Ecommerce AI hit an inflection point.

In the first quarter of 2024, people published 71 ecommerce-AI projects on GitHub. In the first quarter of 2026, they published 1,429. Same three-month window, about 20.1 times the output. This is a short, honest look at what changed, what people are building, and where the line points next.

I work on conversational AI for a living, mostly on the customer side: support, self-service, the agents people actually talk to. So when it started to feel like every store suddenly had an AI something, I wanted a way to check the feeling against data instead of vibes. GitHub is a decent place to do that. It won't tell you what shoppers are buying, but it does tell you what builders are building, in the open, with timestamps. That's enough to draw a line.

Where we were

For most of 2024, ecommerce AI on GitHub was a slow trickle. New projects held flat at roughly 71 to 109 a quarter. The work that existed looked the way it had for years: recommendation engines, scripted chatbots that fell over the moment you went off script, a bolt-on "related products" widget here and there. Useful, narrow, not new.

The three trajectories below tell the same story three ways. Flat, then a bend, then a near-vertical climb. The steepest of the three isn't the broad ecommerce line. It's shopping agents, software meant to browse and buy on a person's behalf, which went from 5 projects a quarter to 470, about 94 times as many.

The turn

The bend starts in early 2025 and turns sharp at the end of it. New ecommerce-AI projects roughly doubled from the fourth quarter of 2025 to the first quarter of 2026, from 708 to 1,429. Put another way: about 41 percent of everything built in the previous fifteen months landed in that single quarter. That's not steady growth. That's an inflection.

The timing isn't a coincidence. The back half of 2025 is when the plumbing for agent-driven shopping got laid down in public. Google announced its Agent Payments Protocol in September, with more than sixty launch partners.^[2]OpenAI and Stripe put out a competing open standard, the Agentic Commerce Protocol, the same month.^[1] Visa and Mastercard each launched their own agent-payment schemes in October.^[3] When the rails for "an AI can complete a purchase" get built, the apps that ride those rails show up right behind them. On GitHub, you can watch them arrive.

What people are actually building

A rising line is only half a story. The more useful question is what all that activity is for. To answer it without putting my own assumptions on the scale, I let an algorithm group the projects by the words in their names and descriptions, then trained a small classifier to sort the whole set into a handful of buckets. Both are described plainly at the end. Here's the split, across the 2,331 most-starred projects from the quarter.

Customer support is the single largest slice, about 37 percent. That tracks with where the money is going. Decagon, an AI customer-support startup, raised at a 4.5 billion dollar valuation in January 2026.^[4] Support is the place where AI has the clearest job, the cleanest data, and the most painful status quo, so it's the place builders and buyers reach for first. It's also, for the record, the work I spend my days on.

The second cluster, search and recommendations, is the old guard, about 28 percent. Shopping agents are smaller at 13 percent, but that number undersells them because they are the fastest-growing strand and the one pulling in the most stars.

The thing that surprised me was hiding in the unsupervised groups: a whole cluster of small projects that wrap real retailers as tools an AI agent can call. MCP servers (small adapters that let an AI agent operate an outside service) for Costco, CVS, Walmart, and Shopline, so that an assistant can "shop wholesale" or reorder a prescription on your behalf. A year ago that idea was a conference slide. In this snapshot it's a folder of working code. That's the agent shift in its rawest form: not a chatbot bolted to a store, but a store turned into something an agent can operate.

You can see the same pattern in what's gaining traction. The most-starred projects in the sample lean toward agent skills, MCP tools, and open-source support platforms rather than another storefront template.

How I counted this, and what it cannot tell you

A trend piece is only worth reading if you can check it, so here's the method in plain terms, with its limits stated up front.

• The data is GitHub, and only GitHub. A script counts public repositories that match terms like "ecommerce AI", "customer support AI", and "shopping agent", by the quarter they were created. That measures what developers build in the open. It does not measure revenue, or what shoppers use, or what happens inside private company codebases. A spike in projects is a signal of builder interest, nothing more.
• Two different views. The quarter-by-quarter counts cover the whole population. The breakdown of what people build is drawn from a sample of the 2,331 most-starred projects from the quarter, so those shares describe the sample, not everything.
• The grouping is real machine learning, and it is imperfect.One model groups projects by their text with no labels from me. A second, a text classifier, sorts them into buckets after learning from a set of keyword-seeded examples. It agreed with held-out examples about 92 percent of the time overall, but it is weak on the smallest buckets, so read those as rough. The keyword seeding also means the score measures how well the model reproduces my rules, not ground truth.
• Search terms catch strays. "AI helpdesk" pulls in a few university IT desks that have nothing to do with shopping. A lot of the projects are small student and portfolio builds with no stars. That is part of the story, the tools got cheap enough that anyone can build with them, but it is not the same as production systems.

All of it, the collection script, the analysis, and the exact numbers behind every chart, is open.^[9] If you think I drew the line wrong, you can pull the same data and redraw it.

Where we are going, or might be

If you carry the recent growth rate forward, roughly 49 percent a quarter, the rest of 2026 lands somewhere between 3,300 and 5,000 new projects in the final quarter, with a midpoint near 4,083. I don't believe the high end. Curves like this one bend back toward earth as a space fills in and the easy projects get built. The honest read is direction, not a number: the building is still accelerating, and it hasn't peaked.

Two things I'll say with more conviction, and one I'll flag as a bet.

First, the center of gravity is moving from "AI that recommends" to "AI that acts." The shopping-agent curve, the MCP-server cluster, and the payment protocols all point the same way. The interesting question for a store is no longer whether an AI can suggest a product. It's whether an AI can complete a job end to end, and who's liable when it does.

Second, the early version of agentic checkout is already getting a reality check. OpenAI launched buying inside ChatGPT, then pulled it back in March 2026 after fewer than thirty Shopify merchants went live with it.^[7] Forrester, watching the same retreat, called it what it was: the loudest name in agentic commerce stepping back.^[8] The market is converging on a calmer shape, where the AI handles discovery and the merchant keeps the checkout. The same week OpenAI retreated, Shopify turned on agentic storefronts for every store by default, on exactly that model.^[6] The hype was wrong about the mechanism. It was right about the direction.

And the bet: by the end of 2026, the phrase "AI customer support" stops meaning a chatbot that deflects tickets and starts meaning an agent that resolves them, with a refund, a reorder, or an exchange it can actually carry out. The funding, the protocols, and the GitHub curve all lean that way. The thing standing in the way isn't model quality. It's the unglamorous work of permissions, guardrails, and knowing when to hand a human the wheel. Which, conveniently, is the work.

What to do about it, if you run a store

You don't need to chase every protocol. A few moves hold up across most of the scenarios above.

1. Make your catalog legible to an agent. Clean product data, clear policies, and a structured feed are now read by assistants, not just shoppers. The store an agent can understand is the store an agent will recommend.
2. Start with support, not a storefront agent. It's where the data is cleanest and the payback is fastest. Amazon says shoppers who use its Rufus assistant are far more likely to buy, and that the assistant drove billions in added sales.^[5] The customer-facing answer is where the value shows up first.
3. Design the handoff before the automation. The projects that earn trust are the ones that know what they cannot do and pass it to a person cleanly. Decide your confidence threshold and your escalation path before you turn anything on.
4. Keep checkout yours. Let AI do discovery and answers. Hold on to the payment, the account, and the relationship. That's where the platforms themselves are landing.

None of this requires betting the company on a forecast. It requires reading the line honestly: a category that was a trickle two years ago is now a flood, the flood is moving from suggestion to action, and the boring parts, trust, handoffs, and judgment, are the parts that will decide who it works for.