The question I get most often from other business owners is not “Which model should we use?” but “Where do I even start?” My answer: with principles that will outlast the next model release, not with tools. Most people start at the wrong end: a chat window, a few prompt tricks from LinkedIn, and six months later a subscription and no clear idea what has changed in their company. The map below goes the other way: principles first, then practical work, then the technical layer, then agents. It is the map I use myself.
Level 1: Principles that outlast the next model
Two books gave me a foundation that has held up so far: Reid Hoffman’s Impromptu: Amplifying Our Humanity Through AI (2023) and Ethan Mollick’s Co-Intelligence: Living and Working with AI (2024).
Neither is about model features or prompting tips. Both are about how to think when you work with something that is neither a tool in the classic sense nor a colleague. Hoffman wrote Impromptu together with GPT-4 and shows the collaboration on the page, including the places where the model falls short. That makes it an honest demonstration rather than a sales pitch, and the book is a free PDF on the author’s own site.
Mollick, a professor at Wharton, is more systematic. He sets out four principles for working with AI, and they are a good example of what I mean by “principle instead of trick.” Always invite AI to the table: try it on every task, including the ones where you expect it to fail, because that is the only way to learn where it helps. Be the human in the loop: the model produces, you decide, and you stay responsible for what goes out. Treat AI like a person, but tell it what kind of person it is: not anthropomorphism for its own sake, but the practical observation that a clearly defined role produces far more useful output than a bare question. And assume the AI you use today is the worst AI you will ever use: whatever limitation you hit right now is a snapshot, not a verdict.
These principles will change over time too. But they last a lot longer than most of the noise around AI, and they give you a way to judge that noise. Mollick also publishes regularly on his blog One Useful Thing, and I read every post.
Level 2: Learning to actually work with it
Once the principles are in place, the next question is practical: how do I work with these systems day to day? I would start with what the model makers themselves put out. Anthropic runs a free learning platform, Claude Academy, with short, plain introductions to working with Claude. They are simple but not trivial, and much of what they teach, such as how to frame a task and where the model’s limits are, applies to other models just as well. In my experience, a little more knowledge multiplies the output. OpenAI has an equivalent in the OpenAI Academy, and Google offers Google AI Essentials; both are free too. Pick the one for the model you use most.
If you want to go further, I send people to Jules White’s Coursera courses. White is a professor at Vanderbilt University, and his Prompt Engineering Specialization consists of three courses: Prompt Engineering for ChatGPT, ChatGPT Advanced Data Analysis, and Trustworthy Generative AI.
The courses are a little nerdy; that is fair warning, not criticism. White shows things rather than describing them, builds up from simple patterns to more demanding ones, and makes you do the work yourself. If you want to move from “I tried it once” to “I know what I’m doing,” this is where I would spend the time.
Level 3: Going deeper
If you want to understand what is underneath, Andrew Ng is the person I would point you to. Ng founded DeepLearning.AI, co-founded Coursera, teaches at Stanford, and previously led Google Brain and served as chief scientist at Baidu. DeepLearning.AI offers a large catalog of courses, from short introductions to material that is technically demanding. Not all of it is for a business owner without an engineering background, so pick carefully.
What I would recommend to everyone, regardless of background, is his newsletter. The Batch comes out weekly and usually opens with a letter from Ng himself. It is part of my required reading every week, because that letter tends to contain a judgment, not just a summary.
Level 4: Agents
The fourth level is where my own work sits: agentic AI, meaning systems of AI “employees” that take on defined roles, hand work to each other, and escalate to a human where it matters. If that is what interests you, I recommend Mike Schwarz of MyZone AI. Schwarz founded the company, which builds AI automation for small and mid-sized businesses and runs an AI Learning Center with guides and articles.
I see him as both a visionary and a practitioner: someone who is building and orchestrating a whole team of agents for work and everyday life. His approach of routing decisions to humans instead of letting agents run unsupervised matches what I have learned building my own setup; the “Building the team” series on this blog will show what that looks like in practice.
What I’d do in the next 30 days
If you are a business owner with limited time, this is the sequence. Read Co-Intelligence first, and while you are reading it, apply the first principle: bring AI into every task you do for a month, including the ones where you expect it to be useless. Keep a simple note of where it helped and where it failed. Spend an evening on the academy of the model you use. Then subscribe to The Batch and One Useful Thing, and unsubscribe from anything that only reports announcements. If after 30 days you want to get good at this, start White’s specialization.
Do not buy a tool yet. The principles come first; the tools change.
Part 2 of this reading map follows, with the sources I use for the technical layer and the weekly tool landscape: Chip Huyen, Victor Dibia and Matt Wolfe.