The 5 levels of AI adoption

In my conversations over the last week, I’ve noticed that companies are falling into 5 different levels of AI adoption. (And that’s ignoring those whose organizations are AI-curious or AI-panicked¹ but actually aren’t doing anything intentional or strategic.)

Level 1: AI as Search Engine — This is where everyone starts, because it fits our existing mental model. We ask questions, we get answers.

Level 2: AI as Autocomplete — Give a prompt some direction and context, and it’ll create a draft for you. Some companies are more sophisticated, giving their prompt access to a folder of helpful documents like quarterly planning, brand guidelines, etc. Other companies are further behind, still living in the world of stuffing everything into a single prompt and getting frustrated with the results. Productivity increases, but still a single player game.

Level 3: AI as Whiteboard — “Look what I did in a weekend” is every founder’s dream and every startup employee’s nightmare. Prototyping with AI is an incredible way to whiteboard an idea or build a POC to user test with. But don’t fall into the trap of falling in love with it — this is not the foundation upon which to build production code² . Code generation is cheap now; throw it away (just like we erase whiteboards), and ask your team to build it properly.

Level 4: AI as Reviewer — “What do you think?” is my favorite question, and it applies to AI Agents too. AI as code reviewer, AI as QA tester, AI as PRD and strategy skeptic, AI as UX design critic. This requires the agents be given all the necessary context, which requires some documentation and AI Ops setup³ , but is a really powerful form of staff augmentation.

Level 5: AI as Squad — You’re chaining these specialized agents together, with one agent using as input the output of the prior agent. If you’re really good, you can have the last agent in the chain evaluate the work against the requirements at the top of the chain, and restart the cycle to fix any detected bugs or missing requirements, looping again and again until finished. But don’t take your lunch break too early; it’s our job to make sure the strategy and requirements actually make sense, the technical design plan is sound, the test cases are comprehensive. We’re here to build the right thing, not build anything.

The leap from Level 2-3 to Level 4 is a big one. It’s a cultural change, a technical operations change, and/or requires a level of clarity and documentation that many companies don’t have. But it’s worth it.

At what level is your company?

The Workshop

This is a newsletter-only section where I share a half-baked idea in hopes that y’all who are smarter than me can work it out with me.

First off, I just want to give a big, heartfelt Thank You to everyone who reached out with excitement that I was writing again. It really means a lot to me. And, I’m grateful to all of you who have invited me into your inbox, when often the last thing we need in our lives is another email and another something to read. I appreciate you.

So I was thinking about OpenAI’s new ChatGPT Pulse app over the weekend. Less about the product itself — I could be wrong but it feels like a throwaway MVP to validate some hypotheses moreso than the Next Big Thing — and more about OpenAIs commitment to building consumer apps and not just be a platform.

When I was at The Knot, one of the questions we’d think about is our defensibility against Google and Meta. Google had all the SEO traffic; Instagram had all of the wedding vendors and engaged couples. How vulnerable were we?

At the time, our thinking was that no, we were safe, as long as our product strategy was to win on user experience. If our browse and search tools were superior, if we made the process of planning a wedding more interactive, if we could find synergies like connecting your guest list to your registry to your stationery so that thank you cards were easier to write, then we’d be OK. Which is a long way of saying, bundling the value chain in ways that deliver a superior UX, and within a market that frankly had too small of a TAM for Big Tech to bother with.

I think there is a similar rationale when competing against OpenAI, but ChatGPT Pulse’s vision of “We're building ChatGPT to help you reach your goals” made me wonder if they would, for the largest TAM industries, be willing to build vertical specific apps that can compete directly on user experience.

1  You’ve heard this before — Leadership yells, “Everyone needs to be using AI!” — but there isn’t anyone in middle management who knows how to translate what is essentially an emotional fear response from Leadership into something actionable or practical for ICs.

2  I think one day we’ll get there. But unless you’re Level 5, don’t make this mistake. Whatever code was generated for your prototype isn’t part of your CI/CD, isn’t built to your coding standards (is it even in the same language?), wasn’t architected with existing systems in mind to minimize code duplication, couldn’t pass a security review, doesn’t use your design system. It’s a communication tool, a way to share the picture in your head with others. That’s a lot of value! And please stop there, your engineers will thank you.

3  Two steps to setting this up: 1) Get an agent whose job is to perform the desired Reviewer task. Here’s a list of 83 agents, for example. 2) Give the agent access to the context it needs, which is probably a folder containing strategy documents, style guides, architecture diagrams, your code base, etc; if you’re more sophisticated, it might also include access to MCP servers for design, compliance, QA automation, etc. Yes, it’s technical right now, you’ll need to learn github.