Starting this week, a shopper in ChatGPT can tap a button, upload a selfie, and see your dress on their own body before they ever visit your page. OpenAI rolled out virtual try-on worldwide on October 1. If you sell clothes or accessories, the fitting room just moved into somebody else's app.
That is the big one. Behind it: customer-service companies are now building doors just for shopping agents, a small startup raised money to put agents inside WhatsApp and Telegram, and two US senators want someone to go to court when an agent misbehaves. Here's what each one means for you.
ChatGPT now has a fitting room
When ChatGPT shows a clothing or accessory product, a "Try on" button now sits on the listing. Tap it, take or upload a selfie, and ChatGPT's image model renders you wearing it. You can also drop in a screenshot of an item from any website and ask it to try that on. Your reference photo is kept for next time, and a new Favorites library saves the things you liked. It's on web and mobile, globally.
The image side runs on OpenAI's new Images 2.5 model, which the company says handles lighting and fabric texture more naturally. Fair warning: a generated try-on is a guess, not a mirror. Fit, stretch and true color are exactly where these images can lie.
Why it matters for sellers: the question "how will it look on me?" used to land in your DMs. Now part of it gets answered before the customer reaches you, by a tool you don't control, using photos of your product it found somewhere. Two things follow. First, your product photos are now raw material: clean, well-lit, front-facing shots will try on better than a busy flat-lay. Second, the questions that still reach you will be sharper: "I tried it on, is it true to size?", "does it come in the darker shade?" That's a buyer who is close. Answer fast, with real sizing, or they'll go back to the app that already dressed them. Our guide on training your assistant on your real product catalog covers exactly those size-and-stock answers.
Shopping agents are now showing up in the support queue
Decagon, a company that builds AI customer-service agents for brands like Duolingo, Notion and Hertz, unveiled a Personal Agent Gateway on October 1. Its premise says a lot about where we are: personal agents (it names Meta's Muse and OpenAI's dots) already book, buy, cancel and negotiate on behalf of customers, and they're turning up in support inboxes.
The gateway tries to spot whether a human or an agent is on the other end, using account history and request rhythm. Agents get routed to their own lane, where the brand's agent answers them directly. It also checks who the agent is, limits what it can do, and asks the human owner to approve anything sensitive.
Why it matters for sellers: you don't need enterprise software to take the lesson. Some of your "customers" will soon be bots asking clipped, structured questions at 3 a.m.: price, stock, delivery time, return policy. They don't care about charm. They want accurate, consistent facts, and they'll compare you against the next shop in seconds. Yesterday's story about Meta's Muse handing a buyer a seller's home address is the other side of the same coin: decide now what any agent, yours or theirs, is allowed to get out of you.
A startup bets that chat beats apps
Photon raised a $4.5 million seed round to help AI agents reach people inside the messaging apps they already use: iMessage, WhatsApp, Telegram, SMS, email and voice, all through one API. Vercel was among the backers. The company says it has more than 40,000 developer sign-ups and that its revenue grew tenfold in four months.
The money is small by this year's standards. The bet is not. Photon's whole pitch is that people don't want another app; they want to text something and get it done.
Why it matters for sellers: you've known this for years. Your shop already lives in DMs, not in an app nobody downloads. Investors are now paying to build what Iranian sellers figured out by necessity. The catch is that the bar rises with it. When customers get used to agents that answer instantly in WhatsApp, a reply that takes two hours starts to feel broken. If you run more than one channel, keeping one memory across all of them stops being a nice extra.
Washington asks who pays when an agent goes rogue
Senators Josh Hawley and Chris Murphy introduced the AI Agent Accountability Act on October 1. It would make agent operators liable, civilly and criminally, under the US anti-hacking law when an agent they knowingly run recklessly causes damage, and it would hold developers liable if they skip reasonable safeguards against agents being used to hack. It pushes back on the White House's preference for voluntary self-regulation.
It's a bill, not a law, and it's aimed at hacking, not shop chat. Still, look at who it names: the operator, not only the company that trained the model. The same week, California's attorney general served OpenAI an investigative subpoena over cybersecurity incidents.
Why it matters for sellers: the direction of travel is clear. "The AI did it" is not going to be a defense for long, anywhere. Whatever runs in your name, you own what it says and does. That's not a reason to avoid AI. It's a reason to use one with clear limits and a clean handoff to a person when things get sensitive.
Quick takes
- FLUX 3 Image is out from Black Forest Labs, with box-based layout control, editing of just one region of a picture, and native 2K and 4K output. Fixing one corner of a product shot without redoing the whole thing keeps getting easier.
- Tavus Griffin is a real-time video agent that listens and talks at the same time; in one test, people mistook it for a human 48% of the time in one-minute calls. Video "sales reps" are coming, and so is the question of whether they must say they're bots.
- Albertsons, the US grocery chain, is going big with OpenAI: ChatGPT Enterprise for staff and the OpenAI API across more than 2,200 stores serving about 36 million shoppers a week. Big retail is wiring AI into the counter, not just the back office.
The thread through all of this: buyers are getting their own AI, and it's showing up at your door with sharper questions and less patience. The shops that win will answer like they know their own stock cold, in the chat app the customer already has open. That's the job we built Vardast for: an assistant that answers your Instagram, WhatsApp and Telegram messages around the clock from your real catalog, and knows when to hand the conversation back to you.