Most people probably have run into the term “CompanyOS” by now, at least if you’re spending as much time on X or LinkedIn than I am, and honestly it’s more than I’d like to admit.
It’s a great term, with a lot of promise. Not sure it’s much more than hype just yet.
But there’s a real idea underneath the buzzword. And we’ve spent enough time building toward it now that I have a few things worth writing down: less a theory, more a set of things that became clear once we’d actually done the work.
First, the idea itself, stated plainly. The premise is that once AI has access to all the information about your company, such as the context, the history, and how decisions actually get made. It stops being a clever autocomplete and becomes something closer to a co-worker. One that builds alongside you, and eventually takes action on your behalf. It’s a genuinely good promise.
The first thing we realized is that the path there is much longer than the pitch makes it sound.
Lesson one: You grow it, you don’t install it
There’s no threshold moment. No point where you connect enough tools and the system “comes online”.
An AI can’t take meaningful action for you until it has enough context and historical data to know what “good” looks like in your company. And that context doesn’t arrive in a few days. It accumulates. So the honest version of a CompanyOS isn’t a product you switch on. It’s something that gets built up gradually, one useful thing at a time.
Which leads straight into the next thing we learned.
Lesson two: People don’t want to “deploy an OS”, they want one annoying task to take less time
What actually moves the needle is small. A specific, irritating task takes less time today. The first steps should never feel like installing an operating system, they should feel like little wins.
And the funny thing about those small wins is that they’re not really small. Each one quietly brings more of your data and more execution capability onto whatever AI platform you’re using. You’re building the bigger thing without announcing it. The “installation time” of a CompanyOS is very long precisely because it can’t be installed — it has to be grown, and people only let you grow it if each step earns its keep on its own.
Lesson three: Keep the decision with the human
Knowledge work has three parts:
1. gathering data,
2. analyzing data,
3. making the decision (or taking the action).
We tend to mash all three together in our heads and call the whole lump “decision-making,” which is why decisions feel slow. Most of that time isn’t spent deciding. It’s the gathering and the analyzing.
Humans are genuinely fast at the deciding part. Give a person the right information laid out clearly and the call takes seconds.
So the most useful thing we found was pointing AI at the front half: gathering, analyzing, the unglamorous setup work, and leaving the decision with people. Not out of sentimentality, but because it’s safer. A human in the loop means a hallucination is far less likely to wander out into the world with your company’s name on it. Keeping the final call human turned out to be the cheapest insurance available.
Lesson four: The human review costs less than you’d expect
The obvious worry is that a person checking every output drags everything back to manual speed. Extra steps, same bottleneck.
In practice it’s far cheaper than that. We’ve automated a lot of sales-focused work, and having someone glance at the output before it ships takes surprisingly little time. Most pre-sales work can be automated, and total output goes up dramatically even with a human reading things over first.
The review is what lets you trust the system.
Lesson five: The real unlock is “what couldn’t we do before”
The easy win is more of the same, faster, the exact same work, ten times the volume. Useful, a little boring, and where most people stop.
The more interesting direction is depth. AI has no problem spending time and energy researching something far past the point a person would tap out. It’s a depth-versus-breadth situation, and AI quietly removes a trade-off you’ve made your whole career. So the question we found most valuable shifted from “how can we streamline what we already do” to “what can we do now that we simply couldn’t before.”
The mental shift underneath all of it: going from what’s the answer to this question. The normal way most of us use ChatGPT, to do this for me, the agentic version. Different muscle entirely, and the one that takes longest to build.
And the most useful trick we found for learning a platform’s real range is almost embarrassingly simple: ask the platform’s own AI what it can do. Most platforms can do considerably more than people realize. We kept discovering capabilities by just asking.
What it actually comes down to
We’re in the genuine infancy of all this. AI-native work, CompanyOS, whatever the next label turns out to be: using AI at real scale across a company is still a non-trivial thing, and anyone telling you otherwise is selling something.
The biggest thing we learned, and the one that surprised me most, is that the hard part isn’t the AI’s capability. AI can do amazing things and it’ll build and test a marketing campaign end-to-end if you ask it to. The hard part is designing the workflows and workspaces that make the pieces
fit together into something real. Most people are working through exactly that, and there’s a frustrating lack of knowledge sharing about how to do it well. Everyone’s quietly solving the same puzzle in their own corner.
So use AI to help design those workflows. Or reach out to people further along, there are more of them than you’d think, us included.
If I had to compress everything we learned into one sentence: just start, pull in as much context as you can, automate the gathering and the analyzing, and keep yourself firmly in control of the decisions.
About the Author
Tuomas Rinta, CEO
Tuomas has been building digital products for the better part of thirty years, his career split between Finland and Silicon Valley. Now he’s building Luo, an AI platform to help people get more work done.