Fairness Is the Most Expensive Way to Fund AI
Equal token budgets feel responsible. They are quietly taxing your best work.
Almost every AI budget begins with the same well-meaning instinct: be fair. Give every team an equal allowance. Cap everyone at the same ceiling. Treat each project as equally deserving of the same handful of tokens. It feels responsible. It even feels principled. It is also one of the quietest and most expensive mistakes an organisation can make. A good AI budget does not need fairness. It needs triage.
Fairness asks, "How do we treat everyone the same?" Triage asks a sharper question: "Where does the next dollar change the outcome the most?" One optimises for the comfort of the people allocating. The other optimises for results.
| Fairness (the trap) | Triage (the fix) | |
|---|---|---|
| The question | How do we treat everyone the same? | Where does the next dollar matter most? |
| Who it serves | The comfort of the allocator | The outcome |
| Your winners | Capped and quietly starved | Funded to run |
| The metric | Cost per token | Value per dollar |
| The result | A balanced spreadsheet, hidden losses | Uneven spend, real returns |
Fairness is a decision disguised as the absence of one
An equal split feels neutral, and that is exactly its trap. Choosing to fund everything identically is not stepping back from a judgment call. It is making one, badly, by deciding in advance that your highest-leverage work and your weakest experiment deserve precisely the same support.
No serious operator runs the rest of the business this way. You do not spread capital evenly across every project. You do not pay every employee the same regardless of contribution. You concentrate investment where it compounds and starve what does not. Then the budget line says "AI," and that discipline evaporates, replaced by a spreadsheet that hands every team the same number because nobody wanted the conversation a different number would require.
What the emergency room understands
Walk into any emergency room and you will see the opposite of fairness, applied deliberately and without apology. The patient with a broken finger waits. The patient with chest pains does not. No one calls this unjust, because everyone understands the alternative: treat all arrivals identically, and the people whose outcomes depend on speed are the ones who pay for it.
Triage is not the abandonment of care. It is the refusal to let the least urgent case consume the resources the most urgent one needs.
Some use cases are chest pains: close to revenue, close to the customer, close to a decision that moves real money, and visibly better with every extra dollar of capacity. Others are broken fingers: worth attending to eventually, but not at the cost of slowing the work that actually changes your numbers. Equal funding treats them as the same patient. They are not.
An equal budget is a quiet tax on your best work
Here is the part that rarely gets said out loud. A flat cap does not simply distribute resources. It redistributes them. Every token of headroom you withhold from a high-performing use case so a low-performing one can have an identical allowance is a transfer, away from what is working and toward what is not.
A flat cap doesn't fund everything equally. It taxes your winners to subsidise everything that isn't working yet.
The cost is invisible on the budget line, which is exactly why it survives. You never see the deal that closed a week slower, the model kept a notch less capable, the team that throttled its best idea to stay under a ceiling set by someone who never measured what that idea was worth. The spreadsheet looks balanced. The opportunity cost is enormous, and it lands entirely on the work you can least afford to slow down.
You are almost certainly measuring the wrong number
Most AI cost conversations orbit a single figure: cost per token, or its cousins, cost per call and cost per seat. It is a tidy number. It is also nearly useless for allocation, because it describes what you spent while saying nothing about whether you should have.
Cost per token tells you how much you spent. Value per dollar tells you whether it was worth spending. Only one should decide your budget.
The number that belongs at the centre of the table is value per dollar invested. What did this use case return, in revenue earned, hours saved, risk avoided, or decisions improved, for each dollar it consumed? A use case with a high cost per token and an extraordinary value per dollar is not expensive. It is underfunded. A use case with a low cost per token and no measurable value is not efficient. It is a rounding error you have mistaken for a saving.
How to run triage on your AI spend
The objection worth taking seriously
There is a real argument for equal budgets, and it is not stupidity. Concentrate spend and you invite politics: the loudest team lobbies hardest, the executive's pet project gets fed, favouritism wears the mask of strategy. Equal allocation is a crude defence against all of that. It is fairness as a firewall.
But equal budgeting does not remove the politics. It freezes them at the default and calls the result neutral. The answer to "triage invites favouritism" is not to stop choosing. It is to make the choosing legible: rank on a transparent value-per-dollar number everyone can see and contest. Politics thrives in the absence of a scoreboard. Triage, done in the open, is the scoreboard.
The organisations that get the most from AI will not be the ones that were fairest with their budgets. They will be the ones that were honest about where the value was, and brave enough to fund it accordingly.
I write about AI, data, and learning regularly at pinaldave.com, and I have been teaching this hands-on in my AI workshops.