There’s an article going around by Alex Lieberman called 30 Features of an AI Native Company. Matt and I have been somewhat obsessed with it internally, enough that we’re going to work through it on the show over the next few weeks. It’s a good list. It asks what the perfect design of an organization would look like if you were building one today, from scratch, to be AI-native rather than AI-adaptive or AI-assisted.
The first item on the list is blueprint your processes. What are the things you do? How do you do them? What are the ins and outs?
I have a caveat on that one right off the bat.
The five steps are an artifact
You should not be trying to replicate every single thing you do in your organization through AI.
If something is five tasks right now, and you’re handing it around between people, and you tell a model like Astra that you want to replicate this process and have all five of these steps, it will probably do that. That’s the part worth sitting with. It will comply. You will get a working system, on time, that does the five things in order, and you will have spent real money encoding a workflow that was designed before any of this technology existed.
Those five steps came from somewhere. Someone was available on Tuesdays. A piece of software required an export before an import. A handoff existed because two departments didn’t share a system. None of those constraints have anything to do with what a model can do now.
So when you blueprint, spend a lot more time on what the input is and what the output needs to be, and much less on preserving the pieces in between.
What changed to make this the right advice
A year ago I would have told you to be more prescriptive, because the models needed the path. Coaxing the technology down the road it needed to follow was most of the work.
Last week I handed Astra a transcript of my prep call with a hospitality client and told it to go figure this out. It opened the client’s site, found appropriate images and downloaded them. It popped open my presentation folder on my desktop and started sorting through past decks, looking for other luxury hotel presentations we’d done. It settled on a design. It built a 45-page deck. It didn’t have the artifacts the early tools left behind. It didn’t need to ask me questions, because I gave it good context up front.
I could have decomposed that into agents. A research agent for the property. Another for imagery. Another modeling the design on prior work. That would have been the blueprint-every-step approach, and it would have worked, and it would have been slower and more brittle than what happened.
The same thing happened with a technical SEO problem. We’re getting killed on searches for AI coaching, where the confusion is between a coach that is AI and a coach on AI. I kicked off a Codex project with Astra and said make this better. In about 45 minutes it built an entirely new page, updated four other pages, wrote seven SEO-optimized articles into draft status, then went to Google Search Console and confirmed the new site index. It needed tweaks under the hood. Getting from nothing to roughly an eight out of ten in that window, without touching it, is a different category of thing.
A few months ago I could have done each of those pieces. They would have been individual pieces. The linkage to WordPress alone would have been challenging.
What this means for your process work
Our job, for those of us who live in this space, is shifting toward identifying the right application for the technology rather than walking the technology through the route. That applies just as much to a leadership team writing down its processes for the first time.
So when you sit down to blueprint: name the thing that goes in, name what has to come out the other side, and note the constraints that are real, meaning security, compliance, approval, anything with a legal or human consequence. Leave the middle open unless you have a reason to close it.
The other item further down that list is reimagine workflows, and it’s worth reading the two together. You will build something, learn from it, and possibly wipe it in nine months. That’s fine now. The sunk cost fallacy isn’t as true as it used to be. We have plenty of things we built in 2023 that are still running and cost effective, and we’re not touching them. But if the only reason a workflow has five steps is that it always had five steps, that’s not a reason.
Worth reading the source list yourself, and then reading your own process docs against it.
