AI Fundamentals: What Is AI, Really?

For all of the investment and hype around AI, it's been difficult - even for someone as embedded as me - to really understand fundamentally what it is, and how it's both similar and different from what's come before. If you've struggled like I have, and want to understand what this thing is that's taking over our world and how it can work for you, here's my best explainer as of June 2026.

There are four main facets to AI as they exist right now: Inference, Data, Automation and Agency. These four systems work together to make the robust AI tools that deliver the most business value. You don't have to use all of them, but as you develop and plan and start layering them together, your AI systems get more powerful.

They also get more expensive, so keep that in mind.

Let's talk through them, then we can talk about how they layer to get stuff done.

Inference

Inference is the thing we all think of when we think about AI - it's the machine thinking like a human. It enables a few key things:

• interpretation of less than perfect information. If your input isn't perfect, it can probably figure it out, identify patterns and understand the semantics, not just the literal data.
• generation of new information. It can create. Not just find synonyms, not just reorganize but make new information based on the patterns that it has seen before.
• decision-making when asked to. If you say "based on X, should I Y or should I Z", it will give an answer that isn't based on predefined rules, but based on the interpretation we mentioned above.

Inference is determined by the Large Language Model (LLM or Model) that represents the "software" that's doing the things above.

Since computers were invented, we've used them to help make decisions, what's changed with the introduction of Inference is that whereas previously, computers needed highly formulated and organized data to be able to understand it and compute, now it can understand a much more human "mess" and infer what was intended, not just what was said. Knowing that inference can work with messy data means you don't have to clean up your data or perfect your prompt to get it working - though it still helps.

Inference has been around longer than people think. ChatGPT didn't invent it, but it did bring it into the spotlight.

Data

Data is the information that Inference works with. This comes from a few different places.

• Trained data. All models were trained on a set of data, and have embedded within their trained data both information and patterns it can draw from. If you ask ChatGPT what happened two days ago, or last week, it won't know without doing a web search. Some Inference models can't access the internet - those models can only provide info based on your input and their trained data.
• Context. Context is "the previous information in the thread" that prevents your AI from being like Dory from Finding Nemo. Without context, it would treat every message like a brand new conversation.
• Input. Every time you put a prompt into your AI, you can give it data. When you say "Help me write this email to be less confrontational", you're giving it two pieces of input: a prompt (the request), and your email draft (the input data).
• Memory. Lately AI is starting to allow retrieval of information between chats, and the generation of explicit memories. These allow the AI to reach beyond the current chat to recall things you told it before.

For most things, data is represented by tokens - basically units of text the model uses and generates. The more data that an AI has to juggle, the more tokens it uses, and the more it costs. That's why data is a balancing act. Too little and the Inference doesn't have enough information to provide a quality answer. Too much and not only does it cost a fortune, but the model can straight up get overwhelmed.

The data that you choose to share with inference, its format and volume play a huge role in whether you find value in AI. Many novices don't provide enough data and expect instant, perfect answers. Many intermediates provide too much, spend all their tokens, and wonder why the model gives them questionable output. It's like cooking. Get the right ingredients in the right amounts and it's delicious, but you can have both too much and too little of a good thing.

I'll be honest, this just takes experimentation and practice. Don't think of it as a search engine with a single answer. Wrestle with it a bit. Go back and forth more like a dialogue, less like an instruction.

Automation

By itself, AI doesn't do anything until you tell it to. Automation is what lets it act without your direct involvement. Automation has been around for decades and usually doesn't require Inference directly - it can leverage simpler mechanisms to take action. Generally speaking, they're broken down into a few types of "triggers"

• a schedule (every 15 seconds, or every day at 2am.)
• a procedural algorithm (if this, do that, not the other thing)
• another action taking place (a new row was added to the database, an elevator arrived at a floor, someone started a machine)

A lot of people confuse Automation with Inference directly, but Automation is really just software - sets of instructions that interact with each other, a clock or another system to get things done. There have always been wizards that have used computers to automate their work through schedules and excel macros and other functional computing - Inference has simply made this easier to work with.

Automation can use Inference as a tool to help smooth out some of the data troubles that have plagued it historically, but whereas Inference is interpretive, automation is reliable. If you can describe how something should be done, don't use Inference. Use Automation.

Automation is how you remove some of the work from people's plates - specifically the kinds of work they have to do over and over again the same way. Repetitive. Rule-based. Regular. Automation is your answer.

Agency

Agency is when the AI system has the ability to act and do things, and not just "think".

When you run ChatGPT in the chat interface and you ask it to help you rewrite an email, it takes the text you give it, and returns different text - it doesn't go in to change your email draft in your email client. But it can do that, and many other things.

You can give your AI system agency in one of two ways - deterministic, or autonomous.

• Deterministic - Sometimes an existing piece of software will use AI as a service but the next step was decided in advance. Your Outlook can suggest that you rewrite your email for clarity. You click the button, it calls an Inference model, and then replaces your email contents with the result. The action that's taken is predetermined.
• Autonomous - AI interfaces are now introducing software called "Tools" for actions that the AI can call directly. The simplest is a web search tool. If the training data for your model doesn't have the answer it needs, it can go search itself. You don't tell it to do that, it can make that determination itself.

These connections to other services - your email, your task system, your customer service queue - they're the things that turn ideas into action. Inference has those ideas. Agency makes them affect the real world.

Choosing what agency to give your AI systems is scary - it puts your money on the table. The two strategies for managing that danger, however, are straightforward: make deliberate choices around Inference and Automation, and keep a human in the loop. These will save a lot of headaches.

Conclusion

Now you know. The next time someone tells you "Just use AI!", you'll have a mental model of what the pieces are, and enough structure to start fitting them together to get stuff done.

Most AI systems that you can use these days layer them together, and the different "Products" that AI companies offer - CoWork, Chat, Agents, Schedules, whatever, they're all just different orchestrations of these four basic principles put together in productive ways.

In the next article, we'll talk about a few of those different orchestrations and how you can practically use them.

‍