We've spent a lot of time in this newsletter talking about AI agents reshaping how people discover and buy products — agents that compare prices, parse reviews, build carts, and increasingly make purchasing decisions on a shopper's behalf. Most of that conversation has focused on what happens online. But the ripple effects don't stop at the browser tab.

When AI agents start redirecting demand — concentrating it, accelerating it, making it less predictable — that doesn't just affect your digital shelf. It affects every piece of inventory on every shop floor that was ordered months ago based on a forecast that assumed humans would keep shopping the way they always have.

Which brings us to physical retail, and the obituary everyone keeps trying to write. 7,325 US store closures in 2024, the highest since 2020. JLL puts announced closures between 2024 and early 2025 at more than 9,900 locations. Cue the familiar headline: stores are dead, online wins, long live the algorithm.

Except store openings have stayed substantial over the same period. Much of the vacated space is being taken by stronger concepts. And some of the loudest "physical retail is finished" takes are coming from people whose entire business model depends on you believing that.

What's actually failing isn't the store. It's the financial arrangement sitting underneath it — one that agentic commerce is about to stress-test in ways it was never designed to handle.

The arrangement under pressure has a name, though nobody outside retail strategy circles uses it: the merchant model. Under it, a seller commits to inventory months before it has any real idea whether customers want it, ties up capital across thousands of individual products, absorbs the cost when the forecast is wrong, and then clears the mistake through discounting.

That end-of-season markdown rack? It's not a promotional tactic. It's a confession. It records the exact moment a retailer accepts it bought the wrong thing, in the wrong quantity, and now has to give away margin to shift it. Every 70%-off sticker is basically a tiny white flag.

Here's where it gets genuinely fascinating: the merchant model is not a physical retail phenomenon. A department store buying ahead of a season is running it. So is a first-party retail line on a marketplace. So is a DTC brand holding inventory in a fulfilment centre, betting on demand it can't yet see. The model is defined by who owns the forecasting risk, not by where the transaction happens.

Physical retail is simply the most exposed version — the most capital-committed, the slowest to adjust. When a bet goes wrong online, stock can be repriced in hours. When it goes wrong in a store network, it sits on shelves until the markdown rack gets wheeled out (accompanied, presumably, by a quiet sigh from the buying team).

There's a pattern in retail, first described in 1960, called the wheel of retailing. New formats enter lean and cheap, undercutting whoever came before. Over time they trade up — better locations, more services, higher costs — which leaves them exposed to the next low-cost challenger arriving underneath. Department stores, malls, discount chains, category specialists, and ecommerce marketplaces have each taken their turn.

The instinctive response has always been to invent a better format and climb back on. The more interesting question is whether the game itself is changing — whether the decision is no longer which format to run, but how much demand risk to own in the first place.

Once you stop thinking of retail as "who merchandises goods best" and start thinking of it as "who should hold demand risk," the current shift makes a lot more sense.

There are two ways to narrow the gap between supply and demand. One is to commit later — order small, watch what happens, reorder once real demand is visible. The other is to move the decision to whoever knows the product best.

A brand understands its own range, its own price ladder, and its own customer signals more intimately than any centralised buyer managing hundreds of competing lines ever could. (No offence to centralised buyers. Some of my favourite spreadsheets are made by centralised buyers.)

This is the logic behind store-in-store arrangements, where the retailer supplies traffic, location, and customer access while the brand controls assortment, presentation, service, and — crucially — inventory. The retailer stops acting as a merchant and starts acting like a platform. It monetises its traffic and real estate, and it hands the risk of forecasting demand to the party best equipped to carry it.

The apparent paradox: a powerful retailer can create more value by giving up some control.

If this sounds familiar, it should. Online retail resolved the same tension years ago. A marketplace doesn't buy and resell every item it lists — it lets sellers reach customers directly and charges for access. The store-in-store deal is the offline version of a bargain that Amazon-era sellers already live inside. In both cases, the risk moves onto the brand. In both cases, the platform wins by designing the rules rather than owning every product decision.

If the merchant model is what's straining, AI agents are the force pressing on the demand side of it — and they press regardless of channel. That's the connection the "physical versus online" framing consistently obscures.

An agent compares products, prices, reviews, and delivery options in real time and acts on a shopper's behalf. They mostly transact online for now (no robots browsing the homeware aisle just yet), but they increasingly own the discovery and comparison that precede a purchase, even when the final purchase is physical. A choice shaped by an agent still strands or vindicates inventory committed months earlier.

On that basis, agents change the merchant model in three specific ways — each more uncomfortable than the last.

First, they disable the merchant model's rescue mechanism. The markdown has always been how a merchant unwinds a bad bet. A discount presented to a human shopper functions as new information — oh, that's cheap now — that pulls demand forward. An agent doesn't experience a markdown that way. It already knows the real price and every substitute across the market, so a late discount rarely discovers buyers it hadn't already found. It mostly concedes margin on a sale the agent would have arbitrated anyway. The safety valve loses pressure at exactly the point demand becomes harder to judge.

Second, agents concentrate demand rather than spread it. Optimisation converges. Asked to find the best option against a set of criteria, agents return a short list, and repeated across millions of shoppers that behaviour thins the long tail and fattens a handful of winners. The merchant model's core hedge has always been the assortment — hold many lines, and the hits smooth out the misses. Agent-driven concentration works directly against that hedge. Outcomes get more all-or-nothing, which is precisely what a broad assortment was built to absorb and now can't.

Third, agents redistribute the information that justified the merchant's role. A central buyer's authority rested partly on privileged access to demand data: point-of-sale figures, footfall, basket composition. When agents mediate discovery, the entity with clean structured product data and a direct relationship to the agent layer often sees demand form first — and that entity may well be the brand, not the retailer's buying desk.

Taken together, these push in one direction. Forecasting gets harder precisely when the penalty for forecasting badly is rising. Whoever owns the inventory absorbs that shock first — and a seller running a store network, holding stock committed months ago with no ability to reprice overnight, is the most exposed of all.

The shift is often framed as pure opportunity for brands — gain physical presence without your own stores, control product presentation, read demand signals directly, protect your equity from a retailer's markdown instincts.

But the same arrangement that hands brands control also hands them the risk that retailers are increasingly keen to shed. A store-in-store deal that looks like distribution access is also a transfer of forecasting exposure — from a party that wanted rid of it to a party that's agreed to take it on. Being nearest to the demand signal cuts both ways.

There's a further complication. The platform bargain sells brands access to a retailer's traffic — but agents act on the discovery that generates much of that traffic. If shoppers increasingly arrive at products through an agent rather than by browsing a retailer's environment, a brand may be accepting inventory risk in exchange for access to an asset that is itself being rerouted. The value of that traffic should be judged against how much of it agents are likely to intermediate, not against how much the retailer commands today.

The century-old habit has been to treat every retail disruption as a prompt to find a better format. The harder possibility is that the format was never the point.

If AI agents are making demand faster and less forgiving than the merchant model was built to handle — in whatever channel a brand chooses to sell — the decision that matters is no longer where a brand sells, but how much of the risk of being wrong it's willing to own once it gets there.

The questions worth sitting with: in which categories is forecast error most expensive, and therefore most dangerous to absorb? That answer isn't uniform — categories where agents can objectively compare specs and prices (electronics, appliances, functional gear) are where demand will concentrate fastest and forecasting exposure will bite hardest. Categories driven by identity, aesthetics, or impulse are more insulated, for now, because agents are worse at optimising for "how does this make me feel." The brand that knows which side of that line its products sit on has a head start over the one that hasn't thought about it.

Where does the brand genuinely know its demand better than a retailer would, and where is it simply being handed a liability dressed as autonomy? What must the retailer still guarantee — in traffic, in standards, in data sharing — before the trade is worth taking?

None of these have clean answers. But "physical retail is dying" isn't one of them, either.

📅 Agenda Hero Magic - Creates AI-generated calendar events

🪄 InMagic - Uncover your personality based on your Instagram profile

💬 Dubbing Studio by ElevenLabs - Precise control over translating videos for global audiences

📧 Klaviyo AI - Create and understand campaigns faster with AI segments, forms, insights and more.

😍It was great, all I needed

🙂Good, not great

🙄It sucked