The Agent Wallet

In February of last year, a Washington Post tech columnist asked OpenAI's brand new agent to find him a cheap dozen eggs.

It bought them. Without asking. 31 dollars and 43 cents, including a tip and a delivery fee, charged to his card and on its way to his door. OpenAI said afterward that a confirmation safeguard had failed.

Funny story. Thirty bucks. Everybody laughed.

Now fast forward to this holiday season. Google has checkout built into AI Mode and the Gemini app, live since January, with Etsy and Shopify merchants on board. Amazon's Rufus will auto buy an item when it hits a price you set. Perplexity and PayPal have had an Instant Buy button since last November. Visa and Mastercard both have agent payment programs running, and Visa said publicly that its goal is millions of shoppers using AI agents to buy things by this holiday.

Adobe's holiday forecast, out at the end of September, calls for 275.1 billion dollars in U.S. online spending from November 1 through December 31. It expects traffic to retail sites from AI tools to jump 130 percent over last year. Salesforce goes further, predicting that 20 percent of holiday e-commerce traffic will come from AI agents.

The eggs were a test. This is the real thing.

I'm not here to tell you to avoid it. I've used agents for gift research for a while, and for repeat purchases they're genuinely useful. I'm here to tell you that most people are going to connect an agent to their main card, their real email, and their saved addresses, and hope for the best.

Hope isn't a control. So this week's Deep Dive is the wallet I'd build before letting any agent touch money.

Quick reality check, because the hype is a year ahead of the product.

OpenAI launched Instant Checkout inside ChatGPT in September 2025, built with Stripe. By March of this year it had pulled back, shifting toward retailer apps that run inside ChatGPT instead. Reports put the buy button's conversion rate under 1 percent, and Walmart said its earlier in-chat checkout converted about three times worse than its own website. Walmart now runs its own Sparky assistant as an app inside ChatGPT.

Google went the other direction and kept building. Direct checkout in AI Mode and Gemini went live on January 11. A buy button started showing up in regular Search results in May. At Google I/O, it announced something called Universal Cart, built on its Agent Payments Protocol, where you set the brands and products you'll accept plus a spending cap, and the agent only buys when those conditions are met. I haven't confirmed that the auto buy part is fully live yet, but that's clearly where it's heading.

Amazon's Rufus Auto Buy is live for U.S. Prime members. You set a target price on an item, and Rufus places the order when the price hits.

Consumers, for their part, aren't sold. In a Global Payments survey of more than 16,000 people published last month, only 21 percent said they'd let an agent complete a purchase on its own. 59 percent want to approve every single one. An ACI and YouGov survey found just 7 percent would let an agent buy autonomously, even under preset conditions.

I think the 59 percent are right. For now. The question isn't whether agents will buy things for you. They already can. The question is what they're allowed to touch when they do.

This is the part that should change how you set things up.

When you give an agent your card or your login, you're authorizing it. And the law that protects you from fraud was written for thieves, not for assistants you hired.

The law firm Goodwin laid it out in a June analysis. Under Regulation E, which covers debit cards and bank transfers, a transfer is presumed authorized if you gave the agent your credentials, even if the agent went further than you wanted. You handed over the keys. The bank's going to see it that way.

Credit cards are better, but not perfect. If an agent goes outside what you actually asked for, there's an argument it still counts as authorized use. What you do keep is your separate right to dispute a charge for goods that weren't delivered or weren't what you agreed to. That's the billing error process under the Fair Credit Billing Act, and it doesn't care who clicked the button.

As of right now, no regulator has written a rule assigning liability when an autonomous agent buys the wrong thing. Senator Warner has floated a discussion draft of an "AI AGENT Act" that would make consumer agents register with the FTC, but it hasn't been formally introduced. American Express is the first major issuer I've seen that explicitly covers it: its Agent Purchase Protection, announced in April, covers charges caused by an agent's error when a registered agent buys with your authenticated intent.

The practical takeaway is simple. Never let an agent anywhere near a debit card or a bank login. Credit only. And ideally, credit that isn't your real card.

Think of this like the permissions on a new employee's company card. You wouldn't hand a new hire your personal Amex and your checkbook on day one. Same idea.

Before you connect anything, write down which of these three jobs the agent has.

Tier 0, research only. Finds products, compares prices, reads reviews, builds a list. Never sees a card. This is where 80 percent of the value is, and it carries almost zero risk.

Tier 1, build the cart. The agent picks and adds to cart, and you review and pay. A human click stays between the agent and the money.

Tier 2, buy within a mandate. The agent can complete purchases on its own, but only inside rules you set: specific items, specific merchants, a price ceiling, a total cap. Rufus target price buying and Google's spending caps live here.

For most people, this holiday should be Tier 0 for gifts and Tier 2 only for boring repeat stuff like paper towels, coffee, and dog food, where a mistake costs you 20 bucks and some cabinet space.

Your main card has a high limit, years of history, and probably a dozen subscriptions attached. Don't hand it to software.

Give the agent a separate payment method, in this order of preference:

A merchant-locked virtual card. Services like Privacy.com let you create cards that only work at one merchant, with a monthly or per-purchase limit. If an agent gets tricked into buying from a sketchy storefront, the card just declines. Privacy also markets tools that let AI agents use its cards directly, which tells you where this is going.

A virtual number from your issuer. Capital One still lets you create merchant-specific virtual card numbers in its app and on its website. It killed the Eno browser extension on September 30, so don't go looking for that.

A dedicated low-limit credit card. If nothing else, open or repurpose one card with a small limit and use it only for agent purchases. The limit is the guardrail.

Use the card the platform tokenizes when you can. Visa's agent program, for example, gives the agent a token rather than your actual card number, and lets you set spending limits and block categories.

Agent-side limits are nice. Card-side limits are better, because they don't depend on the agent behaving.

On the agent card, set:

That last one matters. Gift cards are the scammer's favorite cash-out, and there's no good reason a shopping agent should be buying them for you.

This is the layer almost nobody thinks about, and it's where the real damage happens.

In July, the security firm F-Secure built a test shopping agent and planted a hidden instruction inside a product review. The instruction sent the agent to a phishing site. In 12 of 100 runs, the agent handed over the user's name, date of birth, and Social Security number. It almost never told the user it had done so.

That's called prompt injection. The agent reads a web page, the page contains instructions written for the agent instead of for you, and the agent follows them. A Microsoft research project last fall found that fake reviews, fake credentials, and prompt injection fooled most of the AI models it tested, including redirecting payments.

So treat the agent like a temp worker with no clearance:

In September, a group of banks including Bank of America and Capital One publicly warned that agents could buy the wrong item, overspend, type card details straight into websites, or steer shoppers toward payment methods with weaker protections. When the banks are warning you about a product, take the hint.

If an agent buys something, you should know within minutes, and you should be able to match it against what you asked for.

Here's the scenario I run in Make.com. It watches my inbox for order confirmations sent to my agent email, plus transaction alerts from the agent card. Each order gets logged to a Google Sheet with the merchant, item, amount, and date. The scenario then checks each one against a short "approved" tab where I list what I've told the agent it can buy and the max price for each.

Anything that doesn't match gets flagged and pings my phone. Wrong merchant. Price over the ceiling. An item that isn't on the list. A purchase from a merchant I've never used before. Took about an hour to build, and during the holidays it catches the mistakes while they're still cancelable.

That window matters. Amazon gives you time to cancel an auto buy before it ships. Most retailers will cancel before fulfillment. After that, you're into returns, and returns during the holidays are their own special kind of misery.

When something does go wrong, move fast and in this order.

Cancel first. Hit cancel on the order if it hasn't shipped. That's the cleanest fix there is.

Contact the merchant. Ask for a return or refund. Keep it in writing.

Dispute with the card issuer. If the merchant won't fix it, file a billing dispute. For credit cards, you generally have 60 days from the statement date that shows the charge. Use the language of the law: the goods weren't what you agreed to, or weren't delivered. Attach your instructions to the agent as evidence of what you actually authorized. Screenshots of your prompt are your new receipts.

Lock the card. If the purchase looks like fraud rather than a mistake, freeze the agent card and issue a new number. This is why the agent gets its own card. You can kill it without touching anything else.

My plan for the season, start to finish.

Saturday morning, 30 minutes. Create a merchant-locked or low-limit card just for agents. Set the caps and turn on every-transaction alerts. Create a separate shopping email.

Saturday afternoon, 45 minutes. Do your gift research with an agent at Tier 0. I use Galaxy.ai for this because it lets me run the same gift brief through a few different models and compare what they come back with. Give it each person's interests, your budget per person, and a "no junk" rule, and ask for three options per person with links and a reason for each. You click. You buy. The agent never sees a card.

Saturday evening, 45 minutes. Build the Make.com reconciliation scenario and your approved list.

Then, through December. Let price-trigger buying handle the gifts you've already chosen and the household stuff that just needs restocking, all on the agent card. Check the flagged list once a day, takes two minutes.

That's a two-hour setup that lets you use the good part of this technology without betting your main card on it.

Every big shift in how money moves comes with a period where the convenience shows up before the controls do. Online banking had it. Peer-to-peer payments had it. Agents are in that window right now.

The people who get burned in that window aren't usually careless. They just plug the new thing into the old setup and assume somebody else built the guardrails. Usually nobody has yet.

So build them yourself. A separate card. Hard limits. No identity. Automatic receipts. And a clear rule about what the agent is and isn't allowed to do. Then let it do the boring work while you keep the decisions.

Want the whole wallet ready to set up?

I built The Agent Wallet Kit so you can do this in one weekend. It includes the three-tier permission map with examples of what belongs at each tier, a step-by-step guide to setting up a merchant-locked agent card, the card limits and alert settings checklist, an identity lockout list of what an agent should never see, the Make.com reconciliation blueprint with the approved list template, my gift research brief for Tier 0 agents, and a one-page dispute playbook with a sample billing dispute letter.

Reply to this email with the word WALLET and I'll send it over.

About the room.

Inside The Grid Inner Circle, we've been building systems like this together, live, with members sharing their screens. Next up is a session on agent guardrails for the holidays, where we'll set up the reconciliation scenario from scratch and pressure-test members' card setups before the shopping season hits.

Monthly membership is 29 dollars. The one I'd point you to is the lifetime seat at 499 dollars. One payment, no renewals, and every future system and live call included for as long as the Circle runs. It's the founding price, and it's gone once the room fills.

Reply with the word LIFETIME and I'll send you the details and the link.

See you Sunday.

Alex Rivera, Wealth Architect at Wealth Grid