Before You Let an AI Agent Shop for You, Read the Fine Print
Before You Let an AI Agent Shop for You, Read the Fine Print

Banks, United Airlines, and Amazon are telling personal AI agents “not yet.” Here’s what that means for your wallet, your accounts, and your business, with an action plan for executives, CIOs, CISOs, and consumers.
By Theresa Payton, CEO of Fortalice Solutions and former White House Chief Information Officer | November 12, 2026
Personal AI agents are genuinely exciting. They can book, buy, pay, and answer email while we get our time back. But over the past two months, something happened that most people missed. Some of the biggest banks in three countries, a major airline, and leading online retailers told these agents “not yet.”
At the same time, the companies that make the agents wrote their terms so that if something goes wrong, you are the one responsible.
None of this means agents are bad. It means the smartest move right now is to know who carries the risk before you flip the switch.
Key Takeaways
20 well-known companies in the U.S., Canada, and the U.K. have blocked or restricted personal AI agents from customer accounts, including JPMorgan Chase, Bank of America, United Airlines, and Amazon.
4 out of 4 leading agent platforms (Meta Muse, OpenAI dots, xAI’s Grok Bot, and Instinct) say in their terms that the user is responsible for what the agent does.
$100 to $250 is the typical cap on what the agent maker will pay you if something goes wrong.
Not a permanent ban. Businesses are building controlled “front doors” for trusted agents. The direction is safer access, not no access.
First, What Is a Personal AI Agent?
A chatbot answers your questions. An agent takes action for you. Think of it as the difference between asking a friend for directions and handing them your car keys. An agent can open websites, log in, fill out forms, click “buy,” and send messages on your behalf, often while you’re busy doing something else.
The four leading personal agents all launched within weeks of each other in 2026: xAI’s Grok Bot on August 11, Instinct in an invite-only beta in August, Meta Muse on September 8, and OpenAI dots on September 29.
Who Has Said “Not Yet” to AI Agents
Airlines
• United Airlines (November 11, 2026). United updated its Contract of Carriage (Rule 6.J.6) to prohibit using AI agents to book, change, or service tickets. United cited the risk of exposing personal information when customers share their login, and the chance that automated actions could conflict with FAA and DOT rules. Agents can still show you United’s flights and fares. They just can’t complete the purchase.
U.S. banks and brokerages
• JPMorgan Chase, Citi, Bank of America, Wells Fargo, and Morgan Stanley are all blocking outside AI agents from acting in customer accounts, according to American Banker’s October 27 reporting. JPMorgan CEO Jamie Dimon called it too risky for clients and the bank. Bank of America has told customers not to share banking passwords with agents.
• Charles Schwab is turning away 10 to 20 clients a week who try to use AI agents to move money, according to CEO Rick Wurster.
Canadian and U.K. banks
• Canada: RBC, TD, BMO, and CIBC block agents such as Meta Muse from online banking. Scotiabank limits access, and National Bank is monitoring.
• United Kingdom: HSBC, Barclays, Lloyds, NatWest, and Nationwide have moved to block or restrict agents. These are described as preventive steps, with no signs yet of agents in accounts.
Retailers and marketplaces
• Amazon blocked Meta Muse from browsing and buying on its site in September, saying it was given no notice or choice.
• eBay banned unauthorized “buy-for-me” agents that place orders without human review.
• Yelp allows non-human traffic only when the agent has paid for access.
Why They’re Doing It: The Password Problem
Most of these companies are objecting to one specific thing: an agent logging in with your username and password. Once it does, the website can’t tell the agent apart from you. Anything the agent does looks like something you did.
Having led security and fraud prevention in banking earlier in my career, I can tell you the first question in any disputed payment is always the same: who authorized it? When your agent acts with your password, the honest answer is you. Jamie Dimon described the current setup as a “liability transfer,” meaning the risk quietly shifts onto the customer. Which brings us to the fine print.
What the Fine Print Actually Says
Every leading agent platform reaches the same conclusion: you are responsible for what your agent does. The details are where they differ, and the details are what matter.
Meta Muse. Meta’s terms say you are “solely responsible” for the actions Muse takes, that you accept legal responsibility as if you took those actions yourself, and that you bear all risk of financial loss. Meta’s liability is capped at the greater of $250 or what you paid in the prior 12 months, and that cap explicitly covers unauthorized actions.
OpenAI dots. OpenAI’s Service Terms say you are responsible for the actions agents take and the content they share on your behalf, and you’re responsible for any purchases. The clause does not distinguish between actions you approved and actions the agent took on its own. On the product side, OpenAI hands password changes and money transfers back to you rather than letting the agent finish them.
xAI Grok Bot. Grok Bot’s terms say, in capital letters, that the customer is solely responsible for all agentic actions, whether or not they were intended or authorized. xAI’s liability is capped at the greater of $100 or what you paid. The terms also note that bots on one account share the same environment and logins, so they aren’t separate security walls.
Instinct. Instinct’s terms say commitments the agent makes are binding on you as if you made them, and the company is not responsible for unintended actions. The cap is the greater of $100 or six months of fees. Business users take note: the terms allow personal use only, and model training on your data is on by default.
Other Considerations:
• You can get squeezed from both sides. The agent maker says you’re responsible. Some banks say the same thing from the other direction. Wells Fargo’s agreement treats an agent’s transactions as authorized by the customer, and in Canada, transactions made with credentials you shared with an agent may not count as unauthorized. If an agent makes a mistake or gets tricked, you could be left holding the loss either way.
• A “hacked” agent usually isn’t hacked the way you’d think. The bigger risk is prompt injection: instructions hidden in a web page or email that redirect your agent, just like Dana’s lookalike invoice. Meta’s own security documentation acknowledges this problem is not solved. If you approve something your agent was tricked into suggesting, that’s on you under current terms.
• The contract beats the launch post. Elon Musk publicly encouraged people to let Grok shop for them. xAI’s written terms still say you’re responsible for agent actions and their costs. When marketing and contract disagree, the contract wins.
• The law is still being written. In Amazon’s case against Perplexity, a judge found early on that a user’s permission might not be enough when a website’s terms ban agents. An appeals court has since vacated that injunction, so the question remains open.
• Some of your rights expire in 30 days. Muse users in the U.S. and Canada can opt out of mandatory arbitration within 30 days by mailing a written notice to Meta. Instinct offers a 30-day opt-out by email. Most people never know these windows exist.
The Hopeful Part: A Front Door Is Being Built
Walmart already lets outside agents buy within limits the customer sets, such as $50 or $100. JPMorgan uses AI agents internally and plans customer-facing ones. Banks consistently say they support secure, customer-approved data sharing through open banking providers. Visa and Mastercard are building ways for merchants to recognize registered, trustworthy shopping agents. The goal is a front door with a guest list, not a locked building.
What Would Have Changed Dana’s Ending?
• A separate, low-limit card for the agent would have capped the loss at a number she could live with.
• Reading before approving anything that moves money would have caught the lookalike invoice.
• Five minutes with the terms would have told her, before she started, that the risk was hers.
Your Action Plan: Prepared and Resilient
Being ready for AI agents isn’t about saying no. It’s about deciding in advance who carries the risk and how you’ll catch mistakes. Here is what each leader, and each of us at home, can do now.
Business executives: questions to ask your team this quarter
1. If an employee’s AI agent commits us to a purchase or contract, who is liable, and does our insurance cover it?
2. What do our own customer terms say about AI agents? Do we block them, allow them, or allow them with limits?
3. If a customer’s agent makes a mistake on our platform, what is our refund policy, and will it hold up to public scrutiny?
4. Which vendors are using AI agents with our data, and what do their contracts say about responsibility?
5. Are we only building a lock, or also a front door for the customers who will expect to use agents?
CIOs: Suggested Next Steps
• Map where agents could touch company money and commitments. Start with corporate cards, email, procurement, expense, and travel booking.
• Route agent terms through legal before anyone accepts them on a company account. Flag “solely responsible,” “waive,” liability caps, arbitration clauses, and training defaults.
• Issue virtual cards with low limits for any approved agent use, so a mistake has a ceiling.
• Check your travel program. United now bans agent bookings, so confirm your travel tools and policies won’t trigger cancellations.
• Publish a plain-language agent policy that says which agents, if any, employees may connect to work accounts. Note that some terms, such as Instinct’s, allow personal use only.
CISOs: Suggested Next Steps
• Require a human for anything that moves money. Turn on step-up verification and approvals for payments, vendor bank changes, and password resets.
• Add prompt injection to phishing awareness. Teach employees that an email can now give instructions to their agent, not just to them.
• Run a tabletop exercise: “An employee’s agent paid a fraudulent invoice. Who notices, who calls the bank, and what evidence do we keep?”
• Watch for agent permission grants on company email, files, and messaging, and treat any agent-connected account as higher risk.
• Preserve agent activity records in incident response. Some terms warn that the agent’s own records may not always be accurate, so capture your own.
Consumers: Suggested Next Steps
• Read before you connect. Search the terms for “solely responsible,” “waive,” “limitation of liability,” and “arbitration.”
• Never hand an agent your banking password. Use read-only, bank-approved connections where they exist.
• Give your agent its own wallet. Use a separate card or virtual card with a low spending limit, and turn on purchase alerts.
• Keep approvals on, and actually read them, for anything that moves money, sends messages, or deletes data.
• Start small and keep your own records. Let the agent handle low-risk tasks first, and save confirmations in case you need to dispute something.
• Mark your calendar. If you want to opt out of arbitration, many platforms give you only 30 days.
Your Questions, Answered
Who is responsible if my AI agent makes a mistake or gets hacked?
Under today’s terms, you are. All four leading agent platforms (Meta Muse, OpenAI dots, xAI’s Grok Bot, and Instinct) place responsibility for the agent’s actions on the user. Grok Bot and Muse explicitly extend that to actions you never intended or authorized. The most these companies will typically pay if something goes wrong is $100 to $250.
Why did United Airlines ban AI agents from booking flights?
To protect customer data and stay within aviation rules. On November 11, 2026, United added AI agents to its list of prohibited ticketing practices. It cited the risk of customers sharing their login with an agent and the possibility that automated actions could conflict with FAA and DOT regulations. Agents can still show United flights and fares, but cannot complete a booking.
Which banks are blocking AI agents?
Most of the largest banks in the U.S., Canada, and the U.K. As of November 2026, that includes JPMorgan Chase, Citi, Bank of America, Wells Fargo, Morgan Stanley, and Charles Schwab in the U.S.; RBC, TD, BMO, and CIBC in Canada, with Scotiabank limiting access; and HSBC, Barclays, Lloyds, NatWest, and Nationwide in the U.K.
Can I still use an AI agent with my bank account?
Usually only in a limited, read-only way. Many banks still allow approved connections through open banking providers that can show balances or transactions. What they block is an agent logging in with your password to move money. Follow your bank’s guidance, and don’t share your banking password with any agent.
What is prompt injection, and why should I care?
It’s a way to trick your agent with hidden instructions. Someone can hide commands in a web page, email, or document that your agent reads. The agent may then follow those commands instead of yours, such as paying a fake invoice or forwarding a message. Keeping approvals on for important actions is your best everyday defense.
If my AI agent buys the wrong thing, can I get my money back?
Possibly, but don’t count on it. Contact the merchant and your card issuer quickly. Agent terms generally make you responsible for disputing errors, and some bank agreements treat purchases made through an agent you authorized as authorized by you. A separate, low-limit card for agent purchases is the simplest way to cap your exposure. This is general information, not legal advice.
What should I look for in an AI agent’s terms of service?
Four phrases, one number, and two settings. Look for “solely responsible,” “waive,” “limitation of liability,” and “arbitration.” Find the dollar cap on what the company will pay. Then check whether you can opt out of arbitration (often within 30 days) and whether your data is used for training by default.
Should my company let employees use personal AI agents for work?
Only by deliberate decision, in writing. Some agent terms, such as Instinct’s, allow personal use only. Personal agents connected to work email or files act with that employee’s full access, which most security teams can’t easily see. Decide your policy, communicate it, and offer a governed alternative where you can.
Will banks and airlines ever allow AI agents?
Very likely, through controlled channels. The trend is toward sanctioned access: spending limits like Walmart’s, read-only open banking connections, and payment network programs from Visa and Mastercard that let merchants recognize registered agents. Today’s blocks buy time to build those safer paths.
Where can my company get help assessing its AI agent risk?
A cybersecurity firm with AI and fraud expertise can help. Fortalice Solutions helps organizations assess where AI agents touch their money, data, and contracts, update policies, run tabletop exercises, and train employees to spot prompt injection and agent-driven fraud. Learn more at www.fortalicesolutions.com.
Need Help? Fortalice Solutions Is Here
At Fortalice Solutions, our promise is fortified security: making sure our clients are prepared for an attack and resilient to future threats. AI agents are the newest test of that promise, and the organizations that prepare now will be the ones customers trust.
• For businesses: we can assess your exposure to AI agents, review the agent terms your teams are accepting, update your policies, test your defenses, and run tabletop exercises so your people know exactly what to do when an agent goes wrong.
• For executives and their families: we help protect your personal accounts and digital footprint, so an agent’s mistake or a fraudster’s trick doesn’t become your problem.
Ready to fortify your organization against AI agent risk? Visit www.fortalicesolutions.com or call 877.487.8160 to talk with our team.
Learn more: https://www.fortalicesolutions.com
About the Author
Theresa Payton is the CEO of Fortalice Solutions and served as the first female White House Chief Information Officer. She began her career in banking technology, security, and fraud prevention, and today advises boards and executives on cybersecurity and AI. She is the author of “Manipulated: Inside the Cyberwar to Hijack Elections and Distort the Truth.”
This post summarizes public reporting and published terms as of November 11, 2026. It is not legal advice, and terms change often. Read the current version of each document before relying on it. The opening scenario is a composite created for illustration.
Sources
Primary sources
• Meta Muse Supplemental Terms of Service: https://muse.ai/terms
• Meta Help Center, How Muse works with your guidance and approval: https://www.meta.com/help/artificial-intelligence/1385290430137537/
• OpenAI Service Terms: https://openai.com/policies/service-terms/
• xAI Grok Bot Terms: https://x.ai/legal/grok-bot-terms
• Instinct Terms of Service: https://instinct.com/terms
News and analysis
• Bloomberg, United blocks AI agents: https://www.bloomberg.com/news/articles/2026-11-11/united-blocks-ai-agents-from-booking-flights-ahead-of-holidays
• CBS News, United bans AI agents: https://www.cbsnews.com/news/united-airlines-bans-ai-agents-booking-flights/
• Quartz, United cites data and FAA/DOT risks: https://qz.com/united-airlines-blocking-personal-ai-agents-booking-flights-111126
• PPC Land, United Contract of Carriage Rule 6.J.6: https://ppc.land/united-bans-ai-agents-from-buying-its-tickets-seats-and-upgrades/
• CNBC, JPMorgan blocks outside AI agents: https://www.cnbc.com/2026/10/13/jpmorgan-ai-agents-account-access.html
• Finextra, Dimon on liability transfer: https://www.finextra.com/newsarticle/48585/jpmorgan-chase-blocks-outside-ai-agents-from-directly-accessing-customer-accounts
• Cloud Security Alliance, Banks block personal AI agents: https://labs.cloudsecurityalliance.org/research/csa-research-note-banks-block-personal-ai-agents-20261102-cs/
• PYMNTS, RBC and U.S. bank positions: https://www.pymnts.com/news/banking/2026/rbc-blocks-third-party-personal-ai-agents-over-fraud-risks/
• The Globe and Mail, RBC blocks personal AI agents: https://www.theglobeandmail.com/business/article-banks-ai-agents-client-data-osfi/
• The Telegraph, U.K. banks: https://www.telegraph.co.uk/business/2026/11/08/banks-ban-ai-from-accounts-over-fears-of-rogue-agents/
• FourWeekMBA, Schwab and bank liability terms: https://fourweekmba.com/ai-5-big-banks-block-ai-agents-like-muse-bofa-puts-risk-on-you/
• TechCrunch, Websites blocking agents: https://techcrunch.com/2026/10/06/the-next-hurdle-for-ai-agents-getting-websites-to-let-them-in/
• CNN, Amazon blocks Muse: https://www.cnn.com/2026/09/28/tech/meta-muse-ai-agents-amazon
• eMarketer, eBay user agreement: https://www.emarketer.com/content/ebay-joins-amazon-tightening-control-over-ai-shopping-agents
• Cooley, Amazon v. Perplexity ruling: https://www.cooley.com/news/insight/2026/2026-03-17-court-finds-ai-agent-may-violate-state-federal-law-by-accessing-amazon-accounts-without-authorization
• Finextra, Muse liability and prompt injection: https://www.finextra.com/blogposting/32988/metas-muse-made-agent-payments-a-mass-market-product-the-liability-questions-have-not-caught-up
• Ethical Growth Partners, Grok Bot Terms and consumer liability cap: https://ethicalgrowthpartners.com/blog/elon-musk-grok-credit-card-fine-print
• CellCog, Musk’s purchase pledge vs. xAI terms: https://cellcog.ai/blog/grok-bot-purchase-pledge/
• eesel, Instinct terms summary: https://www.eesel.ai/blog/instinct-ai
• Maarket, Personal agent launch dates: https://maarket.ai/blogs/best-personal-ai-agents
Fortalice Solutions
• Fortalice Solutions: https://www.fortalicesolutions.com
Comments