Which AI shopping features actually worked in 2026

2026-07-16

Major retailers have invested a lot of time and money in bringing AI assistants to shoppers. ​ Let’s look at a number that should make any board think twice.

• Ecommerce
• AI
• Retail Strategy

The big retailers have spent two years and a lot of money putting AI in front of shoppers. Amazon, Walmart, Klarna, OpenAI and plenty of others rolled out chat assistants, automated checkouts, in-app helpers and AI recommendations. It has all been live long enough now that we can stop reading the launch posts and start reading the receipts. And the receipts, good and bad, keep pointing at the same thing.

My read, after watching a fair amount of it land and a fair amount of it flop, is that AI belongs in the shopping journey roughly the way a good sales assistant does. Brilliant at some jobs, useless at others, and the trouble starts the minute you forget which is which. The retailers doing well with it picked a lane. The ones nursing a write-off tried to hand it everything and hoped. Here is how that played out.

The feature everyone announced, and shoppers did not use

The clearest flop of the cycle was in-chat checkout. OpenAI launched Instant Checkout inside ChatGPT in September 2025, so you could buy something without ever leaving the conversation. On paper, the plan was enormous. Over a million Shopify merchants were lined up to join, sitting behind ChatGPT's 700 million weekly users. This was going to be the widest shopping counter ever built, and the whole pitch was that the sale itself would happen inside the chat.

It didn't happen. Come February 2026, Forrester went looking and could count maybe 30 Shopify merchants actually switched on, out of the million-odd that were meant to be on their way. By spring, the reporting had caught up, and OpenAI was quietly rowing back its shopping ambitions. No scandal sat behind it, and nothing had broken. The problem was money, or the lack of it: paying inside the chat converted about three times worse than clicking through to the shop's own site. It turns out people will happily poke around in the assistant and then go and pay where they always pay.

You can see why, if you think about where a shopper actually feels safe. It is on the site they already know. Their card is saved there, the returns policy is exactly where they expect it, the little padlock sits in the corner doing its quiet reassurance. Send them into a chat window they have never used, right at the moment of paying, and you have planted a doubt at the worst possible second. Doubt there is just a lost sale you haven't counted yet. So the whole industry has drifted towards the obvious compromise. AI helps you find the thing. You still pay on the retailer's site.

Klarna, and the difference between the average and the tail

Klarna is the other one everybody cites, and it played out in support rather than on the storefront. February 2024, the company put out a number that did the rounds for months: its OpenAI assistant had got through 2.3 million conversations in a single month, about what 700 agents would handle. It was closing tickets in under two minutes against eleven for a person, and it was set to save tens of millions. Every tech outlet ran the same line, that the robots had finally come for the call centre.

Fast forward fifteen months and the tune had changed. In May 2025, the CEO sat down with Bloomberg and admitted they had cut too deep on people, the service had suffered for it, and they were hiring humans again. This bit gets told lazily, so let me be precise. Klarna did not rip the AI out and rebuild a call centre. It narrowed where the AI sits. Order-status and payment-date questions, the boring high-volume ones, still go to the bot. The messy conversations went back to people: chargebacks, fraud, someone who cannot make a payment and is quietly panicking about it.

Why this should matter to the rest of us has nothing to do with Klarna in particular. It is about which numbers they were watching. Resolution rate, handle time, the usual dashboard fare. All of them averages. And the average buried the one thing that actually counted, which is that support tickets are not all worth the same. Most are dull. Where is my parcel, when does my payment come out, can I still return this. Enormous volume, nothing really at stake, and the bot clears them without breaking sweat. Then there is the small pile, maybe one in twenty, that carries nearly all the risk and every scrap of the regulatory danger, and those are the ones where you want a person who can read the room. Klarna built for the average ticket and got caught out by the rare one. If you take a single thing from the whole saga, take this: work out what a wrong answer costs you on each kind of query before you automate anything, then let the machine loose only where getting it wrong is cheap.

The gap between trying it once and relying on it

Not everything here is a cautionary tale. Some of the adoption numbers are genuinely big, and this is where you have to read slowly, because the headline figures are softer than they look. Amazon says more than 300 million people used its Rufus assistant, now rolled into Alexa for Shopping, across 2025, with monthly users and engagement well up. It also says shoppers who use the assistant are more than 60% likelier to buy on that trip.

Two things pull that back down to earth. One, trying a thing once is not the same as leaning on it. By early 2026, sellers were putting Rufus at under 3% of Amazon shopping sessions, and a Bizrate Insights survey in the middle of the year found the most common reaction to assistants like Rufus and Walmart's Sparky was people who had not touched them and had no interest in doing so. Plenty of awareness. Not much habit.

Two, and this is the one that quietly wrecks a business case: that 60% stat is cheating a little. Look at who is actually in it. The only people counted are the ones who chose to open the assistant, and those people were already halfway to the till before they typed a word. The assistant did not talk them into buying. It walked in on a decision that was more or less made. Some of the uplift is real, sure. It is nowhere near the full 60%. And the same goes for every vendor slide showing AI shoppers converting three or four times better than everyone else. Read those as a hint about the direction things are moving, not a figure you would drop into a forecast and then have to defend.

Where AI genuinely earns its place

None of this makes the money wasted. It means the payoff sits in one particular stretch of the journey. AI is at its best helping people find and compare things. Shoppers are perfectly happy to let it lay out the options, weigh up specs, build a shortlist and answer the questions they would otherwise chase across a dozen browser tabs. That is real friction, genuinely taken out.

The surveys back this up, and they are almost funny in how split they are. Checkout.com ran a report in June 2026 where a third of people reckoned at least a tenth of their shopping would go through AI inside a year. In the same breath, 24% said they would never let AI actually buy anything for them, and 27% said there was not a single company they trusted to run a shopping agent on their behalf. Other studies this year land in the same spot. Ask people whether AI can compare products for them and roughly two in three shrug and say fine. Ask whether it can place the order and only a handful say yes. Comfort holds right up until the money moves, then people clam up, which is the OpenAI checkout story all over again, only this time in a survey rather than on a live site.

The businesses making this pay are the ones splitting the work. Walmart, after its own flat experience with checkout inside ChatGPT, now runs its Sparky assistant in there purely for discovery, and reckons it pulls in about twice as many new customers as ordinary search. Gap tied up with Google's Gemini so it could offer AI help without handing over control of the customer. Amazon's assistant does its best work inside Amazon's own walls, where it has the full catalogue and your history to lean on. Same idea every time. Use AI to walk a warmed-up shopper to the decision, and keep the decision and the payment on your own turf.

How to read this if you run an ecommerce business

So the instruction is not to adopt AI, and it is not to sit on your hands. It is put the right tool on the right job, and be honest with yourself about which job you are handing over.

On the support side, that means giving the bot the dull, high-traffic questions where a duff answer costs you nothing, and keeping a human, with the AI feeding them, on anything involving a dispute or a fed-up customer. And do not just stare at the average, because the average will look perfectly healthy for months while the hard cases quietly rot underneath it. On the shop itself, point the AI at discovery and comparison, tidy your product data so the assistants describe you properly instead of guessing, and leave the checkout where it belongs, on your own site, where people already trust you. One last thing, and Klarna paid for this one. Resist the urge to stand up and announce you have swapped your whole team for AI. Walking that back in public is a genuinely grim way to spend a quarter.

If there is one thing worth carrying out of this year, it is that AI in commerce was never a single switch you flip on or off. It is a box of tools, and like any box of tools, most of them are handy for a specific job and a menace for the wrong one. The retailers who come out ahead will not be the ones who bolted on the most of it. They will be the ones who stayed honest about what it is good at, blunt about what it is bad at, and had the discipline to stop before they broke something that was working.