AI capital allocation: from tracking AI spend to directing it

Most conversations about AI cost are about spending less. That is the wrong frame for a technology whose returns vary by person and by task by an order of magnitude. The right frame is capital allocation: the company has a budget for AI, the budget produces very different returns in different hands, and the job is to move it toward the hands that produce the most. That is a management decision, not a bookkeeping one, and it depends on answering four questions in order.

The four questions

What did we spend? A complete ledger across every provider and every shape: API usage, enterprise seats, plans on cards. Not the invoice total, the ledger it can be taken apart into.

Who spent it? Every dollar attributed to a person, a team and a surface, with what cannot be attributed shown as such rather than spread.

What did it sit next to? The output of those people over the same period, from the systems that record it: closed-won revenue, meetings booked, renewals, code merged.

Where should the next dollar go? Given the first three, which people, teams and workflows show the strongest relationship between spend and output, and how does the budget move toward them.

Most companies are on the first question. A few are on the second. The ones on the third have a story for their board. The fourth is where the money is.

Why the order matters

Each question is impossible without the one before it.

You cannot attribute a total you cannot break down. You cannot put output beside spend you have not attributed to the person who produced the output. You cannot allocate toward what works without a view of what works. Companies that skip ahead, usually to a spend cap, are making an allocation decision with none of the information, and the decision they make is to spend less everywhere, including where it was producing the most.

What "directing" actually looks like

Capital allocation sounds grand. In practice it is a handful of unglamorous moves, made monthly, with the numbers in front of you.

Moving budget between teams. The sales team's spend has doubled and its closed-won per rep is up with it; the marketing team's spend has doubled and nothing moved. The budget goes to sales. That is a decision that is obvious with the ratios and invisible without them.

Buying seats for the people producing. The per-person view shows who is getting results from a personal plan they expensed. Those are the people the enterprise contract should cover first.

Splitting shared keys. A team's shared key is most of its spend and nobody can say who is using it. Splitting it into personal keys costs nothing and turns an unknown into a set of people who can be compared.

Retiring idle seats. Every enterprise product lists its members. The ones with no usage are the ones to drop at renewal, and the money goes to seats for people who will use them.

Protecting the outliers. Every company has a few people spending far above their role's median. Some of them are wasting money. Some of them are the people who have worked out what the tools are for, and they are the most valuable people in the building to understand. The allocation move is to find out which, not to cap them.

What has to be true of the system

A company can do this on spreadsheets for a quarter. To do it every month, the underlying system has to have a few properties.

• One ledger, one shape, every source, so the sales team's API usage and the engineering team's seat usage are the same kind of row.
• Attribution to people by email, retroactive, with an owned queue for the unknown.
• Output from the CRM beside spend, per person, by role, with the fields that count configurable, because every company's definition of "closed" is slightly different.
• Budgets that are thresholds with owners, monthly, that reach the owner where they work and record what the owner decided.
• A short list of findings, generated without anyone asking: spend far above a role's median, spend with no output beside it, spend nobody owns.
• An audit trail, because allocation moves money between people and someone will ask who moved it.

None of this reads a prompt. Usage metadata, people, and outcomes are enough, and a system that wants more than that is asking for something it does not need.

The next step after this one

The four questions above are the first cut: descriptive, honest, and buildable now. What follows them is harder and more valuable: how confident are we that the AI caused the change in output, what was the marginal return on the marginal dollar, and what happens if we act, by raising a budget, restricting a model, or routing a workflow to a cheaper one. Those need more history and more care, and they are where this discipline is going.

But they are built on the first four, and the first four are built on a ledger and a join that most companies have not done. That is the work in front of a finance leader today: not to spend less on AI, but to know what it is buying, per person, and to move it accordingly.

Arbitir is the first four questions as a product: the ledger across Anthropic and OpenAI, attribution to people with an unmatched queue, closed-won from HubSpot or Salesforce beside spend per person, monthly budgets with Slack notices, three nightly findings, and an audit row for every change.