In this guide · 7 sections
Short answer
Treat an AI agent with payment access like a new employee with a company card: give it a separate virtual card with a low limit, require approval for every payment, never give it saved card details or banking logins, and keep logs. As a seller, keep fraud screening on and watch for automated card testing.
What 'agents that pay' means in 2026
An AI agent is an AI assistant that can take actions, not just write text: open websites, fill forms, call tools and, increasingly, make payments. Vendors have started building payment rails for them.
| Development | Date | What it is | Availability |
|---|---|---|---|
| Stripe Link wallets for agents | 29 Apr 2026 | People let agents pay with Link; a one-time-use card is issued per task and each payment is approved | Agent payments for US and Canadian customers |
| Stripe shared payment tokens | 2026 (preview) | Agents pass sellers a scoped token with amount and expiry limits instead of card details | US, Canada and select European countries |
| DBS and Visa Intelligent Commerce pilot | 16 Feb 2026 | Agent-initiated F&B payments with DBS/POSB cards via issuer-controlled flows | Pilot in Singapore |
Stripe's Sessions 2026 announcement describes the Link wallet design: real payment details are never exposed to the agent, a one-time-use card is issued per task, and the person approves each payment. Its Link Agent Wallet docs list agent payments as available to US and Canadian customers. Visa's DBS pilot release says DBS and Visa demonstrated real food and beverage transactions completed by agents.
So for most Singapore businesses, purpose-built agent wallets aren't here yet. The near-term risk is simpler: general AI agents and browser assistants using a staff member's saved cards, logins or password manager.
The risks: wrong purchases, manipulation, and overreach
OWASP's LLM06:2025 Excessive Agency describes the core risk: an AI system with tools can perform damaging actions because of hallucination, a badly worded request, or prompt injection hidden in content it reads. OWASP names three root causes: excessive functionality, excessive permissions and excessive autonomy.
- Misunderstanding: the agent buys the wrong item, quantity or date, or pays a deposit you didn't intend.
- Prompt injection: a web page or email contains hidden instructions that steer the agent to a different seller or to reveal data. Our guide to AI browser agents and security explains how this works.
- Scam sites: an agent searching for 'cheapest supplier' can land on a fake shop as easily as a person can.
- Overreach: an agent with access to online banking, a payroll system or a saved company card can do far more than the task needs.
- Disputes and refunds: proving who authorised a purchase is harder when an agent clicked 'Pay'.
Controls if your staff use agents that can pay
OWASP's mitigations translate directly into payment controls: limit what the agent can do, limit what it can reach, and require a human to approve high-impact actions.
Use a separate virtual card with a low limit
Ask your bank or card provider for a virtual card used only for agent or online tool purchases, with a monthly limit you'd be comfortable losing and merchant-category restrictions if available.
Require approval for every payment
Use tools where the person confirms each charge, as Stripe's Link wallet design does. Turn off any 'auto-approve' or 'remember my choice' options.
Keep saved cards and bank logins out of reach
Remove saved company cards from browsers and assistant accounts. Exclude banking and finance sites from password-manager auto-fill for the agent.
Limit which sites the agent may act on
Where the tool allows it (for example Chrome's auto browse URL allow and block lists), let the agent act only on known suppliers.
Log and review
Keep the agent's activity history, card notifications and receipts. Review agent purchases weekly at first.
Write it into your expense policy
Who may let an agent spend, on what, up to what amount, and who approves.
If you use Gemini in Chrome or Copilot in Edge, our guide to managing browser AI at work shows how to keep Chrome's auto browse off or restricted.
If you sell online: agents as customers
Agent shoppers will mostly look like normal checkouts, but automation also helps fraudsters. Card testing, where bots try stolen card numbers with small payments, is a known problem for small online stores, and our guide to fixing a card testing attack covers it.
Seller checklist for agent-era checkouts
Stripe's Radar documentation explains its fraud screening, and Stripe says it expanded Radar at Sessions 2026 to evaluate sign-ups and usage for AI services. If you want to be visible to AI shopping agents, our guide to selling through AI shopping agents with Shopify covers which programmes apply in Singapore.
A simple policy for agent payments
One paragraph in your IT or expense policy is enough for most small firms.
Allowed
- Agents researching suppliers and preparing a cart for a person to check
- Payments from the dedicated virtual card, approved by a person each time
- Recurring purchases from approved suppliers under a set limit
Not allowed
- Agents using online banking, PayNow transfers or payroll
- Agents using staff members' personal or saved company cards
- Auto-approved payments of any amount
- Agents signing contracts or accepting terms on the company's behalf
Do it yourself or get help?
The controls above are mostly settings and habits: a separate virtual card, approvals on, saved cards removed, banking sites excluded, and a short policy. Staff awareness is the harder part, because agents make it easy to click 'yes' without reading.
Our Security Awareness Training (S$390) sets up a 12-month plan of five-minute lessons in your Microsoft, Google or dedicated training platform, with Singapore-specific scam topics and completion tracking, so staff learn to spot fake shops and payment scams that agents can also fall for. If you sell online, our Online Payment Setup (S$290) sets up Stripe or HitPay in your name with PayNow and cards, checkout on your site and test transactions.
Frequently asked questions
Partly. As of October 2026, Stripe's Link agent payments were available to US and Canadian customers, and its shared payment tokens to the US, Canada and some European countries. In Singapore, DBS and Visa announced a pilot of agent-initiated card payments in February 2026. General AI agents can still use any card or login you give them, so controls matter now.
Giving it more access than the task needs. OWASP calls this excessive agency: an agent with tools can take damaging actions because of a mistake, a vague request or hidden instructions on a web page. An agent with your saved company card or banking login can spend or transfer far more than you intended.
Use a separate virtual card with a low monthly limit and merchant restrictions if your bank offers them, require approval for every payment, remove saved company cards from browsers and assistant accounts, and exclude banking sites from auto-fill. Review card notifications and the agent's activity history weekly until you're confident in how it behaves.
Automation makes some fraud easier, especially card testing with stolen card numbers. Keep your gateway's fraud screening on, add rate limits or CAPTCHA to checkout and sign-up, use 3D Secure for higher-risk orders, and send clear order confirmations so real customers can spot purchases they didn't intend. Review fraud rules after any spike in failed payments.
Not with company cards stored in the browser. Remove saved cards, keep agent features like Chrome's auto browse off or limited to approved sites, and require staff to enter or approve payments themselves. If you do allow agent purchases, use a dedicated low-limit virtual card and a written rule on what may be bought.