Most retail AI assistants answer before they ask. Why agents built for selling fail customers who came to buy, and three shifts to make them assist.
"I want something to cover scars."
Type that into a retailer's new AI assistant and watch what comes back.
No questions. Not one. A warm paragraph of preamble, five products, and a reminder that you're £4.05 away from free delivery.
You didn't ask about delivery. You didn't ask for five products. You asked for help.
That is premature elaboration.
I was taught the phrase in sales training years ago, and I've never found a better one. It describes the urge to tell somebody everything before you know anything. You catch a sliver of a brief, and you're off – the range, the features, the offer, the lot. Good salespeople learn to sit on their hands. Most AI assistants never learned.
It's Mr Search Engine, and he's wearing a mask
Pull the mask off most retail agents and the same face is underneath every time. A query, hooked up to a search index, with a thin layer of agency wrapped round the outside.
It's a Scooby-Doo ending.
That isn't a criticism of the people who built them. It's a description of what they are. We called them agents, and "agent" sounds clever, so we all quietly decided they were cleverer than they actually are. Me included. That word has done an enormous amount of heavy lifting this year.
Try the test I keep coming back to. Take the exact behaviour of your AI assistant and put it on your shop floor.
A customer says she wants to cover some scars. Your colleague asks her nothing, walks her two hundred yards down the aisle, gestures at a shelf and says, "try these."
You'd have a quiet word, wouldn't you?
We've accepted online what we would never tolerate in a store. And then we've called it world-class.
Why everybody did the same thing at the same time
It's worth saying plainly: nobody set out to build something thin.
The clock started in January. Google went to NRF and launched the Universal Commerce Protocol, an open standard for agentic commerce running from discovery all the way through to post-purchase. The room went up. Boards read the coverage. And the same question landed on desks all over the industry – what are we doing about agents?
Retailers move slowly, if they move at all. So, it's taken most of a year for that panic to turn into anything at all. What it turned into, mostly, is version 0.5.
I know of one chief digital officer – since departed, at a retailer you'd recognise – who announced in the autumn that he wanted an agent on the site by Christmas. Not because of a business case. There wasn't one. Because he wanted an agent on the site by Christmas.
So, they built one. It really isn't hard to put together an agent-assisted search tool. The harder question is whether you should have bothered.
Spend like that doesn't appear on the balance sheet as innovation. It appears as sunk cost.
That's the symptom. The cause is knowledge.
Every one of these things has the same hole in it. Not the model. The model is fine. The knowledge.
Let’s go back to the scars.
Somebody who knows the foundation category knows that "scars" is a correction target, sitting alongside acne marks, dark spots and hyperpigmentation. They know that if you're covering a scar, you almost certainly want full coverage. They know that if the scar is on your body, you're probably looking at body foundation too. They know that acne scarring means the formula can't clog pores – Haus Labs make a point of their Triclone concealer being non-comedogenic, and that isn't a marketing detail, it's the answer to the question.
They know that somebody covering a tattoo is going to a wedding, onto a stage, into an interview, or into the water. And if it's the water, waterproof isn't a nice-to-have. It's the entire brief.
They also know that when a customer asks for "Lady Gaga's concealer," she means Haus Labs. Because who remembers the brand? You remember the person.
None of that lives in your product feed. All of it lives in the relationships between things. That is what a knowledge graph actually is – not a taxonomy, not a tidier version of your category tree, but a map of what matters to a customer and what connects to what.
Without it, your agent cannot hold a conversation. It can only search. And searching is precisely what it does, at speed, before you've finished your sentence.
We built for selling. They arrived to buy.
April Dunford, who, in my opinion, is the sharpest positioning thinker working today, opens her talks by asking whether selling is hard. Everyone nods. Then she makes the case that buying is hard too.
Her example is a toilet. She'd bought an old house that needed a bathroom, walked into a store, and found a wall of toilets that all looked identical, carried wildly different price tags, and came loaded with features that meant nothing to her. The first person she spoke to was no help whatsoever.
The person who eventually got the sale didn't pitch her. He taught her how to decide.
That's the pivot, and honestly, it's the whole article.
We have spent twenty-five years optimising the selling experience. Almost nobody has optimised the buying experience. We keep building better ways to present. Customers need better ways to choose.
An agent that lists is still selling. An agent that asks is helping somebody buy.
The knowledge that walked out the door
I used to go to John Lewis at Brent Cross for hi-fi. There was a chap there who knew his stuff. Not the sort who glances at the shelf label and reads it back to you – "yes, that one's got two speakers." He'd watched the televisions. He'd sat in the room with them. He'd tell you this one is selling well, and that one is better, and the reason people love it is the deep blacks.
I bought from him. And then I went back, repeatedly, because of what he knew and because he met me where I was.
Richer Sounds is fifty-odd shops built on exactly that principle, and they say so themselves – walk in and you'll be met by somebody who knows the kit and knows how to make it work in a real house. Snow+Rock was founded by people who ski. Back when they owned Cycle Surgery and Runners Need, the whole anti-Amazon position came down to one thing: the people who work here care about the thing you're buying.
They lived the domain.
Now think about the woman who ran the beauty counter for thirty years. Every formula, every undertone, every regular customer. She retired. There was a card and a cake.
How much knowledge left the building that afternoon?
That's the line nobody puts in the accounts. Retailers spent decades investing in expertise, and all of it sat inside people, and when the people go, it goes with them. Nothing replenishes it. What replaces it is a sixteen-year-old on a Saturday shift with a playbook, doing their best.
An intent-led knowledge graph is that expertise, written down. Structured, connected, and available to every customer simultaneously. It doesn't retire. It doesn't leave.
The bit in the middle
Which brings me to what's actually missing.
We named the category wrong, and I'll put my hand up because we all did it.
"Conversational search."
We took something that wanted to be a conversation and bolted it onto a search. We should have just built the conversation.
A conversation has a middle. Warm up, ask, listen, narrow, and only then present. There's a less polite word for that middle bit and, given the headline, I'll use it once and move swiftly on: foreplay. Skip it, and you get precisely the outcome the headline describes.
Years ago, a man sold me a photocopier. He listened. He wrote everything down. We talked it through properly. Then he went out to his car and came back with the thing, and it was the right thing, and I bought it.
To this day I have no idea whether that was the only photocopier in his boot. I'll never know. I bought it because he'd understood the job I needed doing.
Make these three shifts, if you want your assistant to actually assist
1. Stop selling. Start helping people buy. This is a mindset change before it's a technology change. Ask what a genuinely helpful guide looks like on your site – the Brent Cross bloke, the toilet salesman, the woman on the beauty counter. Then look honestly at what you've currently got and mark the gap.
2. Rebuild the middle. Before you present anything, work out whether you need to clarify, narrow or simply ask. One good question beats five products every single time. And if the customer would rather not explain, let them skip it and carry on – the point is that you offered.
3. Curate. Don't crown. "Here's my top pick" is the biggest cop-out in retail. At least Google Shopping gave me hundreds of products and let me browse to my heart's content. Give people segments instead. You want something for walking the dog. You want something for town. You want the one you'll wish you'd bought when it turns cold. Too few options and they assume you haven't tried. Too many and you've simply rebuilt the aisle.
Ask your own agent one question that your best shop-floor colleague could answer in a sentence. Then count how many questions it asks you back.
If the answer is none, you haven't built an assistant. You've built a search box with better manners.
If you'd like to see what this looks like underneath – an actual knowledge graph, an actual category, actual relationships – get in touch, and we'll walk you through it.
Until Tuesday after next.
Warren.