Agentic Commerce Explained: How AI Agents Will Pay, and What Merchants, Banks and Platforms Need to Know
In short
- Agentic commerce means AI agents that search, choose and pay on your behalf, within limits you set.
- Mastercard, Visa and Stripe have all announced tools for it. Volumes are still early.
- The hard problems are not the payment itself but identity, permission, disputes and who is liable.
- Merchants, banks and platforms can prepare now without rebuilding anything.
- Start with small, capped tests and clear rules for who is responsible when an agent gets it wrong.
Imagine telling an assistant, “reorder the printer toner, but only if it is under $80 and it arrives by Friday,” and the assistant simply does it. It finds the supplier, compares the price, checks out and pays. You see a receipt afterwards.
That is the idea behind agentic commerce, and the world’s biggest payment companies are building the rails for it. This guide explains what it means in plain English, how an agent-led payment works, who is building what, what could go wrong and what merchants, banks and platform operators should do now.
What agentic commerce actually means
“Agentic” simply means software that can take actions toward a goal, not just answer questions. Applied to payments, it comes in three flavours:
- Shopping on your behalf. An agent searches, compares and checks out for a person, within rules the person sets.
- Business purchasing. Agents that handle repeat buying such as supplies, subscriptions and travel for a company.
- Machine-to-machine payments. Agents paying other software for data, computing power or services, often in tiny amounts and many times a day.
How an agent-led payment works
Each network does this differently, so treat the steps below as a simplified picture.
- The person sets the rules. What the agent may buy, a spending limit and perhaps a list of approved merchants.
- The agent receives a credential. Usually a token that stands in for the real card or account number.
- The agent finds the item and starts checkout with the merchant.
- The merchant and the payment network check that the agent is genuine and is acting within its permission.
- The payment is authorised and settled like any other, over cards, bank accounts or, in some designs, stablecoins.
- The person gets a record, and a route to challenge a purchase that went wrong.
Who is building what
| Company | What it has announced | What it is for |
|---|---|---|
| Mastercard | Agent Pay, launched in 2025, and Agent Pay for Machines, launched on 10 June 2026. The latter is built around credentialing, permissioning, transacting and settlement, and launched with more than 30 partners including Adyen, Stripe, Coinbase and Cloudflare | Trusted agents paying merchants, and automated small-value payments between machines |
| Visa | Visa Intelligent Commerce, tools that let developers connect AI agents to Visa’s network, and a collaboration with OpenAI announced in June 2026 | Agent-initiated payments using tokenised credentials |
| Stripe | An Agentic Commerce Suite and an Agentic Commerce Protocol developed with OpenAI | Lets businesses sell through AI agents and accept agent-led checkout |
A word of caution. These are announcements by the companies named, and we have not tested the products. “Announced” does not mean “widely used”: volumes are early, details change quickly and most purchases are still made by people.
What could go wrong
- Identity. How does a merchant know the agent really acts for this customer, and is not an impostor?
- Permission. A spending limit is only as good as its enforcement. A mistaken instruction can repeat many times before anyone notices.
- Manipulation. Agents can be steered by misleading pages or messages into buying the wrong thing.
- Disputes and liability. If an agent buys the wrong item, does the loss fall on the customer, the merchant, the bank or the agent’s provider? Network rules and legal answers are still developing.
- Fraud. Criminals will imitate agents. Our guide to fraud controls for investment and lending platforms covers the basics that still apply.
- Compliance. Agent payments over borders or in stablecoins bring the usual identity checks and anti-money-laundering rules, which differ by country. Our guides to stablecoins for businesses and cross-border payments explain the landscape.
What it means for you
If you are a merchant
- Make your product data clear and machine-readable: price, stock, delivery times and return terms.
- Decide whether you will accept agent checkouts, and how you will tell trusted agents from bad actors.
- Update refund and dispute policies to cover purchases made by an agent.
- Ask your payment provider what it supports today.
If you are a bank or card issuer
- Decide how you will authorise token-based, agent-initiated payments.
- Give customers simple controls to set limits, pause an agent and review its activity.
- Prepare your dispute process for purchases the customer says they did not intend.
If you run a lending, investing or payments platform
- Keep a human step for high-stakes actions such as moving money or accepting terms.
- Keep a clear audit trail of what any agent did, and when.
- If you use AI inside your own product, our guide to AI in credit decisions explains the fairness and explainability expectations that already apply.
- Remember the lesson from embedded finance: responsibility has to be clear before the money moves.
Six questions to ask your payment provider
- Do you support agent-initiated payments today, and through which networks?
- How do you verify an agent’s identity and its permission to spend?
- What happens to disputes and chargebacks when an agent made the purchase?
- What spending limits and approval steps can we set?
- How do you screen for fraud in agent traffic compared with human shoppers?
- What will it cost, and what do we need to change in our checkout?
Use these alongside the hard questions in our guide to running a fair software demo, and ask for the answers in writing.
A simple example: where the checks sit
Take a company whose agent reorders office supplies. This is an illustration, not a description of any one product.
| Step | Check | Who is responsible |
|---|---|---|
| The agent is given authority | Spending limit, approved merchants and an expiry date | The company, set up with its payment provider |
| The agent reaches checkout | Is this a genuine agent acting for this company? | The merchant and the payment network |
| The payment is authorised | Is it within the limit, and is it unusual for this account? | The bank or card issuer |
| The wrong goods arrive | Who can reverse the payment, and how? | Depends on the agreements and network rules, so agree it in advance |
| Something looks suspicious | Can the company pause the agent at once? | The company and its provider |
Notice the pattern: every row needs a named owner. Most of the risk in agent payments sits in the gaps between the rows.
Common mistakes
- Treating it as a technology project when it is mainly a policy and risk question.
- Setting spending limits too high on day one.
- Having no way to pause or revoke an agent quickly.
- Assuming existing dispute rules cover agent purchases without checking.
- Forgetting that agent traffic can look like bot traffic to fraud tools, and the reverse.
- Waiting for perfect standards. Small, capped tests teach more.
A quick readiness check
| Area | Question to ask yourself | You are ready if… |
|---|---|---|
| Rules | Do we have a written policy for what an agent may and may not buy? | A one-page policy exists and someone owns it |
| Limits | Can we set and change spending limits instantly? | Limits can be changed in minutes, not days |
| Identity | Can we tell a trusted agent from an impostor? | Your provider can show you how verification works |
| Disputes | Do we know who bears the loss on a wrong purchase? | The answer is written into your agreements |
| Visibility | Can we see what every agent did, and when? | There is an audit trail you can export |
Our view
Agentic commerce is real, but it is early. The announcements show where the payment rails are heading, not how much traffic will use them next year. The sensible response is neither to ignore it nor to rebuild around it. Run small, capped experiments, settle the questions of identity and disputes before volume arrives, and keep a person in the loop for anything that matters. The companies that do the unglamorous groundwork now will be ready whichever way the market goes.
Frequently asked questions
Is agentic commerce live?
Several payment companies have launched tools and run live transactions. Volumes are still early, and most purchases are made by people.
Is it safe to let an AI agent pay for things?
That depends on the limits, identity checks and dispute rules around it. Start with low limits and an approval step for anything unusual.
Who pays if an agent buys the wrong thing?
That depends on the agreements and network rules involved, and it is still being worked out. Ask your provider before you rely on it.
Do I need a special account to use it?
Not necessarily. Check what your current payment provider supports.
This article is general information, not legal, financial or compliance advice. Requirements differ by country and product, so check them with your own advisers. See our disclaimer. Last reviewed October 30, 2026.
Published by The Funding Journal Editorial. Practical writing on fintech, crowdfunding, investment platforms, SaaS and AI, and the stories of founders who raised capital.
