Shopify Built a Copilot, Not an Autopilot, for Every Merchant

Glen Coates arrived at Shopify the long way. He was a computer-science graduate who spent the early part of his career in video games, then took a sharp left in 2008 and moved to San Diego to run U.S. operations for an eco-friendly shopping-bag company, the bring-your-own-bag kind of business. That company sold a lot of bags wholesale, and the wholesale workflow is what gave him the idea for Handshake: first a sales-rep tool for B2B orders, then a storefront for the buyers themselves. After nearly nine years in New York, Handshake had become, in his words, a wholesale-only version of Shopify.

The acquisition closed in 2019. Coates spent about a year on B2B and wholesale, the original point of the deal, and by May of the taping he would be coming up on five years at the company. In 2020, the first year of the pandemic, he was pulled onto a Code Red for checkout. Since the end of that firefight he has led the core product group: the online store, checkout, the back office, the developer platform, and the App Store, including the AI products the hosts want to talk about.

A Code Red, he explains, is not a metaphor. Tobi Lütke emails the entire company that one thing is the number-one priority, and the operational meaning is literal: if the team on that project asks for help, drop what you are doing. In 2020, checkout was failing in three, four, five different ways. At the peak, two or three hundred people scrambled on pieces of the problem for a year. When the fires were out, Lütke asked the more useful question: how did Shopify get to a place where those failures could happen in the first place?

The answer pointed at the org chart. Shopify had twelve to fifteen small, fractured business units; after the reorganization it had three or four, closer to the actual shape of the product, but each unit was much bigger and had to be led differently. Elad Gil notes something unusual about that leadership bench: a very large fraction of Shopify's management team was acquired in as founders. Coates sees the same pattern one level down. Nine or ten people report to him, each running an R&D org of a couple of hundred, and a very high percentage of them are founders too. He traces the culture back to Lütke, who has stopped being shy about aiming the company. Coates paraphrases the CEO's logic: if he is good at one thing, it is being extremely opinionated, and in internet commerce his batting average is high enough that decisiveness is a way to accelerate the company. Former founders recognize that "all bucks stop with you" feeling and use it to push teams past the designed-by-committee freeze that kills a lot of companies.

Gil's second question is the other half of buying companies: what do you do with the tech? Meta left WhatsApp independent for a long time. Google once spent a year and a half on an identity layer while other teams barely shipped. Coates's frame is an engineer's: look at the stack, and do not want duplicates at the same layer. The trap is that you often fail to notice two things are at the same layer until much later.

Checkout at Shopify is a negotiation engine. It takes a cart, applies the store's business rules, and turns that into an agreement for a sale, with taxes, duties, and shipping. Adjacent tools grew up beside it: invoicing, editing an order after it was placed. Those projects usually start tiny. A customer calls and says they forgot a discount code, so someone stands up a project to add a discount to an existing order. Then the same surface learns to edit shipping, then to add a product the shopper forgot. A few years later the team looks up and realizes it has reinvented checkout.

The cost shows up immediately in other people's roadmaps. A team adding international pricing has to build it into checkout, into order editing, and into invoicing. Work that should have taken a month takes six, because the layer underneath has five things where there should be one. Collapsing those five back into one is the correct answer and also, Coates says, like rearranging the foundations of a house while people are still living in it. You cannot make everyone move out.

The real fix is two habits. First, develop extremely good intuition for when you are accidentally creating a duplicate, so you never get into the problem. Second, when you notice you already did, eat the vegetables and collapse the copies, because every team above that layer will otherwise pay the duplication cost in perpetuity. It is not only R&D that pays. Most teams will implement the two or three most important surfaces and skip numbers four and five, so the merchants who depend on those paths hit something that looks the same and behaves differently. They look under the covers and find the mess.

Shopify was an early AI adopter, which Gil chalks up to having a sharp technical founder in the seat. Coates says the last twelve months have been a hunt for where different models actually apply. The company's mission, in his telling, is enabling entrepreneurs: simplify starting and running a business so more people who would otherwise not get over the hump get a chance to try. Historically that meant software that is simple at the start and reveals complexity as the business grows. The graveyard, he notes, is full of companies that tried to serve small business and enterprise at the same time. Being simple and powerful is the hard problem of Shopify.

The old playbook was imperative. Shopify won by giving merchants the best switches and showing those switches at the right moment, but the merchant still had to flip them. AI opens a different possibility: put a driver in the car who knows how all the switches work. Plenty of people are daunted by learning the admin even when the switches are good. They can talk to a copilot and explain what they want done; they just cannot do it themselves. That, Coates says, is an amazing opportunity to increase the number of entrepreneurs who actually get a shot.

The small examples pile up. How many people get stumped writing product descriptions because they are not copywriters, feel embarrassed they cannot describe their own candle, and never hit the big green launch button? Shopify Magic and Sidekick shipped faster than a lot of large companies managed. Sarah Guo asks the question those companies get stuck on: the outputs are non-deterministic, the old quality process does not apply, so how do you decide something is good enough to ship?

Coates's current principle, which he says might change if confidence intervals get high enough, is that every Shopify Magic feature is allowed to propose changes but not to commit them. The model can generate text; it will not save it until the merchant reads it and saves it. It can draft a reply to "what's your shipping policy?" from what it knows about the store, but the merchant still has to hit enter. Human-in-the-loop is the risk control and the training signal at once. Each suggestion produces a three-part loop: accepted clean, accepted with minor edits, or outright rejected. Lütke, COO Kaz Nejatian, and Coates are all fairly risk-hungry in the founder sense; they get bored when things get too safe. The line they will not cross is hurting a merchant's business. They will take risk onto Shopify and keep the human-must-click-save gate in front of the store. A year or two from now, if a merchant accepts suggestions without edits 99% of the time, the product might offer full auto mode. They are not there yet. You can see the hill-climb from here.

Asked what is coolest in Shopify Editions, Coates does not lead with image generation. He leads with foundations, and warns that he is a commerce nerd. The most central data model in Shopify is products, and for ten or fifteen years it barely changed. Getting to roughly 10% of U.S. ecommerce on a fairly simplistic model is, he thinks, amazing. The weakness is obvious once a catalog gets complex: lots of options, colors, and sizes, and Shopify starts to struggle. They are updating the model to support a much larger number of variants.

The reason that rewrite is so hard is a classic platform tax. Large variant counts force APIs from unpaginated to paginated, which is a breaking change for the entire app ecosystem, and breaking changes are hard. Alongside more variants, they have embedded a standard product taxonomy with standard categories and attributes. When a merchant creates a t-shirt, Shopify will auto-recognize the category, then use AI to guess attribute values: green and yellow from the uploaded images, cotton from the description.

That structured metadata helps search on the merchant's own storefront. More important, when the same products go out to Google, Facebook, and Amazon, the extra structure makes them far more discoverable, which Coates says leads to higher ranking and more money. Get the data model right and keep data quality high, and the effects show up up-funnel in the actual business. Editions also ships a first version of AI image editing, so a merchant can replace backgrounds and get product photos to a professional standard without an expensive shoot. A lot of Shopify's job, he says, is helping someone like his mom at home look as professional as the world's largest companies.

Then there is search, which is personal for him. He only recently learned that LBD means little black dress. Today Shopify search is still very literal keyword matching. He has an internal video of himself infuriated in a store he actually shops: he searched "sweater," the catalog only listed "sweatshirt," and he got zero results. "This cannot still be true in 2023," he told the company. Semantic search is the fix. Search LBD even if those words are not in the listing, and it should still work. Search "Christmas themed shoes" and the team showed him red and green shoes coming back. "Something to wear to a wedding" or "something to wear to the beach" used to fail unless the merchant had typed those phrases into descriptions, which they never do. His bar is simple: why should search on a tiny storefront not be as smart as Google?

The work underneath is unglamorous. The team assembles datasets, fine-tunes, and picks embedding models by category, because apparel, fashion, gifts, and homewares do not behave the same. Multimodal models from the last two or three months let them experiment with how much weight to give text descriptions versus images versus taxonomy attributes. They are still working out the mix. Guo notes that vision models mean a lot of information already sitting in product photography can finally become text and join everything else about the merchant. For a buyer, "things to wear to a wedding" actually working is delightful. For a merchant who has run a store for five years with 10,000 products, being asked to backfill five attributes on every SKU is a nightmare. A model that can make a 90% correct guess is how you bootstrap those catalogs into the current moment.

What is still missing from the perfect AI world for Shopify? Coates's answer is the gap everyone building agents has felt. You can get a Sidekick-esque LLM agent to about 75% in ten minutes, and then it is a brutal hill-climb to 95%. The first stretch is catnip for hackers. The next stretch is the real work. "The last 5% I think is the next five years of tech," he says.

He is almost cheerful about hallucination. When people complain that an LLM is hallucinating, his reply is that hallucination is the job: every time it emits a token it is hallucinating, you just like some of the hallucinations more than others. The name Sidekick is the product positioning. It is supposed to feel like a companion on the journey, a coach, an assistant, someone who has maybe seen this movie before. Getting from 75% to 95% is the process they are in.

On Shopify's side they keep improving training sets and collecting more conversations with user feedback to dial the product in. Then every few weeks a new model appears, and they can take the same training set and evals and maybe get a step function. Innovation is happening inside and outside the R&D team at once, and Coates says you never know what each week will bring. Guo adds the outside view: embedding models are getting better, people are building synthetic-data tools and tooling for the entire RAG pipeline, so the pain of going beyond the demo should ease. She is counting on Shopify to stay at the cutting edge. Coates says that is where they are trying to live.

Outside Shopify, Coates keeps coming back to the Rabbit R1 at CES. He is wearing the shirt; he is a Teenage Engineering nerd, and the hardware hyped him. The interesting question is not the gadget. It is what interface an LLM agent actually learns. For Sidekick, does the agent call the Shopify admin API under the hood, or does it reach into the pixels of the web app and click around the way a person would? Rabbit chose the GUI as the interface and trained a model to operate web apps. If that actually works, Coates says the strategic brilliance is instant: they have the world's largest app store, because the world's largest app store is the web.

Guo had put the same question to Rabbit founder Jesse Lyu. His previous company tried to unify actions across APIs; this time he went the other way, on the view that if you live on someone else's interface you do not own your destiny. Coates thinks the idea also rings true from first principles. Sidekick already runs sidecar in the admin, next to the merchant. The name itself says I am here with you, both of us using this Shopify thing to build a business. An agent that clicks the same pixels you click almost smells right. Every time you ship a new button, a model that is smart enough just learns to use it. He has also heard the industry cold water: cool idea, insanely hard to pull off. As usual, he says, this thing changes every week.

The other fight he is watching is the triangle of Perplexity, Google, and ChatGPT. ChatGPT is strong at the LLM interface and not yet amazing at search. Google is amazing at search and has not really figured out the LLM. Perplexity sits in the middle and bets that LLM-native search is the interesting hill. Buried in that tussle is a product question: do people want hard facts, opinions, or a compressed summary of ten results, and how much looseness will they tolerate if the summary hallucinated? Gil has started to see early players degrade, stacking qualifications and safety language, or sending the model out to gather information in ways that either slow the answer or make it worse. People come to an LLM for a good synthesis of messy web data, not a poor one, and not a search process. They want an answer.

Coates has a test he ran over and over while thinking about storefront search and the Shop app. A friend's daughter was turning six, and he wanted a classic children's book. Google Shopping for "classic children's book for six-year-old girl" was an absolute disaster. Amazon returned nonsense. ChatGPT asked for ten great children's books for a six-year-old girl came back with a list he found amazing. Flip the query: Nike Air Force 1s in size 10.5 in a store in New York City, and ChatGPT is terrible. Gil, who once ran search at Twitter after years at Google, argues the LLM world will need the old one-box instinct: detect a navigational commerce query and route it, instead of pretending the model has to do everything. Coates puts a ruler on it. Whether this shoe is a Nike is not a matter of opinion. Whether this book is great for a six-year-old girl is. Good search has to sit correctly at the midpoint of facts and opinions. Even inside commerce, people do not search for clothes the way they search for industrial parts. Guo, who invested in the search company Neeva and thinks highly of the Perplexity team, adds that Google's distribution chokehold is real, but technology and user expectations are both shifting, and it is not obvious search was ever one market forever.