『How to Actually Use AI in Your eCommerce Business』のカバーアート

How to Actually Use AI in Your eCommerce Business

How to Actually Use AI in Your eCommerce Business

無料で聴く

ポッドキャストの詳細を見る
概要
 AI will not run your ecommerce store on autopilot. Cyril Golub joins Matt Edmundson to explain how to actually put it to work, as a supervised assistant you give a clear goal, not a magician you hand the keys to. It is a practical, often funny conversation about where AI genuinely helps online sellers and where it just burns tokens.SummaryCyril Golub has been building ecommerce software since 2003, so he watches the AI wave with the eye of someone who has seen a few hype cycles come and go. In this episode he and Matt dig into what AI changes for online sellers and what it quietly leaves exactly the same. Cyril's thesis is that "ecommerce" is turning into "AI commerce", where by early 2027 both buyers and sellers have their own agents, and "how do I start selling online?" becomes a single sentence typed to an agent.Along the way they get practical. Most AI startups are building for buyers, leaving the seller side wide open. Marketplaces own your customer data, and on Amazon a buyer is one click from a competitor. Matching the right model to the job matters more than reaching for the biggest one every time. And competitor monitoring at scale lets you act alongside rivals rather than react to them.Key points covered in this episode:Who's building AI for sellers, not just buyers (10:25)Who really owns your customer data (17:41)Which AI model you should actually use (26:55)Using AI to monitor your competitors at scale (31:43)Who's Building AI for Sellers, Not Just Buyers (10:25)Cyril points out something surprising about where the money is going. Roughly three quarters of AI-native ecommerce startups are building buy-side tools, helping shoppers find bargains and streamline their buying. The seller side, by contrast, is far less crowded, which he finds odd given how much easier it usually is to sell to businesses than to chase millions of individual consumers.That gap raises a harder question for any small founder building in the space. If a tool genuinely helps sellers increase their margin, who gets to build it? The giant sitting on terabytes of data, a mid-size player, or a small third party with neither the data nor the audience?"The main question that early-stage investors are asking to startup founders, what if, then put the name of the huge company. What if Amazon will develop it on their own, directly in-house, and then roll out to all of their hundreds of thousands of customers?" — Cyril GolubMatt's hope is that the answer turns out to be the small operator in the garage rather than another platform giant, which is exactly the kind of story this show roots for.Who Really Owns Your Customer Data (17:41)The storefront, Cyril argues, is a habit inherited from the offline world, the digital echo of goods on a shelf. But not every visitor wants to browse a shelf. Some arrive with a clear intention and want to reach checkout in two or three clicks. Others arrive to explore, unsure what they want yet. The problem is that most sellers know almost nothing about which is which, often only the channel or the ad that brought someone in, and rarely anything that ties through to the actual purchase.That blindness is the real cost of not owning your data, and on a marketplace it gets worse."When we are talking about Amazon sales, your buyer is just one click away from buying from your competitor. That's a huge problem." — Cyril GolubHis practical steer is to pull as much data as you can out of whatever platform you sell on, through the richest API access available, so you can connect your own agent and ask it for actionable insight rather than just advice.Which AI Model You Should Actually Use (26:55)This is where the conversation gets genuinely useful for anyone watching their AI bill. Cyril's view is that you should match the model to the job, the same way you would match a hire to a task."It might be like if you are hiring a PhD to just calculate simple numbers in your Microsoft Excel sheet. You must choose the exact type of model that is quite enough, but not more, for your particular goal." — Cyril GolubHe gives a concrete number. Connecting Claude to an Amazon store through the SP-API to process an inventory of around 2,000 SKUs costs roughly 4 million tokens, mostly input. That is far too simple a job to hand a frontier model. His answer is to think in layers:A deterministic layer that prepares the data quickly. If a task is fully deterministic, it is just software, no AI needed.A cheap supervisor model to decide which parts of the data actually matter for the goal at hand.A frontier model reserved only for the work that genuinely needs judgement.Matt notes he is seeing this layered thinking take hold more widely as people wake up to what AI actually costs. Cyril adds a warning about context too, in that the right answer moves over time. What was true a week ago, around a season, a sale or a competitor's campaign, can be wrong the next week.Using AI to Monitor Your Competitors at Scale (31:43)...
adbl_web_anon_alc_button_suppression_t1
まだレビューはありません