Visa, BILL and Billtrust said on 20 October 2026 that they have launched “the first live pilot to enable trusted AI agents to initiate and complete approved buyer invoice payments end-to-end through supplier portals.” In the companies’ release, the agent’s authentication “can satisfy the site protections historically provided by CAPTCHA and other bot blocking mechanisms, without requiring an additional agent login.”
The same day, Yelp added an AI agent that phones restaurants for its users, TechCrunch reported, and an Axios reporter wrote that a Meta agent got stuck at a retailer’s sign-in page. Below we set out what each of these agents can now do with money or a phone line, and the limits each company names.
Business Pill · AI AGENTS
A one-minute explainer of AI agents, from answering to doing. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, the agent that pays in Visa’s pilot is the one with the least discretion: its FAQ says it only executes payments the buyer has already approved.
Update (26 Oct): the banks are on the other side of this. While card networks pilot approved agent payments, major US banks are detecting unauthorized third-party AI agents operating inside customer accounts, Axios reported on 22 October, citing people familiar with the issue (Axios). These are generally cases where an agent logs in with a user’s permission and credentials, which goes against most banks’ rules that humans log in to their own accounts; some agents have contacted call centres or appeared to act independently of user instructions, the people said.
Morgan Stanley has begun blocking third-party agents from its platforms, and JPMorgan, Bank of America and Citi are also restricting them, according to Axios. Most banks have authorized or are moving towards read-only access, which lets agents see account information without taking action. A Morgan Stanley spokesperson said the bank has not observed “rogue AI agents accessing our platforms”, and a Chase spokesperson said customers should not give their usernames and passwords to anyone.
Caveat: Axios gives no count of incidents or affected accounts, and its sources are unnamed.
The release says the flow begins when the BILL Payment Agent, “registered using Visa’s Trusted Agent Protocol and Agentic Directory”, uses tokenized payment credentials and “controls aligned with the buyer’s approved payment instructions to issue payments on behalf of BILL’s customers.” Billtrust then uses the same protocol “to recognize, authenticate, and grant access to the authorized agent”, according to the release.
The companies list five things the pilot enables, including “tokenized payment credentials rather than shared supplier-portal logins” and “payment controls tied to the customer’s approved payment instruction”. In the release’s FAQ, the answer to whether the AI agent is making payment decisions is: “No. The agent only executes eligible payments that have already been approved by the buyer.”
“Businesses should not have to choose between automation and control when making payments,” Visa’s Mark Nelsen said in the release. BILL’s Archana Prasad said that “nearly 500,000 businesses trust BILL”. The release says the pilot “is currently live and processing end-to-end transactions with participating buyers using BILL and Billtrust suppliers.”
Bloomberg’s post on X described the partnership as having AI agents “initiate and complete business invoice payments without the need for human authentication”.
Cost per 1,000 decisions in Nace AI’s own benchmark table (19 Oct): Claude Opus 5.5 $10.17, GPT-6 Sol $3.33, GPT-6 Luna Decisions $0.10 and Drex Max $0.06; rival figures were measured by Nace. Visa, BILL and Billtrust say their pilot’s agent only executes eligible payments that have already been approved by the buyer (release, 20 Oct).
Bloomberg’s post of 20 October 2026 describes Visa partnering with BILL and Billtrust so that AI agents initiate and complete business invoice payments without the need for human authentication. The companies’ release adds that the agent only executes payments the buyer has already approved.
An Agent on the Phone
Yelp’s new feature, Check for Me, will show up on restaurant pages in its app, where users enter their party size and phone number, TechCrunch reported on 20 October 2026. “The AI agent then calls on behalf of the user to check wait times and texts them the results,” TechCrunch wrote.
TechCrunch also noted that “there is a possibility that a user’s agent will talk to the restaurant’s agent”, since Yelp already sells a voice product to restaurants.
Where Consumer Agents Still Stop
In Axios’s 20 October test of OpenAI’s Dots and Meta’s Muse, “Muse attempted to complete the purchase but got stuck at Target’s sign-in page,” the reporter wrote, while “Dots instead directed me to the Target app, where my account and payment information were already saved.” The reporter added: “I’m not yet comfortable authorizing purchases of hundreds or thousands of dollars, or at a point of trusting them enough for consequential decisions.”
Hark’s free-pizza promotion, which we covered on 20 October, offered $30 in DoorDash credit for orders placed through its agent. Hark says it ultimately delivered nearly 20,000 pizzas, while Axios’s reporter got the credit but no pizza, according to Axios.
Update, 28 October 2026: Hark’s own post on X on 17 October said: “Our pizza party was incredibly crazy, tens of thousands delivered.” It gave no exact count; the nearly 20,000 figure is Hark’s number as Axios reported it.
The Decision Layer Underneath
Two products we covered this month sell the step before an agent acts. OpenAI’s Decisions API returns a probability, a choice from a fixed list or a score, with input priced at $0.10 per 1M tokens and no output charge, as we reported on 7 October. Nace AI said on 19 October that Drex Max, its decision model for computer use, costs $0.06 per 1,000 decisions in its own benchmark table, as we reported on 20 October.
Update, 21 October: Mastercard Opens Agent Payment APIs on x402
Mastercard followed on 20 October. In a release timed to Cloudflare Connect, it said its Agent Pay for Machines “builds on x402, an open protocol that lets digital services request payment online, and connects it with the trust, security and reach of the Mastercard network.” After pilots with Cloudflare, Ant International, AIsa, Base, BVNK, Coinbase and Hype, Mastercard said it is “opening Agent Pay for Machines Early Access, giving developers access to the APIs, tools and documentation needed to begin building and testing agent-to-agent commerce experiences.” (Mastercard)
Mastercard described the flow this way: when an agent meets a paid service, the service responds with “402 Payment Required,” and the agent “creates a digitally signed payment voucher, which is a secure promise to pay within its approved limit.” As we read it, both card networks are now building the same control the Visa pilot above relies on: a spending limit set by the business before the agent pays.
As we read it, the pilot moves trust from a password to a registry: the release says the BILL Payment Agent is registered using Visa’s Trusted Agent Protocol and Agentic Directory, and Billtrust recognises it instead of asking for a supplier-portal login.
As we read it, the consumer side is still at the sign-in page: in Axios’s test, Meta’s Muse got stuck at Target’s sign-in, and the reporter wrote that they are not yet comfortable authorizing large purchases.
As we read it, the step before a payment is also being priced as its own product: OpenAI’s Decisions API charges only for input, and Nace AI’s own table lists Drex Max at $0.06 per 1,000 decisions. Visa’s release gives no price for its pilot.
Kunal Patel, Billtrust, in the 20 October release
“Agentic commerce will only move forward if businesses can trust agents to operate across systems with the same clarity and accountability they expect from people”
Three Implications
PRE-APPROVED PAYMENTS ONLY The release’s FAQ says the agent only executes eligible payments that have already been approved by the buyer.
NO SHARED LOGINS The release lists tokenized payment credentials rather than shared supplier-portal logins.
AGENTS ON THE PHONE TechCrunch reports Yelp’s Check for Me agent calls restaurants for users and texts back wait times.
The framework puts it this way: “The limiting factors are organizational readiness, governance, workflow redesign, training, and trust.”
As we read it, Visa’s pilot works on governance and trust: an agent with its own registered identity, paying only what a buyer has already approved. The framework was published in July 2026 and does not discuss Visa, BILL, Billtrust or Yelp.
What Is Not Established
The Visa release gives no transaction count, payment volume, fee or number of participating buyers, and does not say when the pilot could widen. TechCrunch’s report does not say which markets get Check for Me first. Hark’s pizza count is the company’s own, and Drex Max’s cost comparison comes from Nace’s own table.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This piece is part of this week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
In one day, Visa, BILL and Billtrust said an AI agent is paying approved invoices through supplier portals, and Yelp said its agent will phone restaurants. In the Axios test, a consumer agent still stopped at a sign-in page. As we read it, the agents that pay are the ones whose permissions a buyer approved in advance, and Visa’s pilot writes that limit into its FAQ.
A note on sourcing. This piece is built on the Visa, BILL and Billtrust release of 20 October, TechCrunch’s report on Yelp and Axios’s agent test, each read in full, Bloomberg’s post on X, and our own earlier pieces on Hark, OpenAI’s Decisions API and Nace AI’s Drex Max. The decision costs in the chart are Nace AI’s own figures. Nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
