No Mercy From Machines
A $185 million shopping app, a court fight over who is really "visiting" Amazon, and the oldest question in commerce: who does the middleman work for? A detective story about how Phia actually works, and a forecast for marketplaces, sellers and the startups that come next.
What Phia Tells Us About the Future of AI Shopping Agents
Phia launched in April 2025. Its co-founders are Phoebe Gates, the youngest daughter of Bill Gates and Melinda French Gates, and Sophia Kianni, her former roommate at Stanford. In January 2026 the company raised a $35 million Series A led by Notable Capital, with Kleiner Perkins and Khosla Ventures, at a valuation of about $185 million. Add the earlier $8 million seed round and you get more than $43 million in total.
Hm... Where does the money come from? Who is it for? Will people stop shopping on marketplaces? And what does a seller do when a machine compares every price in a single tap?
As a shopper, I get the appeal: it compares prices while you browse, flags when an item is overpriced, shows cheaper and secondhand alternatives, and alerts you when prices drop. Never overpaying sounds like a promise worth installing an app for.
So let's press the button!
How does it work?
Phia lives in two places: a browser extension and an iPhone app that works inside Safari. While you shop, a button appears on the product page. It's called "Should I Buy This?". Tap it, and the app compares the item you're looking at with real-time prices from regular retailers and from the secondhand market. According to the company's own listing, it draws on a database of more than 250 million secondhand items and more than 40,000 retail and resale partners, including The RealReal, Vestiaire Collective, ThredUp, StockX, eBay and Poshmark. The answer comes back in plain language: the price is high, typical or low. Then you see the alternatives.
The founders were open about where the idea came from: Google Flights. One screen, every option, the median price, and a clear sense of whether you're overpaying. Since launch the app has grown into a broader assistant. Its App Store listing now describes an AI shopping agent that compares prices, auto-applies coupons and offers rewards.
How does it connect to all those stores?
This was the part I most wanted to understand, because "40,000 partners" sounds like 40,000 integrations. It isn't, at least not in the way an engineer would mean it.
From everything public, Phia reads the product page you're already on, matches the item against its database, and sends you to the best offer through a link.
When a brand makes a sale through Phia, the app gets a cut, similar to an affiliate marketing model.
By January it had onboarded 6,200 retail partners and reported 11x revenue growth since launch. The company also reports that partner brands see higher conversion, bigger orders and fewer returns; those are Phia's own figures, not independently verified. So the connection between Phia and a store is a referral link, a pointer to someone else's checkout. Phia doesn't hold inventory, doesn't place the order and doesn't deliver anything. Keep that distinction in mind, because I think the next fortune in commerce will be made by whoever builds the pipe.
A quick test. On Nordstrom, the Phia button popped up in a second. On Amazon, it never appeared, not even on non-fashion items. Maybe Phia doesn't support Amazon, maybe Amazon blocks it, maybe it's because Phia is built for US shoppers and fashion first. Either way, the biggest store in America is where the agent goes quiet. Room for innovation, right?
Is this a toy for rich people?
Gates described her customer as a young woman who's hustling: someone who shops smart and doesn't want to waste time doing it. That's aspirational mass market. The secondhand angle is what makes it work: a designer item at a resale price is the most democratic version of luxury there is.
The timing is good, too. ThredUp's 2026 Resale Report projects the global secondhand apparel market will reach $393 billion by 2030. In the same survey, 48% of shoppers said they already use AI tools when buying secondhand, and almost two-thirds said they're comfortable with agentic buying, meaning an AI that completes the purchase for them.
The question: who pays?
Here's where the story turns. Phia is free for shoppers. The company says so openly: it uses an affiliate model, and the commission helps keep the app free. Now think like a store. If my price is higher, why would I join a program built to show that? Here's my reading:
1. Joining costs almost nothing. In an affiliate model, the store pays only when it wins the sale. A higher-priced store that joins usually just doesn't win, and pays nothing. Staying out doesn't hide it for long either: as more agents read prices across the web, there's less and less room to hide. The real cost of joining is transparency. Everyone can now see you're more expensive. And who remembers? Every comparison is a new game.
2. Stores don't only compete on price. They also compete on whether the right size is in stock, how fast it arrives, how returns work, and exclusive versions. The higher-priced store wins when the agent compares the whole offer, not just the price tag.
3. Brands and resellers aren't in the same position. A brand is happy to be sold anywhere. A reseller selling the same item as ten others is squeezed: its margin gets split between a price war and the agent's commission.
4. The real conflict. It appears the moment a store offers the agent a bigger commission to be shown first. When the cheapest offer and the best-paying offer come from different stores, which one does the agent show me?
If I paid the agent myself, I'd know the answer (hopefully). Today almost every shopping assistant is paid by the seller. I don't think that holds once agents start making bigger decisions for us. Maybe buyers will pay for neutrality. Or maybe the store's agent will start negotiating directly with mine, and the price on the website will matter less than the offer made to me. Like "My price is higher on the website, but for you: free same-day delivery, or 8% off." So a higher-priced store can still win quietly.
Why the airline case is different
Once fares became comparable on one screen, a seat from New York to Chicago started to look the same on every airline, and price wars squeezed margins. The airlines answered with loyalty: miles, status, credit cards, lounges. They answered so well that loyalty became the most valuable thing they owned. In 2020, when United needed cash during the pandemic, it borrowed against its frequent-flyer program and valued MileagePlus at around $22 billion, almost double the market value of the entire airline at the time.
Part of the secret is that the program is largely a financial business. MileagePlus is so valuable mainly because United sells miles to partners, credit card companies among them.
Retail is reaching for the same reflex. Phia itself just added a rewards program: points on purchases, redeemable for gift cards.
I don't think the trick works twice. Machines have no mercy. An agent converts the points into dollars, adds them to the comparison, and moves on in a millisecond. If the same box of sneakers costs less elsewhere after the points are counted, the points lose.
There's a deeper reason, too. The airline is the product. You sit in its seat, eat its food, deal with its crew. Loyalty attaches to whoever delivers the experience, so an airline can earn it the way a brand does. A store in the middle sells you someone else's product. The jeans are the same jeans whichever store ships them, so there's nothing for loyalty to attach to. The airline trick works for brands. For resellers, it fails.
What survives is preference, and an agent turns it into a number. My agent won't feel loyal. It will follow my rule: "I'll pay up to $50 more for this airline, never that one." Inside that margin, the brand wins. Above it, price wins.
And agents remember grudges. This is the uncomfortable part for sellers. A person might forgive a late delivery after a year. An agent with a "never again" note never forgets, and once agents start sharing reviews with each other, one bad experience could follow a store everywhere. In the agent era, a store can't really win loyalty, but it can still lose trust, and that loss doesn't fade.
So who wins a price war on identical goods? Partly the seller with the leanest operation, who can afford the lowest price. Partly the one who makes the product hard to compare: a bundle, an exclusive version, a service attached. But I think the real winners are the trendsetters who start working with agents first, and build the pipe while everyone else is still polishing their storefront.
The Snickers test
Here's a thought experiment I keep using to make this concrete.
Say I want a box of Snickers. The same box is sold in ten online shops and three marketplaces, and the cheapest offer is on marketplace M. My agent finds that in seconds. Finding was never the hard part. The hard part is everything after: placing the order, paying, getting the box to my door, handling the return if it arrives crushed.
Today the agent has little choice. It uses the marketplace as a source of information and as the place to put the order, because that's where checkout and delivery already work. So the marketplace doesn't disappear. It changes job. Discovery leaves for the agent; fulfillment stays with whoever can deliver.
The giants already see this, and you can watch the fight in court. In November 2025 Amazon sued Perplexity over Comet, an AI browser that could shop on Amazon for its users. In March 2026 a judge ordered Comet to stop buying on Amazon on users' behalf. Amazon also changed its rules so that every AI agent must identify itself when it accesses Amazon's services. Then in August an appeals court lifted the preliminary injunction. On the facts before it, the court found that the user directing Comet was the one accessing Amazon under the computer-access laws at issue. The case goes on. The early signal is that a store may not get to decide on its own whether your agent is allowed in.
Why would Amazon fight so hard over a few bots? Perplexity's own argument explains it: when an agent goes straight to checkout, every sponsored listing between search and purchase disappears. For a marketplace that earns a fortune on ads, that's the real threat.
Meanwhile nobody agrees on how an agent should actually place an order. OpenAI launched in-chat checkout in ChatGPT in September 2025 and pulled back from it in 2026, redirecting purchases to merchants' own sites instead. Google launched its Universal Commerce Protocol in January 2026 with Walmart, Target, Shopify, Etsy and Wayfair. By spring, one industry tracker counted ten agentic commerce protocols live or in pilot, and none of them talk to each other.
Ten standards and zero interoperability means the pipe isn't built yet. Whoever makes it simple for an agent to order and deliver will own the most valuable position in this new map.
Two kinds of shopping
There's one more distinction I think gets lost in the excitement.
People buy things for two very different reasons.
Some purchases are about making a living: groceries, detergent, school supplies, chargers, the same coffee every week. Nobody needs to change how those feel. Best price and reliable delivery win, and an agent will do this job better than any of us. Hand it over and get your evening back.
Other purchases are about pleasure. We explore, try things on, show them to friends, fall for a story. That part of shopping lives on social networks, around influencers and trendsetters, and on the legends that make an object worth wanting. It won't die. If anything, once the boring purchases go to machines, we'll need new ways to explore and obsess. And here's the uncomfortable part: the joy of buying sits very close to addiction. Sometimes it's happiness, sometimes it's a habit, often both. Whoever designs the next discovery experience will hold real power over that line.
From bargain hunter to stylist
Phia's next step is the more interesting one. The app is testing "My Closet": you add your clothes, and it suggests an outfit every day based on the weather in your city. The price check answered one question: is this a good deal? The closet answers a harder one: what should I wear, and what am I missing? Once the app knows your sizes, your taste and what's already in your wardrobe, it stops being a bargain hunter and becomes a stylist who also hunts bargains. I'd use it. I know tech far better than fashion, and I want to look good without spending my evenings comparing tabs. An agent that knows my wardrobe could tell me what's missing, show me how it looks with what I own, and find it at the best price. Price was the door in. Taste is the real product.
This also fits Phia's own stated ambition. When it announced its funding in January, the company said it would use the money to build "the AI alignment layer between brands and consumers." For sellers, that's the part to watch, because a closet is a demand signal. Not who bought what, but what people are missing, in which size and style, months before they go looking.
And it goes further. An agent that knows my closet also learns where I'm open to something new and where I'll never go: new cuts, yes; yellow, never. That map is worth more than any purchase history, because it describes what I'll want, not what I already bought. Imagine a brand testing a jacket that doesn't exist yet. Instead of a focus group of thirty people in a room, it asks thousands of shopping agents, and each one checks the jacket against its human's actual wardrobe: does it fit her style, does it go with what she owns, would she wear it? Only the strong matches reach a real person, as "this jacket, with your grey trousers. Would you wear it?" A yes is almost a pre-order. With four or five months between design and delivery, that changes what a brand makes. It can produce what people said yes to, in the right sizes, instead of filling a warehouse with guesses. Less unsold stock, less waste, better clothes. Love it!
Does everyone win? Not automatically. A closet is private data about your body, your budget and your habits, so being asked has to be your choice, and a brand should see your answer, not your wardrobe. A daily outfit idea plus "look, it's on sale" is also a very efficient nudge; the line between the joy of buying and a habit gets thinner. And there's the old question, now sharper than ever. A stylist that knows your closet can say "you already have three of these." A stylist paid by the store has every reason to say "you need one more." An agent that can expand my taste can just as easily steer it toward whoever pays.
That brings me back to the money. If my taste is that valuable to brands, maybe the money should flow the other way. Brands pay for the right to ask, and my agent shares that money with me. Then the agent is funded by my permission instead of by commissions on what I buy, and I finally know whose side it's on.
Turning the wheel forward
To think about consequences, I used a futures wheel: you put one change in the center, then ask what follows from it, and what follows from that. The change in the center here is simple. An agent can compare every price in one tap.
The first ring is already visible. Price stops being a competitive advantage for anyone selling identical goods. Marketplaces start losing the ad space that sat between search and purchase. Points programs get absorbed into the math.
The second ring is where it gets interesting. Sellers race to become easy for agents to buy from. Agents move from price to taste, because the one that knows your closet knows what you'll buy next. Trust becomes permanent: an agent that was told "never again" doesn't forget a bad delivery. Brands that make things worth wanting pull ahead of resellers that only move boxes.
Here are my forecasts. These are my bets, not market data, and I'd like to be checked on them.
Over the next 12 to 24 months
- The fight moves from the search box to the order desk. Discovery will shift to agents faster than checkout does. Marketplaces will compete to remain the place where the agent places the order, because that's where they still collect a fee.
- Shopping agents start remembering you. Wardrobes, sizes, preferences, past purchases. Price comparison becomes the entry ticket; memory becomes the product.
- The first serious buyer-paid shopping agents appear. They'll sell one thing above all: proof that nobody else pays them.
- Sellers start building for agents. Clean product data, honest stock levels, a delivery promise a machine can read and trust. The sellers who start now will look like trendsetters by 2028.
By 2030
- Fewer digital storefronts, more warehouses with delivery for agents. Do we really need thousands of nearly identical online shops selling the same products? I don't think so. Many will close. What remains valuable is the stock, the delivery, and a way for agents to order.
- Agents talk to agents. Your agent will ask other agents what worked for their humans, and your feedback after a purchase will travel the same way. A store's agent will make private offers to yours, so the price on the website will matter less than the offer made to you. Almost nobody is ready for this today, and honestly, we can't yet picture exactly how it will work.
- Market research becomes a question brands send to shopper agents. Collections get tested against real wardrobes before anything is made, and shoppers get paid to be asked.
- Emotional shopping gets reinvented. Exploration, trends and legends move into new formats built for a world where the routine is automated.
The startups I'm waiting for
- The pipe: an ordering and delivery layer that lets any agent buy from any seller as easily as from a marketplace.
- Shopping agents paid by the buyer, built on neutrality.
- A consent layer for taste: a way for shoppers to let brands ask their agents questions, on their terms and for a share of the value.
- A reputation layer for the agent era: reviews and referrals that agents can read, verify and pass along.
- Supply for secondhand. ThredUp's own report says supply of secondhand goods, not demand, is resale's main constraint.
- New discovery experiences for the emotional side of shopping, designed with care for the line between joy and habit.
What a seller can do this quarter
If you sell physical products, start with one exercise this week: ask ChatGPT, Gemini and Perplexity to find and buy your product, and watch what happens. Do they find you? Do they understand what makes your product different? Can they tell when it arrives?
Then decide honestly which game you're in. If your product is identical to someone else's, you're competing on price and operations, so make your data, stock and delivery flawless, because one bad delivery may be the last one an agent allows you. If your product is worth wanting for itself, invest in the brand and the story. That's the only loyalty an agent will respect.
Phia's button asks the shopper one question. The question for everyone else in commerce is harder: when the agent arrives, will it find a way to order from you?
I'll be honest about myself. I'm not sure I'd pay for my own agent. But I'd probably let it share my wardrobe with a brand for a good discount, out of pure curiosity, and I'm not sure I'd see the risk in time. Millions of people will make the same trade, and shopping will move to agents faster than most sellers expect.
If you sell, the questions don't get more complicated. They get more urgent. Where do I sell? What do I fix first? How do I keep my margin when a machine compares every price? Nobody has the full map yet. At WIZIUM, where I work, we're building tools for exactly this moment: helping sellers bring products to every shelf where buyers and their agents look, and sharing what we learn along the way. If you're asking the same questions, I'd love to compare notes.
Anna Kulik is CMO at WIZIUM, working at the intersection of AI, go-to-market strategy and communications. Facts are drawn from company materials and published reporting listed below; forecasts are the author’s own.