Every AI-finance story I’ve covered this month has been about access. Claude wanting to read your bank account. ChatGPT already reading it, gated behind its priciest subscription tier. Perplexity doing the same through Plaid. Read access. Look, but don’t touch.

That changed this week. Mastercard announced Agent Pay — a partnership with a startup called Alchemy that lets you connect an AI agent to your actual Mastercard and hand it a virtual card of its own. Spending cap. Merchant restrictions, if you want them. And by default, no per-purchase approval. The agent just buys.

That’s not a bigger version of the same story. It’s a different risk category entirely, and I don’t think enough people clocked the distinction yet.

Quick Facts

Announced	Sept. 17–18, 2026, via Mastercard’s partnership with Alchemy
What you get	A virtual, tokenized, one-time-use Mastercard credential issued directly to your AI agent
Default spending cap	Set by you (a dollar amount), valid for 7 days
Per-purchase approval	Not required by default — you can opt into a “connected” mode that requires passkey approval each time
Where it works	Any online merchant that accepts Mastercard
Dispute trail	Mastercard’s “Verifiable Intent” record, which logs what you authorized vs. what the agent did
Who’s already wired up	AgentCard supports both Visa and Mastercard credentials; Meta’s Muse and xAI’s Grok can already complete purchases through other rails
Liability if the agent messes up	Not specified by Mastercard as of publication

What Mastercard Agent Pay Actually Is

Mastercard Agent Pay is a framework that lets AI agents make purchases with tokenized, disposable Mastercard credentials tied to your existing card — no new account, no new card number to memorize, just a temporary financial identity issued to software instead of a person. The version that shipped this week runs through Alchemy’s AgentCard product, which gives an agent a dedicated email address, phone number, stablecoin wallet, and now, one-time-use Mastercard payment credentials, all set up through a command-line tool in under a minute on the developer side.

For you, the part that matters is simpler. You connect an AI agent — Grok, Meta’s Muse, whatever agentic assistant you’re already using — to your Mastercard. You set a spending limit. Maybe you restrict it to certain merchant categories, maybe a price range, maybe a geography. Then the agent goes and buys things. Groceries, a hotel booking, a subscription renewal. It doesn’t ask first unless you told it to.

Mastercard isn’t first with the concept — Agent Pay itself dates back to an April 2025 unveiling, and a machine-to-machine version for API-level purchases launched back in June. What’s new is the consumer-facing piece: an actual person handing an actual agent a card with actual spending authority, live, this week, through a partner already building the plumbing.

The Default Setting Is the Story

Here’s the sentence to sit with: the default credential requires no per-transaction approval. You set the cap — say, $200 over 7 days — and the agent operates inside that box without checking in. Mastercard does offer an opt-in “connected” mode that locks purchases to specific merchants and amounts and requires a passkey tap before anything finalizes. But that’s the opt-in. The default is closer to giving a very literal-minded employee your card and a spending limit, then leaving the room.

Compare that to what’s already live elsewhere. Meta’s Muse, which launched September 8 with Stripe-powered checkout, uses a single-use virtual card scoped to one merchant and one specific amount — and TechCrunch reported it still asks for your approval before completing a purchase. Mastercard’s default setting skips that step entirely. Whether that’s a feature or the actual problem depends entirely on how much you trust an LLM’s judgment about what counts as “within budget.”

What Happens If Your AI Agent Overspends?

Mastercard built a specific answer to this, and it’s worth knowing before you connect anything:

1. Every agent transaction gets logged against Verifiable Intent — a record, built with Google, documenting what you actually authorized and what the agent executed. It’s the trail Mastercard points to for resolving disputes.
2. The spending cap and 7-day window are hard limits at the network level, not just a polite request to the agent — once the credential’s cap or clock runs out, it stops working until you reissue it.
3. You can revoke access at any time, which functionally kills the agent’s ability to transact going forward, the same way you’d freeze a card.
4. Liability for what happens in between is the part Mastercard hasn’t spelled out. Fortune reported that when asked directly who’s on the hook if an agent makes an unauthorized or mistaken purchase, Mastercard pointed back to the Verifiable Intent record rather than giving a straight answer.

That last point is the one I’d actually worry about. A transaction log tells you what happened. It doesn’t tell you whether you’re getting your money back.

You’re Not the Only One Skeptical

Fortune’s reporting cited a survey finding only 7% of U.S. and U.K. consumers said they’d let an AI assistant make purchases without approval — with over half saying they’re outright uncomfortable with the idea. Most people, given the choice, want price alerts and comparison shopping from an AI, not a standing authorization to spend their money. Mastercard shipped the no-approval default anyway, which tells you this is being built for where the industry expects to be in a couple of years, not where most cardholders are right now.

I don’t think that instinct is wrong, for what it’s worth. We’ve already flagged how unreliable AI chatbots are with numbers that matter — tax refund miscalculations north of $2,000 in independent testing. A model that gets your tax math wrong is a model I’d want double-checking a $200 spending cap before I hand it standing purchase authority.

The Race Behind the Race

Mastercard isn’t moving alone, and it isn’t moving first on the standards side. Days before this launch, Mastercard, Visa, and Ant International announced a joint “Know-Your-Agent” framework, meant to let card networks, wallets, and marketplaces recognize a trusted agent across ecosystems instead of each network running its own silo. It’s trying to reconcile three separate proprietary systems — Visa’s Trusted Agent Protocol, Mastercard’s Verifiable Intent, and Ant’s Agentic Mobile Protocol — into something that actually talks to itself.

Visa isn’t sitting still either. It launched its own Trusted Agent Protocol as part of Visa Intelligent Commerce, and — worth knowing if you’re comparing cards — Alchemy had already integrated Visa credentials into AgentCard before Mastercard joined this week. That means the same AgentCard-powered agent can, in principle, hold a disposable Visa credential and a disposable Mastercard credential side by side. The card networks are competing on whose agent-payment standard wins adoption; the practical effect for you is that whichever card you carry, an agentic version of it is coming whether or not you asked for one.

Read Access vs. Spend Access: Why This Isn’t the Same Story

I want to be specific about why this is a different risk category than what I’ve covered with Claude, ChatGPT, or Perplexity linking to your bank account.

Read access means an AI can see your transactions and answer questions about them. The worst-case failure mode is a privacy breach — your spending history, your recurring charges, the shape of your financial life exposed to whoever compromises the AI account. Bad. Genuinely bad. But it’s passive. Nothing moves without you.

Spend access means the agent can initiate a transaction that moves real money, inside a box you defined, without asking. The worst-case failure mode isn’t just exposure — it’s an actual unauthorized charge, and now you’re negotiating with a card network’s dispute process about whether a bug, a prompt injection, or a model just being confidently wrong counts as “acting within your intent.” Mastercard’s Verifiable Intent record is built for exactly that argument, which should tell you the argument is expected to happen.

Both matter. But if you’re triaging where to be cautious first, spend access is the one with money actually leaving your account before you’ve reviewed anything.

Should You Turn This On?

Depends entirely on what you’re actually trying to automate.

• If you’re using an agent for genuinely repetitive, low-stakes purchases — a recurring grocery order, a subscription you already pay for manually every month — a tight spending cap with merchant restrictions is a reasonable trade. Set the cap low. Set it lower than you think you need.
• If you’re tempted to hand an agent open-ended “handle my shopping” authority, don’t, not yet. The 7-day validity window and dollar cap are real guardrails, but they cap the damage per credential, not the number of credentials you might reissue on autopilot.
• If you’re not sure your card issuer even supports this, check before you assume it’s live — Agent Pay rolls out through Mastercard-certified processors and partner banks, not universally overnight. Citi and U.S. Bank are the first named partners for the expanded AI shopping tools; if yours isn’t one of them, nothing changes for you yet.
• If you want the safer default, use the opt-in “connected” mode that requires passkey approval per purchase until Mastercard names who eats the cost when Verifiable Intent and reality disagree. A dispute record isn’t a refund guarantee.

The Bottom Line

Three card networks are now racing to build a shared way for an agent to prove it’s trustworthy — Know-Your-Agent, Visa’s Trusted Agent Protocol, Mastercard’s Verifiable Intent — and not one of them has published what happens once that trust turns out to be misplaced. That’s the actual gap here, and it’s specific to payment rails in a way read-access risk isn’t: a verified, spending-capped agent can still be a verified, spending-capped agent that bought the wrong thing, and right now the networks have built the record-keeping for that dispute without settling who wins it. Until Mastercard answers the liability question instead of just logging it, I’d leave the no-approval default off.

---

Reporting based on Alchemy’s September 17, 2026 announcement via PR Newswire, and coverage from Fortune, American Banker, and TechCrunch. Terms, availability, and liability policies for Mastercard Agent Pay may change; verify current details with your card issuer before connecting an AI agent to your account.