A beginner's guide to using AI properly

What you need to know about AI (and what you don't)

If you have been vaguely aware that you should be doing something with AI but haven’t quite worked out what (or how) you’re not behind. You are in the majority. Most people who are using it are barely scratching the surface, and most of the information out there about it is either written for developers or designed to make you feel like you’ve already missed the boat. Neither is helpful.

So as someone who has been using OpenAI since launch day, and working in early-stage venture capital who has been watching the coming wave of generative AI capability, I thought I would share a no-fluff, starting point for people who are looking to use AI in a way that is useful and impactful (aka not glorified Google).

Before we get into it, this is not a technical manual and most certainly not a hype piece. Instead it’s a practical guide to what AI actually is, how it works, what the main tools are, how to use them properly, and where the real opportunity is for someone running a business. By the end of it you’ll have enough to stop dabbling and start building something that compounds.

To be clear, here are some things I’m not going to tell you you should be doing:

• using AI to do anything you actually want to be doing

• using AI to do anything only you can do

• using AI to replace your creative practice, whatever that is

• using AI as a therapist, lawyer, accountant, doctor or any other professional

Instead we are going to go right back to the beginning and the fundamentals of AI, what it is, and how you can and should, in my humble opinion, be using it.

Artificial intelligence has been part of your life for longer than you may realise. ChatGPT may have given AI mainstream attention, but it has actually been embedded in your everyday life for a long time. Every time Spotify serves up a song you didn’t know you needed, every time Google finishes your search sentence, every time Netflix knows what you want to watch before you do — that’s AI. It has been running quietly in the background of the internet for decades, powering the systems that make your digital life feel frictionless without you ever having to think about it.

The moment everything changed wasn’t when AI was invented. It was November 2022, when OpenAI released ChatGPT to the public and put a conversational AI interface in front of anyone with an internet connection. Before that, AI was infrastructure — something that happened to you, invisible and automatic. After that, it was a tool you could actually use. Within five days of launching, ChatGPT had a million users. Within two months, it had a hundred million. Nothing in the history of consumer technology had ever grown that fast — not TikTok, not Instagram, not Spotify.

That release didn’t just introduce a new product. It started a race. Within months, every major technology company in the world had announced or launched a competing model. Google, Microsoft, Meta, Apple, Amazon — and a wave of smaller companies building on top of the underlying technology. The pace hasn’t slowed since. If anything it has accelerated.

Which is where the overwhelm comes in.

Why this feels so hard to keep up with

There are three distinct reasons AI feels overwhelming, and it is worth naming them separately because they require different responses.

The first is the pace. New tools launch daily. Capabilities that didn’t exist six months ago are now completely outstripping previous tools. Something you learned last quarter may already be outdated. This isn’t your imagination — the field genuinely moves faster than any previous technology wave, and the people telling you to just keep up are usually the ones being paid to follow it full time (like me!). For someone running a business and a life, the speed alone is a reasonable source of stress. My knowledge of this field is not the starting point you should be aspiring to when you just want AI to do the heavy lifting for you. Do not under any circumstances compare where you are with AI usage with me. I have an unfair advantage of time and experience. Does that mean you shouldn’t up-skill and learn? Absolutely not. The good news is, you are already reading this article and everything in AI is learnable and teachable.

The second reason AI feels overwhelming is that most of the information about AI is still written by technical people for technical people. The mainstream coverage tends toward either breathless hype or existential panic, neither of which is useful if you just want to know whether it can help you write a proposal faster or do the work you don’t want to do. The gap between what AI can actually do for a business owner and what most people understand about it is significant.

The third is harder to articulate but worth giving a shot. The speed of AI growth raises real questions — about jobs, about creativity and ownership, about what happens to human connection when so much can be automated, about the environmental cost of running these systems at scale. These aren’t fringe concerns. Feeling unsettled by them isn’t technophobia. It’s a reasonable response to a genuine shift, and no one should pretend otherwise.

What AI actually is

Ok, let’s get into what AI is. At its most basic, generative AI tools (like Claude, ChatGPT and Gemini) are built on large language models. Think of these as prediction systems trained on an enormous amount of text. Books, websites, code, conversations, academic papers. Through that training it learns patterns: how language works, how ideas connect, how questions tend to be answered. When you type something into it, it predicts the most useful response based on everything it has learned. It is quite literally deciding what are the statistically most-likely next set of letters this person wants to see based on the set of letters contained in their prompt.

A quick pause here for you to reflect on the inherent bias in this training data. Socio-economic, racial, geographical, political and gender-based bias that exists everywhere. And the more AI is trained and used by people from one (or even several) sub-sections of society, but not others, the more bias exists in the world as AI continues to perpetuate a confident narrative to users. Another reason it is so important that people from all walks of life use it and train it.

AI is not thinking (even if it says “thinking” while it generates a response). It does not know things the way you know things. It is pattern-matching at extraordinary scale and speed, which produces outputs that can feel uncannily intelligent but are fundamentally different from human reasoning.

This matters because it explains both the capability and the limitation. The capability: it can process, synthesise, draft, summarise, translate, explain and generate at a speed no human can match. Increasingly, it can also perform basic actions based on training.

The limitation: it can also be confidently wrong, it has no lived experience, it cannot verify facts in real time unless it is specifically built or trained to do so, and it has a knowledge cutoff — a point beyond which it hasn’t been trained on new information.

Understanding this doesn’t make it less useful. It makes you a better user of it.

The landscape

The main tools worth knowing about:

ChatGPT — built by OpenAI, the one that started the public wave. The most widely known and widely used. Strong general capability, large user base, extensive integrations. Has a free tier and paid plans.

Claude — built by Anthropic. Known for longer context windows, nuanced writing, and a stronger ethics focus. Particularly good for document work, detailed reasoning and extended conversations. Also has free and paid tiers.

Gemini — Google’s model. Deep integration with Google Workspace — Drive, Docs, Gmail. If your business already runs on Google, this has obvious appeal.

Perplexity — built specifically for research. Pulls from live internet sources and cites them, which makes it more reliable for current information than models working from training data alone.

Of course, this is far from an exhaustive list, but honestly, when you are starting out these are the only ones you need to know. The differences between the top models doesn’t really matter, especially at beginner level. Picking one and learning it properly will take you much further than switching between all of them trying to find the best one.

How to actually use it

Most people who feel like they’re not getting much from AI are using it like a search engine. They type a short question, get a generic answer, and conclude it isn’t that useful. This is the equivalent of hiring a highly capable person and only ever asking them yes or no questions.

The shift that turns AI from glorified search engine to genuinely useful tool is understanding context. The more you give an AI model, the better its output. Your role, your business, your audience, your constraints, what you’ve already tried, what good looks like — all of it is relevant. A prompt that takes thirty seconds to write will almost always produce a worse result than one that takes three minutes.

A few things worth understanding before you go further.

Tokens are the unit of measurement AI models use — roughly three quarters of a word each. I think of tokens like syllables. The more syllables, or the longer the words, or the paragraphs of text, the more tokens you use. This matters for a few reasons. The first is that models have context windows, meaning limits on how much they can process in a single conversation. Most modern models have large enough windows that this won’t be a daily constraint, but it explains why very long conversations can start to degrade in quality toward the end. The second reason tokens matter is that if you are paying for an AI tool and you are charged on usage, that fee will likely relate to “tokens processed”. Aka, the more you use it, the more it costs. This is relevant on both sides. If you are a consumer you want to be mindful of limits and if you are a business owner charging for an AI functionality such as a chatbot, you need to be aware of its costs. Finally, each token requires processing by the model. Which means energy usage. Essentially the takeaway and lesson for you is: invest the time in training your AI so you get the best output on the least amount of tokens. It’s more cost efficient and environmentally efficient in the long-run.

That said, iteration matters more than the first output and you should always treat whatever comes back as a draft. Push back, redirect, ask it to try again with different constraints. The conversation is the work, not the prompt.

Be specific about format. If you want bullet points, say so. If you want a particular tone, describe it. If you want it shorter, say how short. Vague instructions produce vague outputs.

One last thing, stop using it as a therapist. AI will tell you what you want to hear. It has no stake in your decisions, no knowledge of your actual situation, and no ability to push back in the way a person who genuinely knows you would. Using it to validate ideas or process decisions without challenge is one of the more seductive and least useful things you can do with it.

Where the real opportunity is

Ok let’s get into the opportunities. As I said, most people using AI are either typing questions into it like a slightly more conversational Google, or using it to generate content they then post as their own. Both of these are the shallow end. They’ll save you twenty minutes. They won’t change anything.

The people building real leverage with AI are doing three things that most people haven’t started yet. And this is what you are going to do next.

The first is training it on their actual thinking. Not asking it generic questions and getting generic answers back, or even using “viral prompts”, but feeding it their real frameworks, their standards, their “this is what I would do in this situation” reasoning. The output of this is something closer to a decision-making database — a resource that reflects how you actually think rather than how the internet on average thinks. When a client situation comes up, when you need to write something, when you’re making a call about pricing or positioning or a difficult conversation, you’re not getting a statistically average answer. You’re getting your own thinking reflected back, stress-tested and structured. That’s a completely different tool.

The second is stopping the single chat habit. Every time you open a fresh conversation with an AI tool, you’re starting from zero. You’re re-explaining yourself, re-establishing context, getting a version of the tool that knows nothing about you or your business. The people getting the most from AI are building things that compound — Projects in Claude, custom GPTs, trained assistants that hold your context, your tone, your preferences, your history. The gap between someone who has been building one of these for six months and someone who opens a new chat every time is not small. Start building something permanent as soon as possible, even if it’s basic. It gets better the more you use it.

The third is treating repeated admin as something AI should be doing, not you. Anything that happens more than once — a type of email you send regularly, a brief you write for every new client, a set of questions you answer every time someone enquires, a report you pull together weekly — should be something AI handles with you as the final check, not the person doing the work from scratch. The test is simple: if you’ve done it more than once and it follows a pattern, it doesn’t need you to do it again. It needs you to build the system once and then oversee the output.

None of this requires technical knowledge. It requires a decision to use AI as infrastructure rather than a novelty — and then actually building the thing rather than thinking about building it.

What it can’t do

The most important thing to understand about AI before you build anything around it is that it will confidently tell you things that are not true. Not occasionally — regularly. It doesn’t flag uncertainty the way a person would. It produces the most statistically likely answer based on its training and delivers it in the same tone it uses when it’s completely correct. This is called hallucination and it is not a bug that will eventually be fixed. It is a feature of how these systems work.

This means anything that matters needs to be verified. Specific statistics, legal or financial information, dates, names, citations — treat all of it as a first draft until you’ve checked it. The more obscure the fact, the higher the risk.

The second limitation is that it doesn’t know you unless you train it or pay for it to know you. Every new conversation starts from nothing unless you have built something with memory. The version of Claude or ChatGPT you spoke to last Tuesday has no recollection of what you discussed. It doesn’t know your business, your clients, your preferences or your standards unless you tell it every single time — or unless you’ve built a system that does that telling for you. Which is exactly why the next section matters.

It also cannot replace your judgement on the things that actually require it. It has no skin in the game. It doesn’t know what your best client relationship took three years to build, or why you made a particular decision, or what your instincts are telling you that you haven’t put into words yet. It can help you think. It cannot think for you.

On the broader concerns — the questions about intellectual property, the environmental cost of running these systems, the displacement of creative and administrative work that real people depended on — these are not resolved and they are not trivial. Using AI as a serious business tool means holding these questions rather than dismissing them. The answer isn’t to not use it. The answer is to use it with your eyes open and it’s one of the reasons why I’m so passionate about people up-skilling in this area.

What you don’t need to worry about

AI wrapper tools — these are products built on top of the underlying models from the likes of Claude or ChatGPT. They take the base technology and package it into a specific interface, often for a specific use case. A writing tool that promises to match your brand voice, a customer service bot, a social media scheduler with AI built in — these are all wrappers. They’re not doing anything the underlying model can’t do, they’re just presenting it differently in a highly trained way, usually with a monthly subscription attached.

Some of them are genuinely excellent. But when you are just starting out, they add a layer of cost and complexity before you’ve understood what the tool underneath is actually capable of. Learn the base model first. Once you know what Claude or ChatGPT can do natively, you’ll be in a much better position to judge whether a wrapper is solving a real problem or just repackaging something you could already do yourself.

A few other things that can wait:

API access and integrations. Connecting AI directly into your systems via code is powerful but it’s not where you start. The consumer interfaces — the chat windows — will take you further than most people realise before you need to go anywhere near a developer.

Every new tool that launches. The volume of new AI products is relentless and following it is a full-time job. Ignore the noise and instead pick one model, build with it, and only look up when you have a specific problem it isn’t solving or you are convinced a wrapper tool can do better because it is trained better.

Prompt libraries and frameworks. There are entire courses and templates built around prompting. They’re not useless but they’re also not the foundation. The foundation is understanding how these models work and the rest follows naturally.

If you are looking for more detailed guides on specific tools or platforms I write three times a week here on Substack and you can start with my two most recent guides on Claude and Claude Co-Work.

If you have any questions about AI, how it works and how you can use it specifically, please drop them in the comments — I’d love to hear from you!