This week Stripe launched Link balance for AI agents, starting with Instinct (Announced on X by Stripe). Stripe balance is the “agent control dashboard” where Consumers add money, set a per purchase limit, and their agent can buy without asking for approval every time. Read it. On the surface it is a wallet feature. Underneath it is Stripe building a consumer business, because right now Stripe is the only company that can make agentic payments work.
I think this is brilliant, but Issuing banks, merchants, MSPs and other agentic Platforms should take note of Stripe’s very big expansion into consumer direct interaction
What Stripe Launched
The stack has four parts
- Fund. The consumer loads the balance from a card saved in Link. Stripe Crypto Onramp converts the dollars to a dollar stablecoin. Link’s support pages say the balance is held in USDC. The consumer sees dollars, not crypto. Prefunding the wallet solves the authorization problem and gives the agent’s card a lower cost of funds. That structure produces economics for both Stripe and the agent platform. By keeping linked funds in a “balance” Stripe also circumvent Durbin limits on debit wrapping (Durbin prohibits wrapping a non-exempt debit with any other card).
- Hold. The stablecoins sit in a wallet built on Privy, the company Stripe bought last year. Link offers a self custodial version on Privy, where the consumer can export the private key, and a custodial version held by Bridge.
- Permit. The consumer sets the maximum the agent can spend per purchase without approval. Anything above the limit goes back to the consumer. A limit of $0 turns the feature off.
- Spend. Stripe Issuing gives the agent a virtual card backed by the balance. The merchant sees a normal card.
The launch terms are narrow. US only. Identity verification by SSN, last four digits first and all nine if that fails. Loads of $150 or less per setup link. American Express, Discover and Capital One cards cannot fund the balance. Withdrawals to a bank take 3 to 5 business days.
The demand signal is real. Stripe says agentic purchases through Link grew 38x in the past month, and Meta says more than 25% of Muse’s weekly users shop with the agent (TechCrunch, Oct 21, 2026). I covered both early users in Instinct: Hottest Agentic Assistant, Payments Advantage? and Meta’s Muse: The Agent Becomes the Control Point.
The Control Dashboard
Until now the Link agent wallet required the consumer to approve every purchase. UBS described it at Stripe Sessions in May: a one time use card or Shared Payment Token issued per task, with human approval required for each payment (Timothy Chiodo, UBS, “Stripe Sessions 2026: Recap and Industry Takeaways,” May 2026). That is safe, and it defeats the purpose of an agent. Nobody wants to approve concert tickets at 2 a.m.
Link balance moves the approval from the transaction to the account. The consumer sets the rules once, and the wallet enforces them. That is the same construct Walmart chose for Agent Profiles, where the customer sets a spending limit for each agent inside its Walmart account (Walmart Agent Gateway). The difference is scope. Walmart’s controls work at Walmart. Link’s controls follow the agent to any merchant that takes a card.
Why Prefunding Matters
Prefunding solves authorization. Agentic commerce stalls at authorization. Banks will not let an outside agent act as the consumer (Banks are Blocking Agents: Why?), and merchants own the fraud risk on card not present orders (Explaining the Death of OpenAI’s Instant Checkout). With a prefunded balance, the consumer’s bank sees one load, authenticated by the consumer, and never sees the agent. The authorization decision on the purchase moves to Stripe, which runs the program behind the agent’s card. Stripe knows the consumer, the agent, the limit and the funds. It approves against good money it already holds.
Prefunding lowers the cost of funds. Before the balance, every agent purchase was two transactions. Link charged the consumer’s card, then paid the merchant with a single use virtual card. VGS believes those second leg cards are commercial cards, which carry higher interchange (Timothy Chiodo, UBS, “Agentic Commerce Developments with VGS (Instinct, Muse, Stripe Link, etc.),” Oct 2026). Goldman called the same flow “highly reliant on two step transactions” (Eric Sheridan et al., Goldman Sachs, “Entering the Agentic AI Era for the Commerce Landscape,” Sep 2026).
Prefunding addresses key regulatory issue on Durbin and debit. A non-exempt bank debit ($0.21 + $0..05 .. which will get lower) can’t be “wrapped” by another higher cost instrument. See blog Durbin 2 Wrapping Rules.
The structure still has two legs, but the economics change:
| Per purchase pass through | Prefunded balance | |
|---|---|---|
| Funding leg | Consumer card charged per purchase | One load, ideally from a debit card |
| Funding cost to Stripe | Card acceptance cost on every purchase | Debit acceptance cost, once per load |
| Spend leg | Virtual card, interchange earned by issuer | Virtual card, interchange earned by issuer |
| Consumer approval | Every purchase | Only above the limit |
| Credit risk | Funding auth can fail after the agent commits | Funds in hand before the agent commits |
A commercial card can earn around 300 bps. Blended debit cost of funds are 30bps-50bps. The spread between what the agent’s card earns and what the load costs is the margin that pays Stripe is the margin available to Stripe to create economics for the agent. The UBS and Threadd panel reached the same conclusion from the fraud side: prefunded wallets or agent balances “could preserve customer control and improve economics” (UBS, “Identity, Trust, and Fraud in Agentic Commerce webinar recap,” Oct 2026).
Stablecoins Behind the Curtain
In 2025 I wrote Could Stripe Drive eCom Stablecoin Adoption? (A: Maybe, but they wouldn’t do it). I argued Stripe would not push stablecoins against cards at checkout. When Stripe bought Privy, I expected “invisible” stablecoin wallets inside Link (Stripe Acquires Privy: Link Expands as Does Stripe’s Gatekeeper Position).
Link balance is both. The consumer sees dollars. The merchant sees a card. Stablecoins move and store the money in between, where Stripe controls the ledger. Stablecoins are the plumbing, and cards remain the acceptance network.
Stripe Is Building a Consumer Business
Link had over 250 million users at Stripe Sessions in May (UBS, May 2026). Stripe did not set out to build a consumer business. It is building one because nobody else can make agents pay today. My estimate, and it is anecdotal, is that Stripe is the infrastructure for about 90% of US and EU agentic transactions: ACP with Shopify on the merchant side, and Link inside Muse, Instinct and Grok on the consumer side. A year ago I called ACP the best, and perhaps only, operable protocol because Stripe could use its own risk assets to make it work. Link balance applies the same logic to the consumer.
The big news is that Stripe now holds consumer funds. A KYC’d account, a stored balance, spending controls and an issued card. That is a consumer financial relationship. In Stripe/Advent and PayPal: My Thoughts I questioned whether anyone could build a three party network out of a card dependent wallet. Link balance is a small step toward one. Stripe is the wallet, the card program and often the acquirer, and the balance means the card network carries only the last leg.
This Is Not the End Game
Link balance works now. That does not make it the long term agentic model. Four other scenarios are playing out.
1. Merchants own the risk and set agent access. Walmart’s Agent Gateway keeps payment, risk and customer service inside the merchant account. Agents never see a card. To Walmart’s risk engine, a single use virtual card from an agent wallet looks like a guest order on a new card number. VGS adds that the virtual card breaks card on file, subscriptions, follow on purchases and refunds, and that the chargeback process was not designed for an intermediary as merchant of record (UBS, Oct 2026). Large merchants will prefer agents inside their own accounts. I expect them to set the rules, as I argued in Retail, Roles and Know Your Agent.
2. Amex ACE. The consumer registers intent and permissions with the card, and the issuer scopes the credential to that intent (Why Banks Are Best Positioned in Intent: Amex Updates ACE to Provide the Model, following American Express Breaks the Agentic Commerce Deadlock). Amex can do this because it is the issuer and the network. Visa and Mastercard have similar standards, but US issuers will not let the networks expand into authentication and blocking. I described that box in Payment Authorization: Under The Hood. It is no accident that Amex cards cannot fund a Link balance.
3. Google AP2 and DPCs. A technically elegant answer that relies on cryptography instead of governance. It has no traction, and Mastercard is the salesman. I have covered the gap in Google Rolls out Agentic Payments Protocol (AP2), AP2 as Merchant Signals: 4 Scenarios and DPCs Great Idea with a Long Way To Go. A signed mandate proves the chain was not tampered with. It does not say who pays when the agent is wrong.
4. A consumer centric control layer. The consumer needs one place to manage multiple agents, multiple cards, multiple merchants and multiple domains. Link balance manages one wallet’s agents. VGS made the same point: wallets add value by connecting credentials to many agents (UBS, Oct 2026). In Scenario: Agentic Wallets and Federated Data I wrote that wallets will not become agents but will be the most important interface to all of them, brokering identity, authentication, permissions and data on the handset. Phones will be the platform. Apple and Google already hold the cards, the biometrics and the device.
What to Watch
- Load sources. Whether Stripe steers loads to debit, and whether large issuers start declining loads to Link the way they block agents today.
- Card classification. Whether the networks allow a consumer funded agent card to carry commercial interchange as volume grows.
- Disputes. Who handles the claim when the agent buys the wrong item from a stablecoin balance, and how consumer protections apply to the funding leg.
- Limits. The $150 load cap and the US only footprint. Europe has no date.
- Merchant response. Whether more retailers follow Amazon in blocking agents, or Walmart in keeping payment inside the customer account.
- Agent count. How many platforms beyond Instinct, Muse and Grok turn on the balance.
Conclusion
Stripe has built the first working answer to agent authorization: the consumer sets the rules, prefunds the wallet, and Stripe authorizes against money it holds. It makes money doing it, and it shares that money with the agent platform. That is why Stripe is winning today. The long term model will be decided by who owns the risk. Merchants, issuers and the phone platforms each have a claim, and none has given it up.