4 Levels of AI adoption (where are you?)

👋 Hi!
​
Today I want to talk about a very niche topic nobody's discussing: AI. 🙃

All joking aside: I'm deep in prep for my ggplot2 course launch, and I've decided to add a module on the intersection of R, dataviz, and AI.

That means lots of research, community conversations, and tool-testing. Along the way, I've identified 4 levels of AI adoption: and I think knowing where you stand can be genuinely useful.

---

0️⃣ No AI

AI is a game-changer, but it comes with real limitations: cost, ethics, privacy and more.

I recently ran a poll asking the dataviz community how they use AI. The results were eye-opening: 9% actively refuse to use AI, and another 10% haven't tried it yet.

---

1️⃣ AI through chat

Think ChatGPT. You ask, it answers. It already boosts productivity significantly.

But there's a major drawback for people who code: the AI has no context about your project. You end up copy-pasting constantly, and the model fills gaps with suppositions or hallucinations.

It works, but it's friction-heavy.

---

2️⃣ Agentic AI

"Agentic" means the AI actually does things for you. In our world, that mostly means writing code.

I use Claude Code. You chat with Claude in a terminal that reads and writes directly inside your IDE (VS Code, Positron, RStudio… you name it).

Because the AI reads your entire codebase first, it knows exactly what you're working with. It edits files for you. You just review the result.

Honestly, to me going from no AI to ChatGPT was a smaller leap than going from ChatGPT to Claude Code. It's on another level.

---

3️⃣ Personalization

Once you're in agentic AI territory, you've captured roughly 90% of the productivity gains available. But you can go further.

The first step is a configuration file that personalizes the AI. Mine tells Claude who I am, what my projects are, and small stylistic preferences, like the fact that I don't like em dashes.

Small input, surprisingly big impact.

---

4️⃣ Skills

AI can be weak in niche areas. Many of us feel it's less fluent in R than in more widespread languages like JavaScript for instance.

The fix: pass a skill.md file that gives the AI deep knowledge on a specific topic.

For dataviz practitioners , this is powerful.
​
You can pass a file with dataviz best practices that nudges the AI toward direct labeling over legends. Another that recommends patchwork for plot composition over other libraries. One that enforces your company's design rules.

It removes a lot of friction.

I'm currently building a collection of the best skill.md files to include with ggplot2 uncharted. Students will get them as part of the course.

---

That's the mental model and hope it can add some clarity in this world where AI buzzwords are everywhere everyday. It's useful to know where you are, and to think about what the next step might look like.

If you're an R user with AI tips and tricks, I'd love to hear them! And thanks to everyone who did already.

All the best,

Yan

P.S. ggplot2 uncharted will run as a cohort soon. First students start May 4th! More details coming soon.

P.P.S. Next week I'm heading to Strasbourg to meet Cedric Scherer, the best ggplot2 creator I know. If you have questions for him, send them my way, we might do a live session!