controls (roughly a quarter each).
- Mastercard, Encoding Trust (26,000 respondents, 16 countries): 85% are open to an agent helping them choose, but only 9% to fully autonomous purchasing. The top concerns were privacy, security and loss of control, which Mastercard defines as the ability to override, cancel or reverse the agent's actions.
- UK: in Worldpay's UK data, 48% are open to agent involvement in shopping, security of payment details is the top concern (54%), and a confirmation step before every purchase is again the top trust-builder (34%).
Worldpay, PayPal and Mastercard all sell payments products, so read these as vendor-commissioned research. Worldpay's report also says its year-on-year comparisons are directional because the 2026 survey design changed.
Who is already building this way?
- Stripe Link and Meta's Muse. Meta launched Muse on 8 September 2026. According to Stripe and Meta, checkout runs through Stripe Link, the consumer approves the total in the chat for each purchase, and Muse does not see the underlying payment details. Where a merchant does not accept Link, Link issues a single-use virtual card scoped to the approved amount. Stripe cites more than 300 million Link users and more than a million accepting merchants (Stripe's figures).
- Kernel. Its multi-wallet support works with Stripe Link and AgentCard. By Kernel's account, its vault hands the agent non-secret aliases and resolves the real credential only at the point the browser makes the request.
- Mastercard. In its 2026 report A Short History of the Future of Shopping and Payments, Mastercard's Chief Digital Officer, Pablo Fourez, advises cardholders to set spending limits, name trusted retailers or categories, and decide in advance which purchases need approval. That is a card network describing the same model.
- Tokenisation. A Bernstein note, as reported by Investing.com on 12 September 2026, expects 100% of agentic transactions to be tokenised.
Treat this as adoption evidence, not proof it works at scale. Muse is only weeks old, the blocks above show merchants and banks have not settled how to treat agents, and "credential wallet" covers several different implementations.
What should builders look for?
This checklist is our reading of the evidence above, not an industry standard.
- The agent holds an alias, not the credential. Ask any provider whether the agent can ever read a card number, even a one-time one.
- Limits enforced outside the agent. Spending caps and mandates should be checked by the wallet before payment, not left to the model's judgment or a prompt.
- Human approval that matches the stakes. Approve-each-purchase for higher-value or unfamiliar purchases, and autopilot within limits for routine ones, with strong authentication such as a passkey for approvals.
- Instant revocation. You and your users should be able to cut an agent's access immediately, and cancelling should be easy.
- A record of what was approved. The Payments Association's survey of 100 UK online retailers (14 May 2026) asked who pays if an agent autonomously completes a disputed £2,000 purchase. Answers split: 24% said it depends on circumstances or contract, 21% shared liability, 18% the AI vendor, 15% the payment provider, 13% the consumer, and 9% the merchant. With no settled answer, a clear log of the mandate and the approval is your evidence. (The working group behind that report includes vendors with a stake in agentic payments; the retailer survey figures are the citable part.)
How does Agent Pay Safe approach it?
Agent Pay Safe is a universal agentic wallet. You connect an agent with a single MCP configuration, add a card or a stablecoin or crypto wallet (held non-custodially, never by APS), set spending mandates, and choose assisted mode (approve each purchase) or autopilot (transact within limits). For each payment, APS checks the transaction against the mandate, asks for approval by passkey where required, and completes checkout itself, so card credentials are never exposed to the agent. We are pre-seed and running a closed beta with early agent builders, and the waitlist is open.
FAQ
Is a virtual card the same as a credential wallet? Not necessarily. A single-use virtual card limits the damage if it leaks, but whether the agent can see the number depends on the implementation. Ask the provider directly.
Does tokenisation solve this on its own? Tokenisation replaces the card number with a scoped network token, which helps, and Bernstein expects it to be near-universal. It does not decide whether a given purchase matches what the user intended, so you still need limits and approval on top.
Will users really approve every purchase? Probably not for everything. In Worldpay's data, 59% of consumers open to AI want approval on each purchase. In American Express's Amex Trendex survey (502 US business leaders and 2,005 US adults who shop online, August 2026, run by Teneo), 52% were comfortable letting an agent buy everyday low-cost items but only 15% high-value ones. That points to setting the approval mode by purchase value. Amex sells payments, and its consumer sample is frequent online shoppers, so it will skew more AI-comfortable than the general population.
Does this matter more for UK builders? The UK data points to reassurance, not resistance. Worldpay's UK respondents are open to agents (48%) but put payment-detail security first (54%) and want a confirmation step (34%). If your users are in the UK, show them the controls before you ask them to trust the agent.
Sources
- PaymentsJournal, "AI Agents Face a Long Road, Despite Muse's Strong Downloads," 21 October 2026 (updated 22 October)
- Kernel, multi-wallet support announcement and documentation, September 2026
- ACI Worldwide/YouGov, survey of 2,080 UK adults, June 2026
- Global Payments/Worldpay, The Agentic Commerce Report 2026 (research by The Lantern; 8,027 consumers, seven markets, May 2026)
- PYMNTS Intelligence/PayPal, The Hidden Cost of Checkout Gaps, September 2026 (2,179 US adults, June 2026)
- Mastercard, Encoding Trust (Signals report), 2026
- Stripe and Meta announcements on Muse and Link, 8–9 September 2026
- Mastercard, A Short History of the Future of Shopping and Payments, 2026
- Investing.com, "Agentic commerce: Is Payments the winner this time?" (coverage of a Bernstein note), 12 September 2026
- The Payments Association, "Agentic commerce in UK retail: An unresolved liability question," 14 May 2026
- American Express, Amex Trendex: Agentic Commerce Edition (research by Teneo), 6 October 2026