Everyone's shipping AI features, few are solving AI problems. Why variable cost, freebies and lock-in should worry you, and three things to do this week.
The race to "look AI" is masking a much bigger conversation about value, variable cost, and lock-in. Here's why it should worry you - and what to do about it.
Welcome to issue #8 of The Monday Trading After-Party newsletter.
I want to talk about something that's been bugging me for a few weeks. And it came into focus on a call with a friend in the industry last week, mid-coffee.
A retailer they know had ripped out a perfectly good product recommendations vendor. One who knew the products. Knew the market. Continuously iterated. Put their nuts on the block every quarter to prove the uplift was real.
And replaced it with an internally built AI version.
Bosh. Done.
"There you go. AI. You're welcome."
Then the team moved on to the next thing. Of course they did.
The pattern is everywhere
Look around. It's not one retailer. It's the dominant pattern of the last eighteen months.
Amazon's got Rufus/Alexa. Walmart's got Sparky. Every major retailer is racing to bolt a shiny AI feature onto the homepage so they can tick the "we're doing AI" box at the next board update.
I get it. The pressure is genuinely real. ChatGPT changed the conversation overnight. The big platforms are setting a new bar. The board wants an AI story. And nobody wants to be the retailer left behind.
So, we ship something. We launch the assistant. We put AI in the press release.
Tick.
Here's the thing, though. Most of these features aren't very good.
Don't take my word for it. Try them.
The washing machine test
I went looking for a washing machine recently, using a specific need – a family with a newborn baby. I wanted guidance and genuine help.
The assistant gave me three machines.
No reasoning. No, "Is this for the whole family or just for you?" No, "is it because of nappies, or allergies, or because you live in a flat with shared walls?" Just three machines.
When I pushed for more, it gave me three more. The same machine in different drum sizes.
That's not a shopping conversation. That's a glorified product carousel with a friendly avatar on top.
If I walked into a real shop and asked the assistant on the floor, "Is that really all you've got?" - and they pointed me at the same machine in three sizes - I'd quietly question whether they wanted to keep their job.
That's where a lot of these AI features sit right now.
This is what’s actually going wrong
I don't want to criticise other retailers. They're under enormous pressure, and they're trying. But I think we're misdiagnosing what's happening, and that's worth being honest about.
In my experience, this is the unanswered question:
How much did that AI feature actually make us?
And if the answer is "not much" - is that because the idea was bad? Or because we shipped a v0.1 and never went back to it?
That second question is the one that should keep eCommerce Directors up at night.
Most internally built AI features are launched once and quietly abandoned. They don't get iterated. They don't get measured properly. They don't have a roadmap. And the reason is structural, not lazy.
It's because they're treated as "free."
The freebie trap
Let me explain what I mean.
When the in-house UX person says, "We should put an AI assistant here", the idea is treated as a free improvement. They're already on the payroll. The team's already there. The hosting is already budgeted. So, we go and build it.
And almost no one goes back afterwards and asks, "Did it work?” How much revenue did it add? What's v2? Where's the roadmap?
Now compare that to a paid vendor feature. Anything we pay for gets scrutinised to within an inch of its life. ROI. Conversion uplift. Revenue contribution. "Is it pulling its weight?"
Two completely different standards. One feature is a freebie that "looks like progress." The other is a contract that must earn its keep month after month.
What I'm seeing across the market is retailers replacing paid features that put real money in the register with free features that contribute next to nothing. And calling it “transformation.”
That's not transformation. It's a quiet downgrade with a press release on top.
Intelligence is now a variable cost
Intelligence is now a variable cost
There's another piece of this puzzle, and I hardly hear anyone talking about it. And it's an important one.
Intelligence used to be a fixed cost. Now it's a variable cost.
Time was when you built an intelligent system, the price was fixed. You bought the software. You ran the software. Done. Whether one customer used it or one million, the cost didn't really move.
There's another piece of this puzzle, and I hardly hear anyone talking about it. And it's an important one.
Intelligence used to be a fixed cost. Now it's a variable cost.
Time was when you built an intelligent system, the price was fixed. You bought the software. You ran the software. Done. Whether one customer used it or one million, the cost didn't really move.
That's gone.
Every clever decision an AI makes on your site now has a cost attached to it. Every prompt. Every token. Every API call. You're paying every single time you want to be intelligent.
Multiply that across millions of sessions, and the maths gets uncomfortable fast.
And that's before we even get to the economics of product enrichment. If you want every product in your catalogue to be discoverable across every channel - your site, Google, ChatGPT, Rufus, agentic shoppers - you've got to enrich every single product. Properly. With intent-rich data.
That doesn't scale linearly down to a tiny rounding error. It scales by the count of your range.
Range growth used to be cheap. Push 10,000 new SKUs to the database. Bish bash bosh, on the site by Tuesday.
It's not cheap anymore. Every product you add now carries an enrichment cost - and an ongoing intelligence cost every time someone interacts with it.
Most retailers haven't built this into their unit economics yet. They will. Whether they want to or not.
And then there's the lock-in problem
I read a story recently that stuck with me.
An Illinois farmer called Jake Lee. A single faulty sensor on his combine harvester at the peak of harvest. The combine wouldn't restart. He couldn't fix it himself. He had to wait for an authorised John Deere technician to drive out, plug in the diagnostic tool, identify the faulty part, replace it, and reset the code.
Meanwhile, more than a million dollars of equipment was sitting idle in the field, with rain on the way and a crop to bring in.
That story stuck with me because we're doing the same thing in commerce. Quietly.
We're locking ourselves into single LLM providers. One model. One vendor. One pricing schedule that they control unilaterally.
If your hosting goes down, you fail over to another data centre. Standard practice. Every senior engineer plans for it.
If your AI provider goes down, or quietly doubles their pricing, or changes their terms of service, or deprecates the model your features depend on?
Most retailers don't have a Plan B. They've never even drawn it on a whiteboard.
Pulling it all together
So, zoom out for a second.
Across the industry, retailers are shipping AI features that aren't great, aren't being iterated on, aren't being measured, aren't going to be maintained, introduce brand-new variable costs into businesses with already-tight margins, and lock them into a single vendor with no fallback.
And the press release says, "we've transformed."
I'm not anti-AI. We run a company that uses AI! AI is the most exciting technology of my career. But this is the same conversation we had ten years ago about big data. And five years ago, about machine learning. And before that, about digital transformation.
The technology changes. The wrong question stays the same.
Wrong question: "How can we use AI?"
Right question: "What are the actual problems in our business that are costing us growth, frustrating our customers, or eating our margin? And which of those are now newly solvable with AI?"
That's a completely different starting point. It leads to completely different conversations, RFPs, and outcomes.
It also stops you from replacing positively contributing paid features with non-contributing free ones.
Three things to do this week if you're an eCommerce Director
I want to leave you with three actions. Easy to type. Harder to do. But the retailers who do them will pull away from those who don't.
1. Audit every AI feature you've launched in the last 18 months.
Not "is it live?" Not "does it work?" The real questions are: how much revenue has it added? When was it last iterated? Who owns the roadmap? If you can't answer any of those, the feature is a candidate for either serious investment or a quiet retirement. Don't let "we have an AI assistant" be the end of the sentence - make it the start of one.
2. Start with the business problem. Not the tech.
Before approving the next AI initiative, list the five things that are costing you money right now. Frustrated exits. Dark aisles. Failed search. PDP bounce. Range that never gets seen. Pick the biggest one. Then ask which AI capability solves it - and how you'd measure success. If you can't define the metric in advance, you're shipping a feature for the press release, not the register.
3. Start tracking AI as a variable cost. Today.
Not "how much do we spend on AI vendors?" - that's the easy bit. The harder questions: what's it costing per session? Per intelligent interaction? Per enriched product? Per agentic query coming in from ChatGPT? Build that line into your unit economics now, while the volumes are still small. Because the day will come when somebody else asks first, and the conversation with the CFO is much easier when you've already got the numbers.
We're at a point where everyone in retail is busy launching.
Almost no one is busy compounding.
Be the latter.
See you in two weeks.
Warren
Want to talk about what intent-led commerce looks like for your business?
FoundIt! helps retailers make every product findable by reshaping the experience around customer intent - not your catalogue. Get in touch at foundit.com or DM me for a chat.