The word "AI" has stopped meaning anything specific, which is exactly the problem. It gets used for the spam filter in your email, the recommendation engine on a shopping site, the chatbot on a support page, and also for the thing in movies that decides to take over the world. Those are not the same category of thing. If you run a small or mid-size business and you're trying to figure out whether "AI" is worth your time, the first useful step is narrowing down which of those you're actually being asked to care about.
When people talk about AI right now — in the context of tools like Claude, ChatGPT, or Gemini — they almost always mean a large language model. It's worth understanding, in plain terms, what that is and isn't.
At its core, a large language model starts as a system trained on an enormous amount of text, learning statistical patterns in how language works — what tends to follow what. If that were the whole story, you'd get something like a very sophisticated autocomplete, and for a while that comparison was reasonably fair. It isn't anymore, and this is the part that gets oversimplified in both directions.
After that initial training, these models go through a second phase where they're specifically shaped to behave like helpful assistants — trained on examples of good responses, corrected when they go off track, tuned to follow instructions rather than just continue a pattern. That's why talking to one feels like talking to an assistant instead of watching a sentence auto-complete itself. It's also why the "it's just predicting the next word" explanation, while technically part of the mechanism, undersells what's actually been built on top of that mechanism.
On top of that, the tools you're likely using today aren't limited to what they memorized during training. Many can search the web, pull in a document you've uploaded, or call out to other tools and databases, then bring that information back into the conversation. Some now have modes built specifically for working through multi-step problems — reasoning through a task piece by piece rather than answering in one pass, and revising their own approach along the way. Some retain memory across separate conversations, so they're not starting from a blank slate every single time. None of this makes them human, or means they "understand" the way a person does — that's a genuinely open question, and I'm not going to pretend it's settled. But it does mean the dismissive version of this explanation — "it's just guessing the next word, it doesn't know anything, it forgets everything instantly" — is no longer an accurate description of the tools you're actually being sold.
Here's the honest counterweight, because I'd rather undersell this than oversell it: these systems still get things wrong, sometimes confidently and sometimes in ways that are hard to catch unless you already know the answer. They can be inconsistent from one run to the next. The quality of what you get out depends heavily on what you put in, and on which specific tool and mode you're using for which task — a quick question and a complex multi-step analysis are not the same job, and treating them the same is where a lot of businesses get burned. None of that is a reason to write the whole category off. It's a reason to treat these tools the way you'd treat a genuinely capable but new employee: useful, worth investing time in, and not something you hand the keys to without checking their work.
So if someone asks you "are you using AI in your business," the honest first question back is "which kind, doing what." A chatbot that answers FAQs, a tool that drafts first-pass marketing copy, a system that reasons through a spreadsheet, and a fully autonomous agent making decisions without a human checking them are four completely different levels of risk and usefulness. Most of the value available to a small or mid-size business right now sits in the first three categories — places where a capable assistant saves you real time on real work, with a person still reviewing the output. The fourth category exists, gets a disproportionate amount of the hype, and is where most of the expensive mistakes happen.
None of this requires you to become technical. It requires you to stop treating "AI" as one monolithic thing you either believe in or dismiss, and start asking what specific tool is doing what specific task, and whether the output is being checked by someone who'd notice if it were wrong. That's a much more useful question than whether AI is "good" or "bad," and it's the question that actually determines whether the time you spend on this pays off.
I write about this kind of thing regularly, in plain language, over at 013labs.com — worth a look if you want more of this without the hype.