The invoice arrived. It was bigger than expected.
So someone sent an email, and now every employee in the organization has the same AI token limit. The engineer shipping features daily and the person who opens the tool twice a week are running on identical allocations. I've seen this happen at companies with hundreds of engineers, real AI investment, and genuine usage. The decision took about forty minutes. The damage takes quarters to show up.
That's not cost management. That's panic dressed up as policy. And the reason it keeps happening is that most organizations have the invoice before they have the instrumentation. When the only thing you can measure is total spend, the only lever you can pull is total spend. That's the real problem, and a blanket token cap doesn't solve it. It just makes it invisible.
This isn't a fringe behavior. Microsoft throttled heavy Copilot users in enterprise deployments when usage exceeded planned spend. Salesforce introduced Einstein AI credit limits after enterprise customers blew past projected consumption at scale.
The pattern is the same every time. Large invoice arrives, CFO reacts, blanket restriction goes out. Nobody asks what the spend returned per task, per workflow, per team. The restriction feels like control because the number on next month's invoice is smaller. But the organization has no idea what it cost them to get there, and that's the condition that produces the next bad decision just as reliably as it produced this one.
Every generation of business has learned the same lesson through a different medium. When a campaign starts converting in performance marketing, you don't cap its budget. You scale it. The entire discipline exists to find the creative that works and pour resources behind it. Finding the winner and then restricting its spend isn't caution. It's the opposite of the job.
The same logic runs through sales. Great salespeople get bigger territories, not smaller ones. A company that imposed identical quota ceilings on every rep regardless of output wouldn't be managing fairly. It would be actively destroying the conditions that produce outsized results, and its best performers would be at a competitor within a quarter.
Warren Buffett, who built Berkshire Hathaway into one of the most valuable companies in the world by deciding where capital should and shouldn't go, has made the allocation version of this argument for fifty years. Even distribution of resources feels fair and requires no hard decisions. That's exactly why it's the strategy of organizations that have stopped trying to win. The best allocators find what earns and concentrate capital there, disproportionately and without apology. I keep waiting for someone to apply that principle to AI spend inside their own organization. A blanket token cap is even distribution. It is the thing Buffett spent a career arguing against, applied to the most asymmetric productivity tool most companies have ever deployed.
Blanket caps are a symptom of missing attribution, not a solution to excess spend. The logic is understandable: when a cost center produces an invoice that exceeds its projection and nothing exists to explain the variance, restriction becomes the default response. It is the only lever available when visibility is absent.
The problem is that the restriction gets applied uniformly to a distribution that is anything but uniform. Output within any skilled function follows a power law. A small number of people generate a disproportionate share of measurable value, and AI amplifies that dynamic rather than flattening it. The highest performers aren't just using the tool more frequently. They're using it in ways that compound across tasks, workflows, and decisions in ways that a token count alone will never surface.
Every engineering leader I've talked to about this already knows the answer intuitively. If they had visibility into return per user, per task, per workflow, they would not apply the same cap to everyone. The blanket cap only survives as a policy in the absence of that data. The invoice tells you what you spent. It tells you nothing about what you got. Solving for the invoice without building the instrumentation to read what it produced is how organizations confuse control with visibility. They are not the same thing, and until most companies have the layer that separates them, they will keep reaching for the blunt tool every time the invoice arrives.
The organizations getting this right are treating AI spend the way a performance marketing team treats a media budget. You run the campaigns, you read the signal, and you reallocate toward what's converting. You don't cap the ad set generating ten times return because another ad set is underperforming. You cut the underperformer and fund the winner harder.
The same logic applies directly. High return users, the ones generating measurable output per dollar of AI spend across tasks, workflows, and decisions, should get more budget, not the same as everyone else. Low return usage, repetitive tasks, low-output workflows, tools that aren't moving anything meaningful, gets cut. That's how you spend less and get more. Not by capping everyone equally, but by being surgical about where the value actually lives.
The CFOs and CTOs I talk to already understand this intuitively the moment the frame shifts from invoice management to return on deployment. The reason it hasn't happened yet inside most organizations isn't a lack of willingness. It's a lack of signal. Once you can see return per task, per workflow, per team, the reallocation becomes the obvious next move. The data tells you where to cut and where to concentrate. The blunt tool gets replaced by a decision.
Capping AI spend across the board isn't wrong. It's the responsible first move when you don't have enough information to do anything else. It buys time. It slows the bleed. The CFO who sends that email isn't making a bad decision. They're making the only decision available with the information in front of them, and that instinct is sound.
The question is what happens after the cap. The organizations that treat it as a permanent answer will find themselves managing AI spend into irrelevance while their competitors build the visibility to make smarter allocations every quarter. The ones that treat the cap as a temporary measure, while they build the instrumentation to see return per task, per workflow, per team, are going to move very fast once they have the data they need. That gap between the two groups will not be subtle two years from now.
The goal was never to spend less. It was to spend better. Every major productivity transition in business history has eventually separated the organizations that learned to allocate toward their highest return activities from the ones that managed costs uniformly and called it discipline. AI is not different. The cap creates the conditions to figure out where the value lives. Attribution tells you what to do with that answer.
The cap isn't the enemy. Staying there is. The organizations that move from restriction to precision will find their best performers waiting on the other side, ready to run. The only question is whether you give them the room to do it before someone else does.