AI has taken off with coders and some people like me who love to adopt new technologies. But it’s barely made a dent in most companies. Why?
AI adoption isn't lagging because the models aren't good or the AI isn't improving or intelligent or for a bunch of other reasons people throw out there. Adoption is challenged because the two companies building the best AI in the world have the wrong picture of how enterprises will actually use it.
Their vision runs through apps like Codex for OpenAI and Cowork for Anthropic. And it's not that those apps are bad, or that nothing like them will matter. A version of those apps will matter.
The problem is the bet underneath them: that these tools, more or less as they exist today, are the future of work. I don't think they are. Because of who they're really built for, and where they put all the hard decisions.
They're pushing every decision down to the user
Look at what these tools actually ask a person to do. Which model should I use? How many tokens is this worth? How does skill routing work, and how do I figure out which skill to use in the first place? Do I need to write a slash command for this?
That's a developer's world. And it makes sense, because developers built these tools and developers love that control. But what's happening now is that they're trying to extend a developer's mental model to non-developers: to the business owner, the ops manager, or the person in finance who just wants the work done. You're handing someone a cockpit and asking them to be the pilot, the dispatcher, and the mechanic.
Most people at a company don't want to choose a model or design a routing scheme and they shouldn't have to. Every one of those decision points pushed down to the user is a place where the implementation quietly breaks.
We need to manage AI centrally
All of the work I just mentioned needs to get centrally managed by a group of people whose job is to make the right calls and keep making them as things change. Which models to use, which skills are good, how work routes. Those are organizational decisions.
And it can't be set once and walked away from. The models change every couple of months. The company changes. What was the right configuration in the spring is wrong by the fall. Someone has to own that and keep it evolving.
The traditional SaaS approach isn’t going to work either. Some companies are using genuinely powerful enterprise tools like Glean. They work beautifully, if you have a group of people to integrate and maintain it. Most companies don't have that group and can't conjure it.
And once you have a team that can maintain and customize Glean, why do you need Glean? The software itself is cheap and relatively easy to build. It’s the customization and integration that’s valuable. Something like Glean will exist but its ability to charge premium prices will be limited. The team customizing it has every incentive to just replace it with their own solution as soon as it makes a dent in their profitability.
What's actually missing: vertical implementation partners
As I’ve mentioned in past posts, I believe the future of enterprise AI isn't more apps with more knobs pushed down to the user. It's a whole category of companies whose entire job is to centrally manage AI for a specific kind of business. They will pick the models, find and vet the skills, set up the routing, and stay to keep it evolving as the models and the company move.
They will be operators embedded in a niche, who know what good looks like for a dental group, a regional distributor, a law firm. They will run the parts the AI labs keep trying to offload onto the end user.
The labs are betting the user will become the pilot. They won't. They'll hire someone to fly the plane. Until that layer exists, I don't think we get real success with enterprise AI no matter how good Codex and Cowork get.
— Alan