|
Well, hello, and welcome to this week’s edition of the eCommercer.
I think most of us have a slightly odd relationship with AI right now (especially after yesterday’s nightmare with the PHP bug). We either expect it to read our minds and run the whole business while we sleep, or we get so swamped by the options that we freeze and do nothing. I’ve managed both in the same afternoon.
Match the model to the job
My guest is Cyril Golub, who’s been building ecommerce software since 2003 (one year longer than me, which makes him an even bigger dinosaur). These days he’s building Jinnify, an AI operations platform for Amazon sellers, so he spends his days wiring AI into real stores with real data.
He talks about how we shouldn’t reach for the most powerful model for every job. As Cyril put it, you wouldn’t hire a PhD to add up simple numbers in an Excel sheet. Same with AI.
He breaks the work into layers.
- The genuinely helpful, repeatable stuff is just software, a plain deterministic algorithm, and you don’t need AI anywhere near it.
- Above that sits a cheap, simple model whose only job is to decide which bits of your data actually matter for what you’re trying to do.
- And only at the top, where real judgement is needed, do you bring in a frontier model.
|
Take costs, for example. Connect a top-end model straight to your Amazon store to read an inventory of 2,000 products, Cyril reckons, and you’ll burn through something like 4 million tokens, mostly just reading the data in. That’s a very expensive way to do a job a much smaller, cheaper tool could have handled.
It’s not a wizard
If you’re a roughly million-pound store, you know how to get Claude to fix a few things on your Shopify site, and you’re staring at the whole world of AI feeling a bit overwhelmed, how should you actually think about it?
I loved Cyril’s answer.
| | “Think about AI like your assistant that must be supervised and clearly set with the goal and the expected outcome. Don’t think it is a magician or a wizard that will do everything without you and your supervision.”
Cyril Golub |
So two jobs for you, not for the model. Give it a clear goal. Then supervise the outcome. That’s it. The moment you treat it as a colleague you’re managing rather than a genie you’re wishing at, the overwhelm drops away and you start getting useful work out of it.
Shiny Object Syndrome
I’ve definitely got shiny object syndrome like every founder I know, and my instinct is to throw the biggest, cleverest model at everything, whether the task needs it or not.
I do it with this very podcast. I take the transcript of each episode, hand it to Claude and ask how it applies to my own businesses. That’s a good use of a smart model. But I’ve also been doing the daft stuff the same way, the tidying and the sorting and the list-making that a far simpler tool would do for a fraction of the cost.
Cyril’s layers gave me a cleaner way to think about it. Not “AI or not AI”, but “which bit of this actually needs a brain, and which bit just needs a bit of plumbing”. That feels like a much saner way to run things, especially as the bills start to land.
The full conversation with Cyril is live now. We get into why he thinks ecommerce becomes “AI commerce” by 2027, who really owns your customer data once you’re selling on a marketplace, and how AI could watch a hundred of your competitors at once and match their next move before their first email even lands.
| 
