Some growth investors now count AI spend as headcount, converting a company’s annual AI spend into full-time equivalents (FTEs) when they evaluate the business. The comparison holds because AI is a large recurring cost that takes on work people used to do, and it breaks because AI spend is variable and lacks a shared productivity model. Cost per outcome, marginal cost per unit of work, and adoption breadth are the three numbers that make the comparison defensible in a board meeting.
“[Our lead investor is] now adding AI spend as additional FTEs. If you spend $300,000 a year on Claude Code, they’re adding two additional FTEs to your account.”
That’s how an executive at one portfolio company described the new method their lead investor, a major growth equity firm, uses to evaluate the business, with AI spend dropped straight into the same headcount model it applies to people.
Boards that haven’t raised this yet probably will. Growth investors borrow evaluation methods from each other, and a conversion this simple travels quickly. The companies that walk in with their own numbers get to test the board’s math against real data before anyone treats it as settled.
Why do investors treat AI spend like headcount?
AI spend shares two traits with headcount that boards care about, since it’s a large recurring cost tied to output capacity and it can take on work that people used to do.
Some organizations have already formalized the comparison by managing agents like staff, with regular performance checks and the option to retire an agent that isn’t working. That’s a headcount process applied to software.
The framing also exposes underuse. Every time someone does by hand what the tooling could have handled, the company pays a human hourly rate for work that could have cost a token rate, and the headcount frame is one of the few views that turns that idle tooling into a dollar figure. Most cost reports leave it out.
Where does the headcount comparison break down?
AI spend moves with demand. A person costs the same in a busy quarter and a slow one, while an agent costs whatever it consumes, so AI spend behaves like cost of goods sold even though most companies budget it like payroll. The difference lands in the margin line. Layer variable AI cost onto a headcount model and the margin structure stops holding, because an extra million dollars a year in agent spend only works if it brings in incremental revenue.
AI has no shared productivity model. Decades of imperfect conventions exist for evaluating a person, and nothing comparable exists for AI yet. The usual substitute is token consumption, which does real damage once it becomes the proxy, because developers learn to burn tokens so they look busier and the metric ends up rewarding the wrong behavior.
AI spend can change faster than anyone plans for. Hiring takes months. Severance follows a predictable schedule. Agent spend can double in a month without anyone approving the increase, and substitution often happens as an accounting outcome with no plan behind it, because token spend has to come from somewhere and people cost is usually where it comes from.
What three numbers should you show the board?
Three numbers let the headcount comparison hold up under scrutiny.
Cost per outcome. This is the AI cost of one unit of finished work, such as a resolved ticket or a qualified lead. Because it’s measured in the same units as a person’s work, it’s the metric that makes the FTE comparison legitimate. Cost per token and cost per seat can’t support that comparison.
Marginal cost per unit of work. The trend tells a board more than any single reading. If later units of work cost less than earlier ones, the investment is compounding. If they cost more, there’s a scaling problem the headcount frame will hide, since people don’t get more expensive per unit the more a company uses them.
Adoption breadth. This measures underuse directly, covering the share of the eligible team using the tooling and the spread between the heaviest and lightest users. Concentrated usage means the company paid for organization-wide capacity that a handful of people consume.
Some boards already ask for a version of this, usually token usage by engineer mirrored against pull requests. That metric carries the incentive problem described above. Pairing it with an outcome measure corrects for it.
What should you bring to the board meeting?
We recommend one page with three charts.
The first shows cost per outcome over the last four quarters, with the outcome defined in terms the business already uses. The second puts total AI spend next to the headcount equivalent the board is applying, so both numbers sit side by side before anyone has to ask. The third shows adoption breadth as a distribution, because an average hides concentrated usage.
Expect the follow-up to be what the same output would have cost in people. When that figure is higher than the AI spend, the headcount comparison works in the company’s favor and there’s little reason to argue with it. When it’s lower, the marginal cost trend has to show AI cost per unit falling toward the people cost, and that’s the version of the meeting worth preparing for.
How does Revenium show you these numbers?
Revenium is the AI Economic Control System, the system of record for AI economics, and it attributes each AI transaction to the team, agent, and business outcome behind it. All three numbers come from that transaction-level attribution data, which most companies already generate and don’t yet aggregate.
Analyze ROI & Unit Economics is where cost per outcome and marginal cost get built. AI Outcomes tracks the business result of each agent run, so the outcome side of the calculation comes from measured results. Attribute Spend to Teams & Departments handles the breakdown by team that boards ask for once they see a total. AI Assistants covers adoption breadth for coding tool spend, and Build Custom Dashboards & Charts turns the three charts into a live view that doesn’t need rebuilding every quarter.
A page built ahead of time gives the team room to check the numbers, and to settle on its own FTE conversion, before the board sees them.
Frequently asked questions
What does it mean to count AI spend as headcount?
It means converting a company’s annual AI spend into full-time equivalents (FTEs) and adding them to the headcount model an investor or board uses to evaluate the business, so AI tooling gets judged on the same terms as the people it supports or replaces.
Why doesn’t the headcount comparison fully hold for AI spend?
AI spend moves with demand and behaves like cost of goods sold, while headcount is a fixed cost, so treating the two the same way distorts margins. AI also lacks a shared productivity model, and its spend can double in a month without anyone approving the increase.
What is cost per outcome for AI?
Cost per outcome is the AI cost of one unit of finished work, such as a resolved support ticket or a qualified lead. Because it’s measured in the same units as a person’s work, it’s the metric that supports a fair comparison between AI spend and headcount.
Is token usage a good measure of AI productivity?
Token usage on its own is a weak measure, because developers learn to burn tokens so they look busier and the metric ends up rewarding activity over results. Pairing it with an outcome measure, such as cost per resolved ticket or per merged pull request, corrects for that incentive.
What should a company show its board about AI spend?
We recommend one page with three charts covering cost per outcome over the last four quarters, total AI spend next to the board’s headcount equivalent, and adoption breadth shown as a distribution across the eligible team.
