Your AI agent can research suppliers, compare quotes, draft the purchase order and write the email to the vendor. Then it stops, and a person has to log into the portal, enter card details and click pay. The same thing happens with refunds, subscription renewals and invoice payments. If agents are so capable, why can they produce almost anything but still cannot complete a transaction?
The short answer: agents are not held back by intelligence but by trust infrastructure. Payments and purchases require proof that a specific person authorised a specific transaction, a way for the merchant to verify that the request is genuine, and clear liability when something goes wrong. Card networks, banks, procurement systems and fraud tools were built around humans proving those things. New protocols for agent payments are emerging to fill that gap, but for most businesses today, the right design is agents that prepare transactions and humans, or tightly bounded rules, that authorise them.
The three problems a transaction has to solve
When Google announced its Agent Payments Protocol (AP2) in September 2025, developed with more than 60 payments and technology companies, it framed the challenge in three parts:
- Authorisation. Proving the user gave the agent authority to make this particular purchase.
- Authenticity. Letting a merchant verify that the agent’s request reflects the user’s real intent.
- Accountability. Knowing who is responsible if a transaction is fraudulent or wrong.
Today’s systems answer those questions with passwords, one-time codes, card verification, device checks and a human clicking “confirm.” An agent acting on its own breaks each of those assumptions.
What the new agent payment protocols do
AP2 uses “mandates”: cryptographically signed records of what the user authorised. An intent mandate captures the original instruction and any rules, such as a price limit; a cart mandate records the exact items and price the user approved. For delegated tasks, the agent can create the cart mandate automatically only when the pre-agreed conditions are met.
The Agentic Commerce Protocol, announced by Stripe and OpenAI later in September 2025, focuses on the merchant side. Merchants remain the merchant of record and can accept or decline orders, while a shared payment token lets an AI application initiate a payment without exposing the buyer’s card details. Each token is scoped to a specific merchant and cart total.
Both approaches share a principle worth copying even if you never use them: an agent should carry proof of a narrow, specific authorisation, not a general-purpose credential.
Why businesses should not wait for the protocols
Consumer checkout inside AI assistants is one thing. Business transactions add more layers:
- Purchase approval thresholds and budget owners
- Supplier onboarding and bank detail verification
- Segregation of duties between who orders, who approves and who pays
- Audit trails for finance and tax
- Contract terms that no protocol checks for you
Giving an agent a company card or payment API key to “just handle it” skips every one of those controls. It is the financial version of giving an agent admin rights, a risk we covered in why AI agents should not get more access than your developers.
A design that works now
| Step | Agent | Human or rule |
|---|---|---|
| Find options, compare, check contract terms | Does the work | Reviews summary |
| Prepare purchase order or payment run | Drafts in the ERP or payments tool | — |
| Authorise | Cannot | Approves in the system with their own credentials |
| Low-value, pre-approved purchases | Executes within a hard cap | Rule enforced by the payment system, reviewed weekly |
| Record, reconcile, file documents | Does the work | Samples for accuracy |
A hypothetical example: a 60-person company lets an agent renew approved software subscriptions under $200 using a virtual card with a merchant lock and monthly limit, while anything else becomes a drafted request that the budget owner approves in the finance system. The agent saves hours; the controls stay intact.
What to watch
- Whether your payment provider, card issuer or ERP supports agent-initiated transactions with scoped tokens.
- How liability for disputed agent purchases is allocated in provider terms.
- Whether your auditors accept agent-prepared, human-approved workflows as meeting segregation of duties.
Agents that do the work up to the point of trust
The fastest wins come from letting agents do everything around a transaction and keeping authorisation where it belongs. AB7 Solutions builds AI agents and business process automation that connect to your ERP, CRM, procurement and payment tools through APIs, with scoped credentials, spending caps, approval workflows and audit logs designed in. We can also help you evaluate emerging agent payment options when they fit your stack. If a simple approval workflow would deliver most of the value, we will recommend that first.
Tell us which transactions your team handles repeatedly, and we will show you how much of the work an agent can safely take on.
Email: ab@ab7solutions.com | director@ab7solutions.com
Phone: +91 9878067778 | +1 321 341 7733
Website: www.ab7solutions.com
Sources: Google Cloud, Announcing Agent Payments Protocol (AP2), September 2025; Stripe, Agentic Commerce Protocol co-developed with OpenAI, September 2025.