TL;DR
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
AI agents are already making purchases on behalf of people, although the volume remains small, FIS executive Troy Mann told PYMNTS. Banks lack a universal way to verify which agent initiated a payment and whether it followed the buyer’s instructions, leaving permission tracking and data sharing as open challenges.
AI agents are already shopping and making purchases on behalf of people, but banks lack a universal way to verify which agent initiated a transaction and what the customer authorized it to do, FIS executive Troy Mann told PYMNTS. The gap could complicate fraud checks and payment approvals as consumers use stored credentials in systems being adapted for agent-driven purchases.
Mann, FIS Global’s senior vice president of product management, described agentic commerce as active but still small in volume. In many cases, a person first sets the intent and constraints, such as a price ceiling or shipping requirements. An agent then searches or acts within those limits. Payment systems, however, were largely built to identify accounts, credentials and transactions, rather than the software that makes a purchasing decision.
That leaves banks with questions beyond whether a payment credential is legitimate. They may need to establish which agent used it, whether that agent had permission to make the purchase and whether the transaction followed the buyer’s original instructions. Mann said FIS is working with financial institutions to identify agentic transactions, apply different fraud scoring and reporting, and incorporate signals across the payment life cycle.
The relevant information is spread among parties. A merchant may know what a customer requested, an AI platform may hold the instructions given to the agent, and a bank may see the account and payment history. Those signals do not necessarily travel together. Mann described tokenization as an important security layer, while stressing that tokenized credentials alone do not establish the agent’s authority.
Why Agent Permissions Matter
If a bank can verify a card but cannot tell whether the software using it was authorized for a particular purchase, a familiar credential may provide an incomplete picture of a transaction. That makes permission and intent relevant to fraud reviews, not just the identity of the account and the validity of the payment details.
The issue also affects the competing interests around automated shopping. Merchants may want agents to turn customer instructions into completed purchases; banks want to protect accounts and retain a role in payment decisions; customers want convenience without giving up control of their money. A system that connects authentication, permission records and transaction data could help address those needs. The PYMNTS report does not describe a universal system already in place.
For now, safeguards may add steps to checkout. Mann said a degree of friction could be useful while the industry develops ways to make an agent’s authority visible. That is a shift for payments systems designed to make checkout faster, but the report presents it as a possible near-term trade-off rather than a settled industry policy.
As an affiliate, we earn on qualifying purchases.
From Stored Cards to Agent Mandates
Many consumers have already stored payment credentials in digital wallets or online shopping accounts. In an agentic purchase, the agent may use those credentials after receiving instructions from the person. The payment still draws on a customer’s account, but another decision-maker is involved between the customer’s request and the transaction.
Existing payment controls offer some building blocks. Mann pointed to tokenization and raised the possibility of virtual cards designated for agentic commerce. A business could set limits by amount, merchant or time window so the credential’s permitted use reflects the authority granted to an agent. The report also mentions dynamic security credentials as a possible approach. It does not say these measures have been adopted across the market.
Mann recommended a “crawl, walk, run” approach, with people involved in establishing intent and authorization as automation develops. He told PYMNTS that “intent has to be expressed and captured.” The report frames auditability and data sharing as part of the challenge: payment participants need enough information to assess what an agent was allowed to do.
““Intent has to be expressed and captured.””
— Troy Mann, speaking to PYMNTS
payment permission management systems
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The report does not identify a shared standard for proving an agent’s identity and authority across merchants, AI platforms, payment networks and banks. It also does not specify how those parties would exchange or retain a buyer’s instructions, or who would be responsible when the transaction record is incomplete.
Current transaction volumes are described only as small; the report gives no count, time window or comparison baseline. It also does not provide adoption figures for agent-specific virtual cards or dynamic credentials, or name a timetable for broader deployment. Mann’s comments describe work and possible approaches, not a completed industry-wide solution.
transaction authorization software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Building Controls Alongside Automation
FIS is working with financial institutions on transaction identification, fraud scoring and reporting, according to Mann. The report points to more information moving among merchants, agents, networks and issuers as a requirement for better assessment, but does not give a schedule or name a common protocol that will make that exchange possible.
In the near term, businesses and payment providers may consider restricted virtual cards or other controls that limit where, when or how much an agent can spend. Mann’s suggested gradual approach keeps people involved in setting intent and authorization as the systems develop. Whether those controls become common, and how banks will confirm that a purchase matches a customer’s instructions, remain open questions.
As an affiliate, we earn on qualifying purchases.
Key Questions
Are AI agents already making purchases?
Yes. FIS executive Troy Mann told PYMNTS that agents can shop, negotiate and transact for people, while describing current volumes as small. The report does not give a transaction count or time period.
Why might a valid payment credential not be enough?
A valid credential can show which account is being charged, but it may not show which agent used it or whether that agent was allowed to make the specific purchase.
What safeguards are being considered?
Mann discussed tokenization and possible virtual cards with limits on amount, merchant or time window. The report also refers to dynamic security credentials. It does not say these options are universal or widely deployed.
What does the industry still need to resolve?
Payment participants need ways to connect the customer’s instructions with the agent, credential and resulting transaction. The report says those signals are split among merchants, AI platforms and banks, and does not describe an industry-wide method for sharing them.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
