236: ChatGPT Wallet Is Here. OpenAI Wants to Be the Middleman When You Shop.

0:00You can now tell ChatGPT to go and buy something for you. And it will. With your card. The feature is called ChatGPT Wallet. OpenAI announced it on November the 13th. Friday the 13th. On that same day, they published instructions for the stores. And one of the first things they say to the stores is that ChatGPT can already use their websites. And they don't need to do anything to put it on a list of approved visitors. So the store does not get asked. ChatGPT will show up either way. This is no hacks. AI is changing the way you use the internet and nobody has asked you about it. So I check what AI is actually doing to it.

0:40The good, the bad, and the pointless. Should you let ChatGPT buy things for you yet? I expected the wallet, the money part, to be the risky part. And I was wrong about that. The wallet is good. The problem is what's being built around the wallet. Step by step, OpenAI is putting itself between you and the stores you buy from. That's a middleman. Or as OpenAI calls it, buying directly. AI shopping is a new way to discover and choose what to buy without ever looking at a store's website. You ask, the AI finds, compares, and now buys for you. And if you never look at the store yourself, whoever does the looking for you matters a lot.

1:22Right now, almost nobody shops this way. DoorDash, one of the biggest food delivery apps in the US, says outside AI agents have placed maybe several hundred orders in total. In total. And still, every few weeks, somebody builds another piece of this puzzle. It's a bit like self-checkout when it first showed up. Supermarkets put the machines in. Most people still went to the cashier. And the machines kept getting better anyway. And the rules for it are being written right now. OpenAI's agents are called dots. A dot is ChatGPT going off and doing a task for you in a browser, in the cloud, not on your phone or your computer.

2:00It reads pages, it clicks buttons, fills in forms. A purchase goes like this. You add your card once. The dot goes to the store's website and fills the card like you would. Before it pays, it asks you. You see the items, the total, and the card, and you approve it with your face ID or your fingerprint every single time. Honestly, that part is really well designed. And then Visa or MasterCard or American Express makes a stand-in card, a different card number with its own security code, its own expiry date, it's the same idea as Apple Pay except this one is good for that one purchase only. The Dot uses that stand-in at the store's normal checkout and the store's payment company charges it.

2:41The money comes off your real card and the store never sees your card number. For now, this is only ChatGPT Pro, OpenAI's expensive tier in the US and only through Dots. It started on November 13th, so there's no number yet on how many people have used it, and OpenAI has not named a single store that's been paid through it. And the pro part? I wonder if that's a look at the future. Google's newest model in its AI search is only on its most expensive plan. Meta says its AI only makes phone calls for you if you pay. The AI that actually does useful things shows up first for the people who pay extra.

3:21Maybe that will change in the future. Maybe it will not. Either way, your card is fine. But the buying itself, it happens on the store's website. And that's new because this is not OpenAI's first try at shopping. The last one was called Instant Checkout. You bought inside Chad GPT and the store had to build something to be there. Six months after it was announced, about 30 Shopify stores had, according to Forrester, a research firm, used it. And then in March this year, OpenAI dropped it. Its page now says it is moving away from a standalone instant checkout experience. And that was the right call.

3:58Last time, OpenAI asked the stores to come to it, and very few did. This time, OpenAI is going to the stores. And those instructions OpenAI has for stores, this is what they're for. OpenAI calls them a playbook. You upload your products with prices in stock. You make sure your security lets OpenAI's agent through. You talk to the company that checks your orders for fraud. Look, uploading products is not new, but now it's one of the first things OpenAI asked for. Back in June, I had Malte Lanver from Peak AI on the show. Peak is a company that tracks what AI tools recommend. His team, together with an independent researcher, looked at ChatGPT's shopping results.

4:39The products, the prices, the ratings, they matched Google Shopping. ChatGPT was showing you Google Shopping data. Now OpenAI is asking every store for its products directly. In the playbook, there's a list of questions stores ask with OpenAI answers. So a store might ask, can I require that a human confirm the checkout? OpenAI says it does not have features for that yet. And it adds that across all its users, it requires human review before any purchase like that. So the human who confirms this is you if you're buying. The store cannot ask for anything more. A store might ask, can I keep agents away from certain pages or products?

5:22The answer, the exception that lets the agent in can apply site-wide. And whatever page level controls the store already has stay in place. So there's no special switch to keep the agent away for some products. A store might ask, who handles payment or order errors? OpenAI answers, the customers are ultimately responsible for the actions they take. So if the agent orders the wrong size to the store, that's you ordering the wrong size. And then there's fraud. Fraud tools are built to stop bots and OpenAI's agent is a bot. So purchases can fail because of this. OpenAI's fix is to start sending the store extra information about you.

6:05Your IP address as OpenAI sees it, your country, your time zone, your language. The store will see most of that if you visited yourself. but now the agent carries all that data for you. And if the store doesn't like any of this, well, there's a question for that too. How will this affect my relationship with my customers? OpenAI's answer is that the websites that blocked access may lose users or customers who turn to other sources to make decisions. This is important. OpenAI also says that a billion people use ChatGPT every week. That's their own numbers. Some businesses are saying no to this kind of setup anyway.

6:42This week, United Airlines said AI agents can show you United flights, but they cannot book them for you. United's reasons are privacy, purchases people did not mean to make, and federal rules. When you book a flight, you have to confirm things. Like that question, do you have anything dangerous in your bags? An agent cannot do that for you. And then United CEO Scott Kirby gave the other reason at an event on Tuesday. According to the Wall Street Journal, he said, To be honest, I don't want to create another middleman. The way I read that for an airline, the middleman is whoever you book through.

7:21They get your email, the upgrade offer, the next trip. An AI agent that books for you would be a brand new one sitting on your phone. On the wallet page, OpenAI says shoppers transact directly with businesses which keep their customer relationships and control over pricing, fulfillment, returns, and customer service. And I don't know who's right here, OpenAI or United, but obviously this is a conflict because OpenAI never says they're not the middleman. I think OpenAI is trying to get in as a middleman without addressing the fact that they're a middleman. And my test for what a middleman is, is simple.

7:56Do they take a cut? Do they collect the data or anything else? If yes, they're a middleman. So OpenAI, a cut. OpenAI told CNBC it's not taking a cut to begin with. and it doesn't even process the payments the store's payment company does. Data? That's a different story. A store that uploads its products sends OpenAI's prices in stock. And the conversation where you decided what to buy happens in ChatGPT, not on the store's website. They can't even see your page views in analytics. Anything else? Also, yes, the product list the store uploads can also feed ChatGPT's ads if the store switches that on.

8:38So that's a middleman. On that same page, OpenAI calls this buying directly from the business. That's directly through OpenAI's agent paying with OpenAI's wallet in OpenAI's browser. I'm not sure OpenAI is trying to avoid the word, but calling that directly is not honest. And a middleman is not necessarily bad. Apple Pay sits in the middle too, but Apple Pay only carries the payment. You still pick the store. You look at the product. You decide what to buy. Here, the AI does the looking and the choosing inside ChatGPT. And the store sees the agent and the order, never the conversation that chose it.

9:22That's the difference. Apple Pay is a tool you use. This is a tool that does things for you and making the choice is one of the things it does. So what does having another middleman cost you? If the agent gets it wrong, the mistake is yours with a store that never agreed to the agent in the first place. And the choice of what to buy and where happens somewhere the store cannot see and you cannot check. and that's why for me this is a not yet not for anything that matters most people cannot use it yet anyway but when it gets to you i would do three things the one moment you do control is the review screen the items the total the card that's the last thing between you and the order read it the way you would read it if you fill the cart yourself because for the store that's exactly what happened.

10:15Use it for things you'd be fine getting slightly wrong. A reorder from a store you already know is fine. A flight isn't and neither is anything you'd need to return or argue about. And if it fails, it might not be broken. The store might have said no. The store's website still works. If you happen to sell anything online, there's a setting that has probably decided this for you. I checked my website security dashboard on Cloudflare, which sits in front of a lot of websites. There's a section called personal agents. It lists four AI agents by name, Claude Duet from Anthropic, Dots from OpenAI, Gemini Spark from Google, Muse from Meta, and there's a fifth one coming soon.

10:59Each one is a dropdown. All four were set to allow for me by default. I never touched those settings. That's how they came. And for agents, Cloudflare even marks allow as recommended. On my website, on Nohack's website, that is completely fine. I don't sell anything there. But if you have a store, that dropdown is the kind of decision United Airlines made. The difference is that they made it on purpose. If your website is not on Cloudflare, the same setting usually sits with whoever runs your security layer. If I had a store, it would depend on the store. If the orders are small and returns are easy, I'd let them in.

11:38And then I'd tell my fraud provider about the extra information OpenAI sends so the agent's orders do not get declined. If a wrong order is expensive or there's something the customer has to confirm themselves, I'd do the same thing United Airlines did, at least for now. The product upload is a separate decision. It helps ChatGPT show your products. It also puts your catalog with OpenAI. If ChatGPT sends you customers, I would do that. If it doesn't, it's probably okay to wait. On November 10th, the CEO of an airline looked at AI that books for you and called it a middleman. Three days later, OpenAI launched AI that buys for you and called that buying directly.

12:23One of those is wrong. Thank you for tuning in. I would like to remind you that every week I have a newsletter at nohacks.co slash subscribes. subscribes. And as far as this episode is concerned, please consider rating, reviewing, and sharing it with someone who would benefit from it. I will talk to you next week.