Should companies restrict AI access to control costs?

No. Every knowledge worker should have baseline access to an enterprise AI assistant, the same way everyone gets email. Universal access builds organization-wide literacy and lets valuable use cases emerge from the work itself; you cannot find the best opportunities if most of the organization is never allowed to explore. The strategy is baseline access plus targeted additional investment.

Why allocate AI budget by team instead of by role or department?

Because job titles do not create returns; work systems do. Two teams with identical roles and identical token budgets can produce completely different outcomes depending on how well they convert AI capacity into results. Role-based allocation looks disciplined because it matches existing budget structures, but it funds titles instead of the work systems that produce value.

How should AI ROI be measured?

At the team level, with business and delivery metrics rather than AI metrics: cycle time, release frequency, defect rates, customer response time, manual work eliminated, revenue timing, operating cost, and capacity created for higher-value work. Individual metrics collapse into weak proxies like tokens consumed or lines of code produced, which are easy to collect and easy to misuse.

Is high token consumption a sign of waste?

Not by itself. A team that spends twice as much while creating five times the value is underfunded. Low consumption can look like discipline when the team has never experimented enough to find where AI improves the work. Consumption without outcomes is waste, so judge the return.

Does more AI budget automatically make a team faster?

No. More tokens create potential throughput. A team with unclear priorities, weak ownership, or slow approvals will mostly use AI to create work faster than the organization can absorb it. Management still has to remove the constraints around the team before additional capacity converts into results.

Put Your Tokens Where Your Revenue Is

Imagine two engineering teams in the same company, the same size and with the same roles, running the same models on the same AI budget. A year from now, one of them has cut delivery time by 40 percent and retired a backlog of manual work, and the other has a very impressive token dashboard. Most enterprise budgeting processes would fund those two teams identically next year, and that is the problem this article is about.

The instinct behind that budget is old, and I understand where it comes from. For years, enterprise technology budgeting has been built around access and fairness. Every department gets a budget and every employee gets roughly the same tools, with resources handed out by headcount, role, or last year's spending. That works well enough for software licenses, and it breaks down with AI.

AI is productive capacity, and the value of capacity depends on where it's applied and whether the team around it can turn faster execution into business results. The companies that come out ahead over the next decade will be the ones that put AI capacity where it produces the greatest return, which has very little to do with who spends the most.

AI budgets should be managed like an investment portfolio, not an IT expense.

Universal Access Is Table Stakes

None of this means restricting AI access. Every knowledge worker should have an enterprise AI assistant with a sensible baseline allocation, because AI is becoming part of the standard operating environment the way email and collaboration software already are.

That baseline builds literacy across the organization and lets useful ideas come out of the work itself. The team that finds your best use case next year is probably not the team you would pick today, and you won't find it at all if most of the organization never gets to try.

The Problem With Fair Allocation

Once everyone has the baseline, the question is where additional capacity goes, and most organizations will be tempted to hand it out by role. Engineering gets one budget and marketing gets another, each sized by headcount. That looks disciplined because it fits the way companies already budget, and the allocation spreadsheet is easy to defend in a planning meeting. It is also too crude, because job titles don't produce returns. Work systems do.

Look at what those two teams did with the same models and roughly the same number of tokens. The first pointed that capacity at its delivery system. It automated test creation, shortened incident recovery, and cleared a backlog of repetitive maintenance work, and delivery time fell by 40 percent. The second generated more code, more drafts, and more documents, and its throughput, quality, and customer outcomes stayed where they were. Nothing on either team's org chart explains that gap, since what separated them was whether they could turn AI capacity into results somebody could measure. The thing doing the damage is the fairness model, and specifically its assumption that spreading capacity evenly is the same thing as investing it well. Spread evenly, a team that has shown it can turn tokens into shipped work gets the same allocation next year as a team that turns them into drafts.

Why Team-Level ROI Is the Better Measure

The case for the team as the unit of investment starts with measurement, because individual AI productivity is hard to measure well. Companies that try usually fall back on proxies like token usage, prompts submitted, or lines of code, numbers that are easy to collect and look great on a dashboard, which is why they get misused, even though they say nothing about how the work gets done.

A developer may generate code faster while the team is still blocked by unclear requirements, slow reviews, or a fragile deployment pipeline.¹ A product manager can write specifications more quickly, and it won't matter if the team builds the wrong thing. A support analyst can close more tickets while customer satisfaction declines.

Team outcomes connect to the business directly. Did cycle time improve, and did release frequency go up? Are defect rates down, and are customers getting answers faster? Did the team bring revenue in sooner or cut operating costs without adding headcount?

Every one of those is a business or delivery metric the team should have been watching before anyone bought a token, which is the point.

Strong Teams Strengthen Weaker Individuals

Team-level investment also avoids a common mistake, which is treating AI capacity as a reward for the strongest individual performers. That looks efficient, and over time it weakens the organization.

Good teams use AI to raise the performance of the whole group, so when a strong engineer with AI builds a better test harness, or a review checklist the agent runs against every pull request, everyone on the team gets the benefit, including the newest person who doesn't yet know where the sharp edges in the codebase are. A team that learns to use AI well narrows the gap between experienced and less experienced people without pretending everyone has the same skill level.

Hunting for a handful of AI superstars misses where the return shows up, which is in the system around the heaviest user. Knowledge gets easier to transfer and mistakes get easier to catch, and the lift reaches people who wouldn't have produced those results on their own.

Measure What the Spending Produced

The point of an AI budget is to get as much value as possible out of the capacity you paid for, and that calls for different questions than a spend report answers.

Which team consumed the most AI is the least interesting of them. Ask which team cut delivery time, took a recurring operational chore off someone's plate, or brought revenue in sooner. Tokens are the input, and the return is what you're buying.

A team that spends twice as much and creates five times the value is underfunded.

A team that barely uses AI can look disciplined when it hasn't yet found where AI could improve the work.² Consumption without outcomes is waste, and the budget conversation has to catch the quiet team that never experimented as well as the loud one burning tokens on drafts.

Build an AI Investment Portfolio

The best AI budgets will look less like software expense management and more like portfolio management. Everyone gets baseline access and every team gets room to experiment, and additional investment flows to the teams that show they can turn AI capacity into measurable business results.

Some of those investments will fail. Teams will try ideas that don't work. Some workflows will produce less than anyone hoped, and that's how investment works. The failure worth worrying about is continuing to fund low-value activity because the budget was assigned by headcount, title, or politics.

Capital should move. Teams producing strong returns should get more capacity, and teams with promising use cases should get enough room to prove them. A team that isn't producing value should be expected to change its approach, fix the surrounding system, or give the budget back so it can go somewhere better.³

Budget Reviews Should Become Investment Reviews

Traditional budget reviews ask how much was spent and whether the department stayed within plan, questions that tell you almost nothing about whether the money was well spent. An AI investment review should ask what the organization got back in revenue, cost, and delivery speed, at the team level.

That's a much better use of the meeting than arguing over token limits by job title, and it doesn't need perfect attribution. Most enterprise investments can't be traced to a single clean number either, so the standard is enough evidence to make a better investment decision. A team should be able to explain what changed and what AI contributed, what is still in the way, and what more investment is expected to produce.

The Management Responsibility

AI budgeting ends up as a management allocation problem. Leaders have to decide where additional capacity will create the most value, and they have to accept that AI won't fix a broken operating model on its own. Give more tokens to a team with unclear priorities, weak ownership, and slow approvals, and it will mostly produce work faster than the organization can absorb it. More tokens buy potential throughput, and management still has to remove the constraints that keep the team from turning that potential into results.

So AI budgeting can't be handed entirely to IT or procurement. Those groups are good at vendors, security, and contracts, but deciding where the next unit of AI capacity will produce the best business return is a portfolio decision, and it belongs with the leaders responsible for strategy, execution, and results.⁴