Say "AI in ecommerce" to most people and they picture one thing: a chatbot, or a little widget recommending a product. That's a small slice of what's actually happening. Underneath it are a handful of distinct technologies, at very different stages of being ready for real use, and they don't get talked about separately nearly enough. Here's what each one actually is, in plain terms, and what it changes for a store owner or a shopper.
Agentic AI: an assistant that can actually do something
Ask a regular AI model a question and it writes you an answer. That's it. It can't check whether the answer is still true, and it can't lift a finger to act on it either way. What changes with an agentic system is that the model gets tools, real ways to reach into other software, read what's actually there, and sometimes make a change, plus the ability to loop: check what happened, decide what to do next, take another step. Put those together and you get something meaningfully different from a chatbot. Not an assistant that tells you what your store's problems usually look like, but one that goes and checks your store, finds the specific thing that's wrong, and fixes it.
MCP, and standards like it
Here's a problem that doesn't get talked about much: connecting an AI model to a piece of software used to mean somebody building a custom bridge for that exact pairing. Do it again for the next model, the next service, forever. Anthropic published MCP, short for Model Context Protocol, in late 2024 to get rid of that: one shared way for a model to find out what a service can do and use it, no matter who built either side. A few other AI labs have since picked it up too, which is the part that matters. A standard nobody else uses is just one company's API with a nicer name. This one's turning into plumbing everyone can build on.
Retrieval-augmented generation, or just "look it up first"
A language model only knows what it was trained on, which is a snapshot from some point in the past. Left to its own devices, it'll answer "what's my stock on this item" with something that sounds confident and might be completely wrong. Retrieval-augmented generation, RAG if you want the shorthand, is the fix: have the model check the real record first, the actual stock count, the actual price, then answer from that. It's a much smaller idea than agentic AI, more of a lookup than an action, but it's the whole difference between an answer that sounds right and one that is right.
Computer vision, for the searches words can't do
Not every shopper can describe what they want. Try typing "the chair like this one, but lower" into a search box and see how far it gets you. Models that match images instead of keywords let someone search with a photo, and they let a catalog get checked automatically for the kind of thing a written description would never catch, a product shot at a strange angle, a variant with the wrong photo attached to it.
Demand forecasting, the quiet one
Guessing what will sell is an old problem, nothing new about that. What's changed is that models can now weigh seasonality, price, and traffic together and catch things a simple reorder rule never would, a product trending early before its usual season, a slow seller about to pick up because of a spike in something unrelated. It's the least flashy technology on this list, and for a store carrying real inventory, probably the one with the most money riding on it.
The question none of them get to skip
Every one of these gets more useful the more access to real data it's given, and the more it's allowed to act on that data. And every one of them runs into the same question the moment it does: what stops it from getting something wrong. Nobody in the industry has fully answered that yet. The systems worth trusting are the ones built around the question rather than bolting an answer on afterward, every action shown before it happens, a person saying yes to it, nothing running on its own overnight. Skip that part and you haven't built a more capable version of the idea. You've built a different, riskier one.
Stop asking “how many orders did I get?” Start asking what to fix first
Connect Shopify and Google Ads and see where the money is going — the best seller about to run out, the stock that has stopped moving, the search terms you pay for that never sell. Every fix comes with what it's worth, and getting it done is one word away.
- See the biggest wins first
- Finds it, then fixes it for you
- Catch lost sales before they happen