AI Shopping Agents Hit Two Walls on My Real List

When I was a kid, my mum used to hand me a list and send me to the shop. Veggies. Bread. The basics. The list was for me, not her. She already knew what things cost. I could buy outside that list if I wanted to, a nice thennilavu or a biscuit pack. I would walk, buy exactly what was written, and come back. Job done.

I was her agent.

Nobody thought of it that way then. You send the kid, the kid comes back with the stuff. Trust, a list, exact money, a short walk. That is the most ordinary thing in the world. It is also exactly what every household, office, school, and workshop still runs on repeat. Weekly groceries. Monthly toner. Term-start school books. Workshop consumables. Same loop, different scale. Someone makes the list. Someone runs it. That loop is what I handed to AI.

I should tell you where I am, because it matters. I am in New Zealand. I buy from Jaycar, PBTech, marvle3d.co.nz, and a handful of international niche suppliers no global platform has ever indexed. I AM the long-tail problem.

The List

I was setting up a small resin printing operation. The list: a resin printer, FEP film, digital calipers, nitrile gloves, a respirator rated for fumes, a curing turntable, assorted smaller bits. Items spread across six stores. Some big retail. Some specialists that exist only in a small country at the bottom of the Pacific. These are not Amazon sub-brands. They are tiny shops with hand-built websites, run by people who know their product better than their SEO.

I wrote the list plainly. The way you would write any shopping list. Then instead of opening twelve tabs and doing the rounds myself, I gave it to an AI and said: figure it out.

What Happened

For the big retail pages, it worked cleanly. In, out, prices, links. Recognisably the same shape as my mum's errand, just faster and across more shops at once.

For the smaller stores, the NZ specialists, the international niche suppliers, it struggled. Slowed down. Made mistakes. Sometimes stopped entirely. Not because the agent did not know the store existed. It can load any URL. The problem was those sites were built for a human clicking slowly. No structured product data. Checkout flows that break on automation. Variants nested inside variants with nothing underneath for a machine to grab.

The local NZ stores were exactly where it hit the wall. And those are the stores I actually need. The global platforms do not stock a resin printer compatible with the FEP film from a specialist in Auckland. That is the long-tail problem made personal.

Three Tools, Same List

Tool one: fast on big retail. When it broke, it broke quietly. Wrong product. Outdated price. A polite "I could not finish this" buried near the end of a long task. You would not notice unless you checked every item.

Tool two: more cautious. It asked before acting, confirmed before clicking. Better for oversight. Slower for speed. If you want control, this is the shape. If you want to hand off the list and walk away, it is not there yet.

Tool three: not really an agent. A search-and-compare layer. Fastest for research by a clear margin. It does not add anything to a cart. It does not buy. For finding and comparing, it was the most useful of the three. For doing, it does nothing.

None did the whole job. All did some of it.

The Part Nobody Is Saying Out Loud

One of the biggest AI companies in the world launched buy-inside-chat last year. Major retailers signed up. The pitch was that AI would handle the entire loop, finding through to payment, without you ever leaving the conversation.

Six months later: stale stock data, merchant onboarding harder than expected. They stepped back. Now the agent hands you to the retailer's own site to pay.

That is the receipt for what I found on my own list. The split, AI finds and you buy, is not a temporary phase. It might be how this works for a long while.

The Two Walls

Wall one is the long tail of suppliers. Big stores worked. Small ones did not. Most of commerce is not on major platforms. Most things people actually need come from somewhere specific and small.

Wall two is the recurring list. The one-off purchase is a small problem. The interesting version is the list that comes back. A household's Saturday groceries. An office's monthly toner and paper. A school's term-start books. A workshop's consumables. Someone runs each of these lists manually: same tabs, same suppliers, same comparisons, every week or month or term. The shape an agent could actually fill is not "buy me a printer once." It is "this list again, best prices this week, across these six suppliers, ready for me to approve." We are closer to that than people realise. Further from it than the marketing suggests.

The Flower Nobody Searches For

Somewhere in the world a particular flower grows. It blooms. It dies. The value it carries passes back into the ground. Nobody outside that place knows it exists. No supply chain reaches it. Not invisible because it lacks value. Invisible because the system was never built to find it.

That is the long tail at its most extreme. And it is also most of the world. A maker in a remote region with no web presence. A grower with a product nobody has ever searched for. A specialist shop in a small country. The constraint is not software alone. It is connectivity and structure. When connectivity reaches everywhere and agents go looking instead of waiting for a search query, those things enter the system for the first time.

That is the progress worth watching. Not checkout flows. The slow expansion of who gets to be in the system at all.

What Actually Changed

When I was my mum's agent, I went to one shop. The list was simple. The trust was complete. She knew what things cost before she handed me the money.

The AI goes to many shops at once. The trust is partial. I trust it to find. I do not yet trust it to pay. But the shape is the same. Someone sends a list. Something goes. Something comes back.

The agent was never the new idea. Kids to shops. Office juniors to suppliers. Procurement teams to vendors. What changed is who can be one. For now, it handles more than you expect and less than the marketing suggests. That is not a failure. For anyone running the same procurement list every month, the discovery and comparison work alone is real time back. Where it stops is where the interesting problems start. And I am not just observing. I am going to keep building toward the missing piece.