AI Adoption and the Distribution of Intelligence Problem

When we hire, we don't just hand out a paycheck and hope that good work happens. We think about where to concentrate talent, what problems deserve the most capable people, and whether we've structured the team to get the best work out of it. Headcount is a resource, and we manage it like one.

We need to treat AI the same way.

When we buy AI capacity, we're not buying software in the traditional sense. We're buying intelligence — a finite pool that gets drawn down every time someone opens a session, runs a prompt, or hands a task to an agent. The decision about where that intelligence goes, who gets access to it, and what it's pointed at is a resource allocation decision. And it's one that most organizations are making by default rather than by design.

The invoice tells you how much, not why

Organizations with mature AI deployments can tell you, roughly, who their biggest consumers are. The vendor dashboard surfaces total token consumption by user, but that's typically where the visibility ends.

What you can't see is what those users are actually doing with those tokens, whether the consumption maps to work worth investing in, or whether the budget is being concentrated in the right places at all. A developer who tops your usage chart might be running a frontier model on tasks a lighter one could handle just as well. A team burning through tokens every month might be doing it on work that has nothing to do with what you'd expect. You have no way to know, because the context that would tell you — the prompts, the sessions, the work actually being done — isn't visible from the outside.

Think about what that looks like when your engineering team and marketing team are both heavy AI users. Engineering may be running coding agents through complex, multi-step tasks that include reviewing pull requests, writing production code, and debugging across your codebases. The context windows are subsequently large, they’re using frontier models, and typically, the work warrants it. The marketing team may be deploying similar processes with frontier models to draft email and blog content and consume the same amount of tokens. Could they accomplish the same with a different model?

The result of this ambiguity and lack of clarity is that AI budgets get distributed the same way water finds its own level. They go where people happen to use them, not where the organization has decided they should go. That's fine when the spend is small enough not to matter. It becomes a real problem when it isn't.

Adoption is not the same as proficiency

Within any team, including my own, there are people who have genuinely maximized what AI can do for them. They've built the right habits, they provide the right context, and they produce meaningfully better outputs than colleagues doing the same work with the same tools. The gap between these users and everyone else is rarely effort. It comes down to method, and in most organizations, that method goes nowhere.

The person who figured it out keeps using it. Their colleagues keep working the way they always have. You end up with wide variation in how people on the same team extract value from the same investment — not by design, but because there was never a way to see the gap.

With most work, the better approach is at least visible. You hear the best pitch. A useful process gets noticed. The best output speaks for itself. With AI, the best workflows are opaque. They live inside individual sessions on individual devices that no one else can see. There's no natural mechanism that surfaces them or moves them across a team.

To expand on our earlier example, consider that two people on your marketing team are both using AI to build out campaign materials. One of them has figured out how to load brand context, previous campaign performance, and audience specs into their workflow upfront. The brief comes out structured, on brand, and ready to hand off. The other opens a chat window and starts typing. The output requires multiple rounds of revision, pulls in someone else for review, and takes three times as long to reach the same place. Both show up as active AI users burning the same pool of tokens. It has no visibility into the fact that one person has developed a skill the other hasn't, or that closing that gap would likely cost far less.

Your AI investment is almost certainly producing a wide spectrum of outcomes across people with equivalent access to it. Some portion of your team is getting a lot out of it. Another portion isn't. You likely don't have a clear picture of which is which — which means you're also not in a position to get more intelligence to the people who can do the most with it.

How would you know?

Both of these problems: where the budget is going and whether the people using it have figured out how to use it well, come down to the same thing: you cannot comprehensively observe what's happening with computer use agents without the right foundation in place.

The data to answer these questions exists. It's in the sessions, the prompts, and the agent trajectories accumulating every time someone on your team does work with AI. That data lives entirely on the device. It cannot be meaningfully captured from a vendor dashboard or reconstructed from a network log.

The full chain of what an agent did, on whose behalf, and toward what end is only visible from the endpoint. Capturing that and synthesizing it is precisely what we're building at Origin, so that distributing intelligence across your organization stops being an open question and starts being a managed decision.