AI Agent Pay: Why AI Agents Will Rewrite Payments Strategy
and how you can make sure you are not left behind.
For thirty years, the networks have sparred over loyalty perks, cross‑border fees, and token form factors. Last week, they did something bigger: they plugged autonomous AI agents straight into their rails.
Visa calls the program Intelligent Commerce, Mastercard unveiled Agent Pay, and PayPal rolled out an Agent Toolkit aimed at independent developers.
In a fast-developing AI race, we have gone from prompting to letting AI Agents do tasks for us in the background, with payments becoming the next frontier.
This means that once the screens disappear completely and the software can spend for itself, the rail holding the credentials wins. A future that looked far, far away has quickly become real, and could potentially shift the gravity from user interface to machine-to-machine execution.
This leaves all of us building payment companies wondering where we fit in this new era.
In this newsletter, I will break down what an AI Agent really is, how the agentic rails introduced by Visa, Mastercard and PayPal differ from each other, why they decided to go all-in on this, and what the impact will be for others, if they don’t act, and share what I believe the strategy needs to be for payments companies looking to get in on the action.
Let’s dive in…
What an AI agent really is
Payments professionals know automation; they live with rule engines and batch scripts.
An AI agent goes further.
OpenAI’s field guide describes it as software that can decide, act, and finish multi‑step tasks without asking for every instruction. It carries three essentials:
A model that reasons.
Tools, APIs it can call, such as “create_token” or “void_auth”.
Instructions and guardrails that limit scope, route edge cases to humans, and cap risk.
Agents are a step beyond chatbots: they break work into subtasks, call other agents, learn from results, and return when the job is complete.
In plain English, an agent is a junior colleague you hire in code. Give it a budget, a rulebook, and the keys to safe APIs, and it comes back with a settled transaction.
Why now?
Cheap multimodal models, structured tool‑calling APIs, and memory primitives all converged in just the last quarter. Meaning that the barrier to writing a “buy‑this‑flight” workflow dropped from a month of brittle integration work to just five lines in a framework.
Compare this to the lengthy documents merchants would receive when implementing a PSP back in the early 2000’s, versus when Stripe came out with just 3 lines of code.
The new agentic rails and how they differ
This week was especially eventful, as both Visa and Mastercard released their versions of AI Agent Pay.
Let’s break down what each one actually is, as well as PayPal’s Agent Toolkit, which was released in April.
Visa Intelligent Commerce wires tokenization, authentication, and risk scoring into one endpoint.
The agent never handles raw card data; it receives a network token and spending limits, then executes on VisaNet. They have partnered with OpenAI, Microsoft, and Nvidia to ensure this becomes a success. Basically, Visa is betting that if it owns the credential vault, every future shopping bot will remain a Visa cardholder.
Mastercard Agent Pay anchors on the Multi‑Token Network.
The agent requests a “Pay on behalf” authorisation that can settle through a classical card, a bank‑linked token, or a regulated deposit token. Mastercard markets the programme to banks and large merchants that want issuer‑level visibility and control. Focusing on a more merchant-driven rollout, Mastercard has decided to partner with Salesforce, Shopify, and Adobe Commerce, a signal that they are more focused on a push into embedded retail flows.
PayPal Agent Toolkit chooses speed.
It ships a Python and TypeScript library that snaps into popular agent frameworks like OpenAI’s SDK and LangChain. The Toolkit wraps order, invoice, dispute, and payout APIs so any vertical SaaS developer can drop in a full payments stack without waiting for scheme approval. PayPal concedes network share of interchange but aims to win developer mindshare when the long tail of agents needs a ledger tomorrow.
Three philosophies emerge:
Credential first (Visa): Protect the PAN‑to‑token map and keep scheme rules intact.
Rail‑agnostic (Mastercard): Abstract the rail, win by optionality.
Developer‑speed (PayPal): make installation the differentiator, not interchange.
What was the strategy behind this move?
In a very short period of time, AI companies have learned that Agents convert intent actions. For payments, that means that Agents can convert an intent directly into revenue.
Today, ChatGPT can recommend a product, but a user still needs to fill out a checkout form. Tomorrow, an AI assistant can book the item, file the expense, and confirm shipping all inside one conversation.
The company that clears the payment captures the value while the interface fades.
In hindsight, we have learned that for the networks, the tokenization lesson still stings. Apple Pay proved that whoever writes the token spec controls the front door to the consumer.
Agents are the next token wave: a tokenised instruction set for autonomous spend.
Networks cannot risk Amazon or OpenAI standardising their own “pay” command on top of open banking APIs or stablecoins. Plugging their rails into agent frameworks is a defensive play dressed as innovation.
PayPal’s stride is a different story. It lacks a card network to defend, so it optimises for the fastest on‑ramp.
The company hopes thousands of niche agent workflows, property management pay‑outs, healthcare invoicing, and game micro‑transactions will adopt the Toolkit before acquirers even finish their risk reviews. If that happens, PayPal’s wallet could become the default ledger for a number of autonomous commerce.
Who pulls ahead, who falls behind
Early advantages will be beneficial to PSPs that already expose granular APIs.
An agent only needs milliseconds to “authorise”, “capture”, and “refund”; platforms that still batch into nightly files cannot respond in real time.
Vertical software firms, such as construction management or B2B marketplaces, gain a chance to introduce hands‑free billing tiers that settle, reconcile, and post to accounting without human clicks.
Meanwhile, incumbent acquirers tied to mainframe batch windows will have to watch processing volume flow to agent‑ready competitors.
Banks that ignore non‑card instant rails risk losing high‑frequency micro‑payments to wallets that route over ACH Instant, Pix, or SEPA Insta.
Manual review teams become a drag when autonomous checkouts promise zero‑click settlements and lower abandonment.
What should your strategy be?
Don’t get it twisted, the decision to adopt, partner, or build needs to come from the boards and executive teams, not engineering leads.
Here is how you do that:
First, audit the customer journey.
Where would an autonomous checkout collapse friction and unlock revenue? Subscription renewal, loyalty redemption, or supplier disbursement often top the list.
Second, test the resilience of your stack.
Can your core gateway accept an authorisation with no browser data, no 3‑DS challenge, and a risk callback in under 200 milliseconds? If not, any agent strategy begins with modernising the risk engine, not model selection.
Third, open a structured dialogue with your scheme partners.
Visa and Mastercard are courting select pilots. Early access can translate into preferred fee schedules and influence the API specs. Late entrants will consume whatever is published.
Fourth, explore alliances beyond the schemes. PayPal, Adyen, Stripe, and emerging orchestration layers will each push their own agent wrappers. A multi‑rail hedge preserves negotiating power.
Fifth, build a roadmap for proprietary intelligence. Even if you run over scheme APIs, ownership of the decision models, fraud scoring, financial optimisation, and dynamic settlement choice keeps margin in your hands rather than the network’s.
Three Strategic Paths to follow
Essentially, there are three paths that a Payments company can follow in the next 12 to 18 months.
The full‑builder path suits a handful of large PSPs and top‑tier acquirers with deep engineering benches.
They will craft their own agent orchestration layer, embedding proprietary risk and routing logic. Capital expenditure will be heavy and time‑to‑market slow, but it will allow them to keep control of the fees they will set and allow them to white‑label their partners.
The fast‑adopter path appeals to mid‑market providers and vertical SaaS firms.
They will wrap Visa Intelligent Commerce, Agent Pay, or PayPal’s Toolkit, then differentiate on the domain workflow rather than core payments. For them, speed to launch is the edge, but dependence on partners’ roadmaps is the trade‑off.
The aggregator path belongs to orchestration start‑ups and switch vendors.
They will aim to integrate every agent, rail, schemes, wallets, and real‑time bank payments into a single API. Merchants will gain hedging flexibility while the integrator will shoulder reconciliation complexity and dual risk reviews.
Each path solves a different equation of control, capital, and speed. What matters is that leadership picks one deliberately instead of drifting.
In Conclusion
According to McKinsey’s macro analysis, enterprise value from generative agents is pegged at up to $4.4 trillion yearly. Payments will capture a big part of this slice because every agent interaction ends with a settlement event.
The shift from click‑based commerce to intent‑based commerce compresses the funnel and magnifies the role of the underlying rail.
Expect three waves:
Embedded assistance (next 12 months). Conversational copilots in travel, retail, and SaaS trigger payments through scheme APIs. Volume is incremental but proves trust.
Invisible purchasing (12–36 months). Agents placed inside consumer banking apps negotiate bills, roll subscriptions, and arbitrage prices without notifications. At this stage, networks absorbing risk cement their relevance.
Agent‑to‑agent marketplaces (beyond 36 months). Enterprises run fleets of specialised agents that settle with one another across multi‑rail layers, including tokenised deposits and regulated stablecoins. Clearing priority and liquidity management become algorithmic; the front‑end brand is almost irrelevant.
History shows that innovations with low physical friction and clear economic upside accelerate. Tokenisation took five years to mainstream; Tap‑to‑Pay took three once smartphones shipped. Agentic payments ride pure software rails, so adoption curves will be steeper.
Agents shorten the distance between desire and transaction to one API call.
Whichever rail answers that call first writes the future margin structure of payments.
Decide now whether to build the rail, integrate the rail, or route across rails, standing still is the only losing move.
Thank You for Reading. Please like, Comment, Share, or Post on Your Social media. I appreciate all the feedback I can get.
P.S. If you want to work with me in a larger capacity, either speaking, advisory, or consulting, feel free to email or DM me.