What a head of AI should put in front of the board each quarter

The head of AI role is new enough that most people holding it have never reported to a board on it, and most boards have never asked. That is ending. The AI line is growing faster than any other, the press is full of both productivity claims and disappointments, and directors want to know which one their company is. The report they need is one page, five numbers, and a short list of what is being done about them. This is what goes on it.

The five numbers

Total AI spend for the quarter, by shape. API usage, enterprise seats, and plans on cards, as three figures and a total, with the same three for the prior quarter. The shape split matters because directors will assume the contract is the whole bill, and it usually is not.

Spend per person, as a distribution, not an average. The median, the top decile, and the number of people above three times the median for their role. An average hides everything interesting. "Most people spend under fifty dollars a month and eleven people account for sixty percent of the bill" is a sentence a board remembers.

Spend beside output, for the roles that have a countable output. AI cost per closed-won dollar for account executives, per meeting booked for SDRs, per renewal for customer success, per merged pull request for engineers. As a ratio per role, with the spread across people. This is the number the board actually came for, and it is honest as long as it is presented as what it is: spend next to output, not spend causing output.

Unattributed spend, and its trend. The share of the bill that cannot be tied to a person or a team. It should be small and falling. If it is large, the rest of the report is built on sand and the board should hear that from you rather than work it out.

Budget outcomes. How many team budgets exist, how many were exceeded, and what was decided each time: raised, investigated, fixed. A quarter of "exceeded and raised because the output was there" is a strong position. A quarter of "exceeded and nothing happened" is a governance gap, and better named than found.

The three to leave out

Hours saved. Every AI vendor's deck has them, every board has seen them, and nobody believes them, because they are self-reported and unpriced. If a rep saves two hours a day, the number that matters is what the rep did with them, and that is the output number above.

Adoption percentage. "Eighty percent of employees have used AI this quarter" is a fact about logins. It says nothing about value and it rewards the wrong thing. Show usage as spend per person instead, which at least says who is doing how much.

A single ROI figure. "AI delivered a 340 percent return" cannot be defended on a year of data with no control group, and a director who asks how it was calculated will not like the answer. Show the ratios and the trend and let the board draw the conclusion. It will draw it.

The short list of actions

Below the numbers, three to five sentences on what is being done, each tied to a number above.

• The eleven top spenders are being interviewed to understand what they are doing that works, and two of their workflows are being written up for their teams.
• The enterprise contract renews next quarter; twenty-three idle seats are being dropped and the seats reallocated to the people currently on personal plans.
• The sales team budget was raised after closed-won per rep moved with spend; the marketing budget is being investigated because it did not.
• Unattributed spend fell from eighteen percent to six percent as shared keys were split; the remaining six is three service accounts now marked as infrastructure.

That is a page a board can read in four minutes and ask good questions about.

What it takes to produce it

None of these numbers exists in any one system. Spend is at the providers, in three shapes. People are in the CRM and the identity provider. Output is in the CRM and the engineering tools. Budgets and decisions are wherever someone wrote them down. Producing the page means having joined them, per person, by email, and having kept the join current every night for a quarter.

The companies that can produce this page are the ones that did the unglamorous work first: one ledger, attribution to people, an owned queue for the unknown, output beside spend, budgets with owners. The page is the reward for it, and it is the thing that turns the head of AI from the person who spends the money into the person who allocates it.

A note on tone

Boards respond to candor. A report that says "we spent this, most of it is here, we can see it next to output for these roles and not yet for those, here is what we are doing" is more credible than one that claims a return. The goal of the first few quarterly reports is not to prove that AI is working. It is to establish that the company knows what it is spending and who is producing with it, which is the position from which every later claim can be made.

Arbitir produces the numbers on this page from a company's own connections: spend across Anthropic and OpenAI by shape and by person, the distribution by role, spend beside closed-won and meetings from HubSpot or Salesforce, the unattributed share, and every budget event with what was decided.