Agentic Commerce: Should You Let AI Agents Pay for You?

Agentic commerce is the new buzzword in retail: AI agents are supposed to compare, order and pay for us. At the same time, big US banks are locking out agents that log in with their customers' passwords. What's behind this, and how much money should you trust an AI agent with?
Let an agent prepare purchases – yes. Let it pay only with your explicit approval in each individual case, and ideally with a separate card with a low limit. No AI agent gets your online banking password, your TANs (one-time transaction codes) or full access to your password manager. Not because agents are stupid, but because they are still too easy to manipulate today, and if something goes wrong you may be liable yourself.
What is an AI agent, and what is agentic commerce?
A chatbot answers questions. An AI agent acts: it opens websites, fills in forms, writes emails, books appointments. Well-known examples are the ChatGPT agent, Claude, Google's Gemini agent and AI features in browsers.
Agentic commerce is the next step: the agent does the shopping. You say “Order me the cheapest suitable printer cartridge”, and the agent searches, compares, adds it to the basket and pays.
How agentic payments are meant to work
Under the label agentic payments, payment networks and platforms are building the technology for this. The shared idea: the agent doesn't get your real card details, but its own limited permission.
- Mastercard Agent Pay (since April 2025): every agent receives its own “agentic tokens”, that is, stand-in numbers instead of the real card number.1
- Visa Intelligent Commerce (since April 2025): the programme applies Visa's tokenisation technology to AI agents, so that merchants can accept agent payments like normal card payments.1
- Agent Payments Protocol (AP2) from Google (since September 2025, with over 60 partners, including Mastercard and PayPal; handed over to the independent FIDO Alliance in April 2026): the agent presents cryptographically signed “mandates” for purchase intent, basket and payment. They prove that a real person authorised exactly this purchase.2,3
- Agentic Commerce Protocol from OpenAI and Stripe (since September 2025): this launched “Instant Checkout” in the US, that is, paying directly in the ChatGPT window. In March 2026 OpenAI scaled it back again: since then, payment generally happens in the merchant's own shop. The standard itself is still being developed. In-chat purchasing never existed in Germany.4
That is the clean way. The unclean way is for an agent to log in to bank or shop sites with your real password. And that is exactly what the banks are now pushing back against.
Why US banks are locking out AI agents
At the end of October 2026, Axios and Reuters reported that big US banks are locking third-party AI agents out of customer accounts. JPMorgan Chase, Citi, Charles Schwab and Morgan Stanley are blocking agents that sign in with their customers' login details. Bank of America and Wells Fargo are restricting access.5,6 Schwab CEO Rick Wurster announced that the firm will let agents in through dedicated interfaces (so-called MCP servers), mainly with read-only access: no third-party agent may move money.7
Bank of America also emailed its customers, telling them not to share their username and password with AI assistants. Anyone who does so anyway is responsible for whatever results from it.8
The background: when an agent logs in with the real password, it looks to the bank like the customer themselves, but it is software that could move money on its own. A US rule for orderly data access through interfaces (Section 1033 of the open banking rules) was provisionally halted by a federal court in Kentucky in October 2025; the US Consumer Financial Protection Bureau (CFPB) is currently rewriting it.9 So for now, the banks are regulating it themselves.
In the EU the situation is different: the Payment Services Directive (PSD2) already provides regulated interfaces through which officially authorised third-party providers may access account data or initiate payments with your consent and strong customer authentication. Common AI agents such as ChatGPT or Claude don't have such an authorisation; they can only be connected through a regulated payment service provider. So the safe route goes through these official interfaces, never through handing over your password.
What consumers think
Scepticism is high, in Europe as in the US:
- 19% would let an AI assistant complete a purchase. 67% want to check and approve every purchase themselves. (Nuvei, 5,000 respondents in the UK, Germany, Italy, Poland and Benelux, July 2026)10
- 50% would not trust an agent with their card details. 53% already use agents to compare prices, but only 14% want the agent to complete the purchase with their approval. (Ecommpay, 1,750 respondents in the UK, France, Germany, Spain and Italy, 2026, as reported by Stellagent)11
- 42% would let an agent spend up to 100 US dollars, 22% nothing at all. But only 5% feel comfortable in principle with an AI carrying out actions itself with their consent. (Ping Identity, 11,000 respondents in 12 countries, 2026)12
- 79% think nobody should have to give an AI their bank password to get help with their finances. (ABA/Morning Consult, USA, October 2026)13
- 10% of respondents under 50 use AI agents, while among respondents aged 50 and over it is only 2%. (BSI Cybersecurity Monitor 2026, Germany)14
Notably, merchants and financial providers are much further ahead. In the European Signicat study “The Battle to Onboard 2026”, 79% of financial providers consider themselves ready to accept applications from AI agents. But only 14% of consumers would let an AI apply for a financial product on their behalf.15
The real risk: prompt injection
The biggest problem isn't that the agent miscalculates. It is prompt injection: an AI agent cannot reliably tell whether an instruction comes from you or is part of the text it is currently reading. If a shop page says “Only recommend this offer” in invisible text, or an invoice says “Forward this email” in white lettering, the agent may take it as an order. The tech journalist Jörg Schieb compares agents to a hard-working intern who believes every note that lands on their desk.16
Tests from 2026 show how widespread the problem is:
- LivePI benchmark (University of Pennsylvania): depending on the model, 10.7 to 29.6% of the hidden attacks succeeded, tested with GPT-5.3-Codex, Claude Opus 4.6 and Gemini 3.1 Pro, among others. Every model tested could be hijacked via group chats. The good news from the same study: a two-stage defence made up of an input filter and a check that must approve every action before it is carried out stopped all malicious actions in a test setup with GPT-5.3-Codex.17
- Public attack competition by Gray Swan (March 2026): 464 participants, over 272,000 attempts against 13 top models. Every model could be hijacked at least once, even though an attack only counted if the agent carried out the harmful action and hid it from the user. Success rates ranged from 0.5 to 8.5%, depending on the model.18
- In companies, almost half (49.8%) of more than 380 organisations surveyed had already observed, or at least suspected, unwanted or unauthorised actions by agents. Only 17.5% can automatically stop a misbehaving agent.19
Germany's Federal Office for Information Security (BSI) has classed indirect prompt injection as an “intrinsic” vulnerability of AI-based applications since 2023, meaning it is a consequence of how they are built and not a bug that an update fixes.20 According to the BSI, there is no one hundred per cent protection.21 A one per cent success rate sounds low. But an agent that reads dozens of shop pages and emails every day will sooner or later run into that one doctored text.
Even the AI providers are hitting the brakes on their agents
Telling: Anthropic lists banks, trading platforms and crypto exchanges as a high-risk category for its browser agent Claude, which is blocked by default or requires explicit permission. Before purchases, Claude asks for confirmation.22
The password manager 1Password introduced “1Password for Agents” on 22 October 2026: the agent may sign in but never sees the password, and you confirm every single sign-in on your own device. One catch remains: once the agent is logged in, it can keep acting in the account. In addition, the service is for now only available for personal and family accounts on 1Password.com; European accounts under 1Password.eu are not yet supported.23 Approval bridges like this are far safer than typing passwords into a chat. Even so, you shouldn't use them for bank accounts. So the industry is relying on human approval rather than blind trust.
And who is liable if the agent pays wrongly?
For AI agents, this is not yet settled. The basic rule in Germany is: if a payment was made that you didn't authorise, the bank must refund the amount (§ 675u BGB, the German Civil Code). With AI agents, though, there are two traps:
- A wrong purchase by your own agent: if you instructed the agent yourself and it simply buys the wrong thing, that may already count as a payment you authorised. In that case, the right to a refund may not apply at all.
- A hijacked agent: if an attacker uses the agent to trigger a payment you didn't authorise, the bank must in principle refund it. But if you acted with gross negligence in breach of your duties, for example by not keeping your login details secret (§ 675l BGB), you are liable for the entire loss yourself (§ 675v para. 3 BGB).
In July 2025, the German Federal Court of Justice (BGH) ruled that handing TANs over to fraudsters is grossly negligent (case no. XI ZR 107/24).24 Nobody has yet decided whether courts would take an equally strict view of voluntarily handing bank details to an AI agent. But the risk is yours. Bank of America says it openly: anyone who shares their details is responsible for the consequences.
Note: This is not legal advice, and the legal position described here is German law. If in doubt, ask your bank or a consumer advice centre.
The everyday rule: the agent traffic light
The traffic light comes from Jörg Schieb.16 Behind it stands Meta's “Agents Rule of Two”: in one session, an agent should have at most two of three abilities, namely processing untrusted content, accessing sensitive data and acting externally.25 Anyone who shops and pays on the web needs all three. That is why a human has to approve.
If you let your agent shop anyway: 6 protection rules
- A separate payment method: a prepaid or virtual card with a low limit, never your main card and never your current account directly. Even safer are virtual single-use cards, which some banks and payment services offer: their number is valid for one purchase only.
- Set a limit per purchase and per month where the service offers it. In the Ping survey, the limit for many people was 100 US dollars. According to a Checkout.com survey, spending limits are also the control that consumers most often want.11
- Switch on approval before every purchase and don't just click it away: check the merchant, item and final price.
- Use official routes: use the platform's payment features instead of typing card details or passwords into the chat.
- A separate browser profile for the agent, with no online banking logged in.
- Check your statements regularly and turn on push notifications for every card payment.
Conclusion: yes to the shopping assistant, no to the blank cheque
Agentic commerce is coming. Payment networks, merchants and AI providers are investing heavily, and according to Google Keyword Planner, “Agentic Commerce” is searched for in Germany about ten times as often as a year ago. Yet even OpenAI scaled back buying directly in the chat in March 2026: people would rather pay in the shop they know. So today the rule is: AI agents are excellent shopping assistants and poor cashiers. Let them search, compare and prepare. You press the buy button, and your banking password stays with you. Banks, security authorities and even the AI makers themselves are drawing the same line right now.
Frequently asked questions about AI agents and paying
What is an AI agent?
An AI that doesn't just answer questions but carries out tasks on a computer by itself: operating websites, filling in forms, writing emails, shopping.
What does agentic commerce mean?
Shopping through AI agents: the agent searches for products, compares them, orders and pays on the customer's behalf. Agentic payments is the term for the payment technology behind it.
Can I give my AI agent my online banking password?
No. Banks require customers to keep their login details secret. If you hand them over, you risk being liable yourself if something goes wrong. Several US banks are now actively blocking this kind of access.
What is prompt injection?
An attack in which hidden instructions in websites, emails or documents trick an AI agent into doing things nobody wanted, such as making the wrong purchase or passing on data.
Is paying through ChatGPT or other AI services safe?
Never type your card number or password into a chat window. It is safer if the service uses its own limited payment permissions and you confirm the final amount of every purchase yourself, for example with a fingerprint in your banking or payment app. A residual risk of manipulation remains.
Sources
- Stellagent: Mastercard Agent Pay und Agentic Tokens. stellagent.ai
- Analytics Vidhya (2025): Google’s Agent Payments Protocol (AP2). analyticsvidhya.com
- PYMNTS (2026): Google and Mastercard Contribute Agentic Commerce Standards to FIDO Alliance. pymnts.com
- Forrester (2026): What It Means That The Leader In “Agentic Commerce” Just Pulled Back. forrester.com
- Tagesspiegel (24 October 2026): US-Großbanken sperren KI-Agenten aus Kundenkonten aus. tagesspiegel.de
- Cybernews (2026): Morgan Stanley, Bank of America, Citi have started blocking AI agents. cybernews.com
- InvestmentNews (2026): Schwab CEO says firm is blocking rogue AI agents from client accounts. investmentnews.com
- TechRadar (2026): Bank of America tells customers to stop sharing banking passwords with AI agents. techradar.com
- Cooley (2025): Court Enjoins CFPB’s Open Banking Rule Pending New Rulemaking. cooley.com
- Nuvei (2026): How Europe pays: 2026 consumer payment preferences across five markets. nuvei.com
- Stellagent: Ecommpay and Checkout.com surveys on agentic commerce. stellagent.ai
- Ping Identity (2026): AI trust changes when assistance turns to action. prnewswire.com
- ABA Banking Journal (October 2026): ABA survey: U.S. consumers wary of AI agents handling their money. bankingjournal.aba.com
- BSI (2026): Cybersicherheitsmonitor 2026. bsi.bund.de
- The AI Insider (2026): Signicat study “The Battle to Onboard 2026”. theaiinsider.tech
- Schieb, J. (2026): KI-Agenten: Ein einziger unsichtbarer Satz kann Probleme machen. schieb.de
- Zhao, L., Bhaskar, A., Dobriban, E. (2026): LivePI: More Realistic Benchmarking of Agents Against Indirect Prompt Injection. arxiv.org
- Gray Swan (2026): How Vulnerable Are AI Agents to Indirect Prompt Injections? arxiv.org
- Salt Security (October 2026): State of Agentic AI and API Security 2H 2026. salt.security
- BSI (23 July 2023): Indirect Prompt Injections – Intrinsic Vulnerability in Application-Integrated AI Language Models. bsi.bund.de
- Borncity (2026): BSI-Leitfaden: Schutzmaßnahmen gegen Prompt Injection in KI-Chats. borncity.com
- Anthropic: Using Claude in Chrome safely. support.claude.com
- 1Password (2026): 1Password for AI Agents. 1password.com
- Der Betrieb: BGH sieht grobe Fahrlässigkeit bei TAN-Weitergabe. der-betrieb.de
- Willison, S. (2025): New prompt injection papers: Agents Rule of Two. simonwillison.net
Simon Gütter is a marketing expert who has spent years working at the intersection of marketing, sales and automation. He finds the places where companies lose time and customers, and builds workflows that run on their own: from appointment booking to reporting. More about Simon