AI Agents

Personal AI agents: block them or let them in?

United restricted personal AI agents, Shopify opened every store to them and OpenAI gave them a wallet. Here's how to decide what agents may do on your site.

S

SelfAgentic Team

7 min read
A person holds a bank card over a laptop keyboard while shopping online.

In one week, three companies gave three different answers to the same question: what should a business do when customers send personal AI agents to act for them? On November 11, United Airlines announced it would restrict unauthorized third-party AI agents from its booking systems. Two days later, Sierra released version 1.0 of its Personal Agent Protocol, which every Shopify store now publishes automatically, and OpenAI launched ChatGPT Wallet so agents can pay without seeing a card number.

If you sell anything online, you now have an agent policy, even if you never wrote one down. The right answer can differ from one action to the next. Sort what an agent can do with you into reads, reversible changes and irreversible actions. Open the reads, watch the reversible changes, and put a human confirmation in front of anything that moves money or can't be undone.

What happened this month with personal AI agents

A personal AI agent is an assistant that acts for one customer, such as booking, buying or checking an account, rather than an agent the business runs itself.

United's announcement says it is limiting access both through its own digital platforms and through travel agencies and aggregators, to stop use by "unauthorized, third-party personal AI agents or assistants." It says customers who share their login with an outside AI tool can expose sensitive information. Agents can cause "unintended purchases, account changes, or use of MileagePlus miles." And some steps, like confirming at check-in that bags hold no banned items, have to be done by a person under FAA and DOT rules. The announcement describes no approved route for agents.

Sierra's Personal Agent Protocol, nicknamed Poppy, goes the other way. It gives a business a standard way to tell a visiting agent what it offers and how to connect, and it is built largely on existing standards such as MCP, OpenAPI and UCP. Customers sign in with OAuth (the familiar "allow this app to access your account" screen) and decide what their agent can and can't do. According to The Next Web's report on the release, Meta's Muse agent, which more than 10 million people have set up, now checks for Poppy first on every site it visits. The same report notes that the spec is still labelled a draft and may change in ways that break existing setups.

ChatGPT Wallet deals with the payment step. Before each purchase, the user reviews the items, the total and the card, then approves with Face ID or a fingerprint. The business receives a payment token instead of the card number. For now it is limited to ChatGPT Pro subscribers in the US.

Three ways to handle AI agents on your website

Approach

Who is doing it

What you gain

What you give up

Block unauthorized agents

United Airlines

Protects customer accounts, purchases and miles from agent mistakes

Sales from customers who prefer to delegate, and you still have to spot agents to block them

Publish what agents may do

Shopify stores via Poppy; design partners such as Mercury and Turkish Airlines

Agents find a supported path instead of scraping your pages

You depend on a draft spec and must keep your declared actions accurate

Let agents act, but a human approves payment

ChatGPT Wallet; Turkish Airlines upgrades

The customer confirms every charge, and you never see card details from the agent

One extra step on every purchase

These approaches can be combined. Sierra's examples do exactly that: Turkish Airlines lets agents check fares and find upgrades, and the customer approves before paying. Mercury lets agents start payments, but they go through the account's own approval rules.

Which agent actions should need a human?

United's list of worries is a useful map, because each item is a different kind of risk. Checking a booking is a read: nothing changes. Most account changes can be reversed. Spending miles or buying a ticket moves value and is hard to take back. Confirming what is in a bag is a legal statement only a person can make.

Group your own actions the same way:

  • Reads, such as order status, invoice downloads and opening hours. Let agents do these freely; they save your support team time.
  • Reversible changes, such as updating a delivery address before dispatch or adding a seat to a plan. Allow them, send the customer a confirmation, and log that an agent made the change.
  • Irreversible or money-moving actions, such as payments, cancellations that delete data, refunds and anything with a legal declaration. Require the customer to confirm, in your interface or through a payment flow like ChatGPT Wallet's.

We wrote more about drawing this line for your own automation in what needs a human's approval. The same test works whether the agent belongs to you or to your customer.

A worked example: a three-person SaaS team

Say a three-person team sells scheduling software on monthly and annual plans. Customers already ask their assistants to "sort out my subscriptions," so agents are showing up at the billing page and in the support inbox. Here is how the team could set a policy in an afternoon.

  1. List every action a customer can take in the account area. They find 14, from downloading an invoice to deleting the workspace.
  2. Mark each as read, reversible or irreversible. Invoice downloads and plan details are reads. Adding or removing seats mid-month is reversible, because the next invoice is prorated. Switching to an annual plan, cancelling, and deleting the workspace are irreversible.
  3. Open the reads to agents and write them up in a short page that agents and people can both use.
  4. Allow the reversible changes, but email the account owner each time with an "undo" link.
  5. Keep the irreversible actions behind a confirmation the account owner has to click in the product, so an agent can prepare the change and a human finishes it.
  6. Review the agent log once a week for a month and move actions between groups based on what actually happens.

Under this policy the agent handles the paperwork and the account owner makes the decisions that cost money or data.

Personal AI agents FAQ

What is the Personal Agent Protocol?

It is a protocol from Sierra that lets personal AI agents connect with businesses through their websites, APIs and support agents. Version 1.0 was released on November 13, 2026, after a first draft on October 9. Customers grant access with OAuth and can change what their agent may do over time.

Should I block AI agents from my website?

Blocking makes sense when mistakes are costly and regulated, which is the reason United gives. For most small businesses, a blanket block also turns away customers who delegate routine tasks. A middle path is to allow reads and require human confirmation for payments and deletions.

How does ChatGPT Wallet pay without a card number?

The user adds a card once and approves each purchase with a biometric check. The payment network then issues a token that stands in for the card, so neither the agent nor the business sees the real number.

Do I need to do anything if I'm on Shopify?

Sierra says every Shopify store now publishes a Poppy file automatically, with no extra setup for merchants. It is still worth checking which actions your store allows and how your checkout asks the customer to confirm.

The same rules apply to agents you run

If you also use AI agents inside your own business, the questions are identical. United's first complaint was about customers handing their password to an outside tool. The safer pattern is a scoped grant: in SelfAgentic, connecting a tool is a separate decision from authorising a particular agent to use it, and each agent has hourly and daily action limits. By default, anything that would post publicly or change an ad needs your approval, and on unattended runs it waits in your Approvals inbox until you decide. The agent prepares the action and a person confirms it.

This week, list your customer-facing actions, sort them into the three groups, and decide where a person has to click. If you want to try the same model with your own team of agents, you can start building on SelfAgentic.

Keep reading

Ready to build your AI team?

See how a team of agents handles real work, and checks with you before anything goes out.

Free to start. No credit card required.