7 min read
Procurement finally makes a useful mistake.
They meant to order 100 standard corporate laptops. You know the machines. They can run Slack, Outlook, and one Docker container before the fan begins negotiating a severance package. Instead, they bought 10 premium MacBook Pros. More memory than most of your developers have ever seen. Enough local compute to turn builds, containers, emulators, and experiments back into work instead of scheduled afternoon activities.
Before anyone removes the plastic, the fairness machinery starts. HR suggests a lottery. The Agile organization recommends rotating the machines through sticky teams. Development Experience warns about the paved road. Every manager wants one. A vice president quietly asks whether the laptop is good for email.
Then the COO has a better idea.
“Give the machines to our 10 best engineers.”
The COO turns to the Chief Product Officer. “Tell me the five most important goals in the company. Not the fourteen priorities everyone agreed were number one. Five.”
Then the COO works with the CTO to move the top 10 engineers onto those five goals. Two engineers per goal. Clear their calendars. Move ownership. Drop the local backlog work that becomes less important now that someone has been forced to say what important means.
For one dangerous moment, the operating model reflects the company’s priorities instead of its reporting lines.
That is not favoritism. It is capital allocation.
Right now, your AI strategy does the opposite. You take those 10 MacBook Pros, build a reservation system, give every developer eight hours a month, and spend the quarterly business review wondering why the laptops have not produced measurable ROI. That is 800 scheduled hours across roughly 1,600 available workday hours. You rationed the machines and left half the capacity idle. Someone blames adoption. Someone recommends more training. Someone calls Apple to ask how to save money on electricity. By Friday, engineers are told to charge the MacBooks at home and run them on battery at work. Facilities reports savings.
Nobody blames the arithmetic.
Your AI inference budget works exactly the same way.
Equal Distribution Is Not a Meritocracy
Suppose you have 100 engineers and $10,000 a month for AI inference. The easy answer is $100 each.
Most readers also read: If you cannot afford the tokens, can you afford to build it?
The number fits neatly into the expense policy. Procurement reports universal availability. The transformation office calls access democratized. The CFO gets a predictable total. The CTO gets almost no useful concentration of machine capacity, but the pie chart is lovely.
A meritocracy allocates opportunity based on demonstrated ability and results. Giving everyone the same ration before the work begins is egalitarian distribution. That may be right for health insurance or parental leave. It is a strange way to allocate scarce production capital.
The recommendation is simple: fund your most productive engineers, not all your engineers.
Do not accept applications. Do not create a request form. Do not reserve a participation pool so the policy feels inclusive. The leadership team chooses the engineers, chooses the problems, and gives that team the entire budget.
If you made inference scarce, act like it is scarce.
Build the AI All Star Team
The NBA All Star Game does not give every professional player 45 seconds on the court so the league can report universal participation. It selects people who have already demonstrated that they can perform at the highest level and lets them play.
Do that.
Choose your 10 most productive engineers. Not the people with the most Jira activity. Not the people who dominate architecture meetings. Not everyone with “principal” in a title that survived three reorganizations. Choose engineers who understand the business, finish difficult work, make sound technical decisions, protect quality, and improve the system around them.
That is the roster. Leadership picks it.
Give the roster all $10,000. That is $1,000 per engineer, ten times the equal allowance. Give them the strongest approved models, enough context, and enough inference to operate differently for a quarter. Do not spread the money across the other ninety seats so someone can rewrite email with more executive presence.
AI is an accelerator. It amplifies judgment, domain knowledge, and initiative. It also amplifies weak judgment and bad priorities. A strong model pointed at a bad premise gives you a polished bad premise faster.
The engineers outside the roster are not less capable or less valuable to the organization. They are rational people working inside the system you built. A small allowance does not repair that system.
The roster is not a lifetime achievement award either. Pick it for the quarter. Fund it. Measure it. Change it when the evidence changes. Leadership owns the selection and the consequences of getting it wrong.
This concentration will open your eyes. One exceptional engineer fully funded with AI can probably produce the work of five to ten average corporate developers on suitable work. Not because that engineer becomes ten times smarter. The machine collapses typing, research, testing, documentation, and coordination that your old operating model spread across a team.
Some of you are going to say there is no peer-reviewed science proving the five-to-ten number. If that is your first response, you probably have not used AI to build serious software or thought very deeply about what happens when an exceptional engineer can hand the machine the typing, research, testing, documentation, and coordination. Have you looked at how few engineers the AI-native startups doing this work actually have? Start there. Use the tool properly for ninety days, then bring data to the argument.
None of that becomes visible when you average the investment across 100 people.
CPO, Pick Five Things That Matter
The All Star team creates a useful problem: Product has to decide what “most important” means.
Most roadmaps contain twelve number one priorities, six strategic imperatives, three executive promises, and a modernization program everyone supports as long as it does not interfere with anything. That is how the portfolio ends up with 140 essential things.
Now you have one fully funded team. Does it protect a $4 million renewal, remove $600,000 in manual billing work, cut eight weeks from the release path, or rebuild the compliance evidence process blocking a new market? All are defensible. They cannot all be number one.
The CFO states the economic result. The CPO distinguishes company value from stakeholder volume. The CTO explains what is solvable and whether the organization can absorb the result.
Do not bury the team inside one product area so a vice president can clear a local backlog faster. Put it on work the whole company recognizes: revenue, cost, risk, market access, customer retention, or the speed at which strategy becomes production.
Concentrated inference does more than test AI ROI. It tests whether you have a product strategy. If three executives cannot agree where the team spends the next ninety days, the company has not decided what matters.
It will also reveal what does not matter. What you do with all that newly visible roadmap filler is a different conversation.
The Machine May Cost More Than the Operator
If the CFO fixes inference at $10,000 a month, fine. Working capital is real, and “the models are cool” is not an investment case.
But inference should begin as a function of loaded labor and total delivery economics, not as a polite allowance per employee. As agents take on more production work, machine capital may eventually exceed labor capital.
That is okay.
The pilots cost less than a 747. You still do not fly without them.
Human judgment remains valuable because it directs the expensive machine toward a safe, accepted outcome. Software leaders are simply not used to tooling becoming a material production cost. Labor was the factory and tools were a rounding error. That relationship is changing.
Give the All Star team the entire budget for ninety days. Record the old path first: implementation, review, testing, release delay, defects, rework, and total cost. Then measure what reaches production.
If implementation falls from six days to two and the work arrives four days earlier, you have the beginning of an economic case. If implementation falls by four days but cycle time falls by four hours, you learned something equally valuable: code review, testing, security, architecture, or release management is now the constraint.
Run it for a quarter. Continue for six months if the economics hold. Finance sees machine spend beside accepted value. Product sees which priorities survive evidence. Technology sees where the operating model must change.
That is a more useful executive conversation than “AI usage rose 18%.”
This May Just Be Accountability Making Noise
Your directors and vice presidents will complain that their teams are not being invested in. Fine. Change management is real. Hear them.
Then ask what they brought to the decision: the All Stars and one of the company’s best bets, or a local backlog and an argument about fairness? The scarce investment went to the strongest roster against the most important work for ninety days.
That noise does not prove the allocation is wrong. It may just be accountability making noise.
What Happens Next Quarter
Here is my bet: next quarter, the AI budget will not be nearly as constrained.
Not because AI has ROI. AI is a production input. The work has an economic return. Concentrating the budget forced the organization to choose work important enough to produce one.
Every CFO understands the formula:
Accepted value created > total investment
When that inequality holds, fund the next bet. “We invested less than we created in value” is the kind of AI strategy a CFO can approve without a transformation deck with 40 slides. Once the numbers are credible, that same CFO can usually find more money. Finance is remarkably creative when an investment returns more than it costs.
The fastest way to get a larger AI budget is to stop giving the small one to everyone.
Companion
