Let’s be direct about something most AI writing avoids. Deploying AI agents is not free. Every agent interaction consumes tokens — the unit of computation that AI providers charge for. At small scale this cost is invisible. At enterprise scale, across departments with multiple simultaneous initiatives, each running multiple specialized agents, it becomes a significant and poorly-attributed budget line that will eventually land on someone’s desk with a question attached: what did this produce? Organizations that cannot answer that question cleanly will find their AI budgets quietly reduced. This article is about how to build the governance model that keeps the answer ready.

The Multiplication Problem.

The series has discussed span of control — the idea that a manager who once supervised eight people may soon coordinate eight people plus twenty-five AI agents. That framing is useful at the individual level. Now scale it:

None of that arithmetic is hypothetical. Organizations beginning serious AI deployment are already encountering it. The question is whether they built the governance to manage it before they encountered the bill, or after.

What an API Key Actually Means.

When an organization connects to an AI provider — Anthropic, OpenAI, Google, or another — it uses an API key. That key is the billing relationship. I have them, but I also have a Pro account against which tokens can also be drawn. Using this method is predictable and potentially less costly. Every prompt, every agent run, every retrieval, every multi-step reasoning chain is metered in tokens and charged against that key’s account, or Pro usage limits. Understanding three things about this matters for organizational governance:

The Chargeback Vacuum.

Organizations solved this problem in a different domain twenty years ago. When cloud computing matured, Finance discovered that infrastructure costs were accumulating in IT budgets with no connection to the business outcomes producing them. The answer was showback and chargeback models — attribution systems that traced cloud spend to the department and initiative generating it. AI agent costs need the same solution, and most organizations have not built it.

The governance vacuum looks like this: Marketing deploys twelve agents for a campaign. Operations deploys eight agents for a process audit. HR deploys six agents for workforce analytics. All of them run against the same API key. End of month, a single invoice arrives. Finance sees the total, has no attribution model, and labels the entire line as “IT infrastructure.” The CEO asks how AI is performing. Nobody can answer — not because the work wasn’t done, but because the cost and the value were never connected.

The Agent P&L.

The governance mechanism that answers all three CFO questions is treating each agent cluster as a cost center with its own accountable P&L. Not at the individual agent level — that granularity is impractical. At the initiative and department level: a defined set of agents deployed for a defined purpose, with tracked input costs and measured output value.

The right column is not science fiction. It is the organizational infrastructure that every company deploying AI agents at scale will need — most just haven’t built it yet because the costs were small enough not to matter at the pilot stage. Scale changes that relationship abruptly.

Span of Control at Agent Scale.

Return to the span-of-control question and apply the economics to it. In the traditional model, every direct report costs a known amount and produces visible work. The manager’s job is to direct that work and hold individuals accountable. With agents, the span expands dramatically — but the accountability structure doesn’t automatically expand with it.

A manager with eight employees and thirty-six agents across four initiatives is not running the same job with more helpers. They are running a fundamentally different operational system — one where cost accumulates across many invisible threads, where individual agent failures can be masked by aggregate output quality, and where the question “is this worth what we’re spending?” requires a measurement apparatus that has to be deliberately built. The expanded span without the expanded governance model is where cost and accountability separate.

The Measurement Frameworks That Already Exist.

The good news is that the tools for building this accountability layer are not new. The discipline that this series has referenced throughout — Kirkpatrick’s evaluation model, Phillips’ ROI methodology, Total Alignment, Continuous Performance Improvement — was designed to connect activity with measurable value. These frameworks have always had to answer the CFO’s question in a human-workforce context. They answer it just as well in a mixed-workforce context, with one substitution: where you once tracked training hours and development activities, you now track agent runs and token consumption.

That five-level evaluation applied to agent deployment is the answer to the CFO’s question — and the mechanism that keeps AI investment justifiable through every budget cycle.

The CTO’s New Mandate.

The Chief Talent Officer’s emerging role in the series has been to architect capability — human and digital. Add the economic dimension and the mandate expands further. The future CTO is not just designing the capability architecture; they are helping the organization build the governance model that keeps that architecture financially accountable. That means partnering with the CFO to build the attribution model, with IT to establish per-department key allocation, with operations to define agent sunset criteria, and with every department head to establish value measurement definitions before agents deploy. This is not a technical function. It is an organizational leadership function — which is exactly where it belongs.

What to Do Before You Scale.