Software was purchased by seat, project, contract, or infrastructure footprint. AI introduces a more fluid unit: cognition-like work consumed in variable amounts by people, applications, and agents.
01 / The Signal
The cost moves when the work moves.
A human may initiate one task that triggers research, tool calls, model routing, retries, evaluation, and follow-up work across several agents. The visible action is small. The resource chain behind it may not be.
At the same time, AI cost does not map cleanly to traditional headcount, SaaS seats, or cloud infrastructure. It can be centralized or embedded, fixed or variable, interactive or autonomous. The cost can sit inside a vendor subscription, an API bill, a workflow platform, a product margin, or an employee expense account.
Conceptual basis
This note develops Karl Ohlemann’s prior public CFO framing: the mandate is not simply to spend more or less on AI. It is to govern intelligence as a new economic input whose consumption, authority, and business value need to be visible.
02 / Why It Matters
Unmeasured intelligence becomes unmanaged operating leverage.
AI can reduce cycle time, improve decision quality, expand capacity, or create new revenue. It can also generate plausible work no one needed, repeat expensive loops, duplicate tools, and conceal cost inside apparently efficient automation.
The finance function will be asked a familiar question in an unfamiliar form: what are we getting for the money? Answering it requires more than token counts. A low-cost workflow can still destroy value if it acts on poor data or without authority. A costly workflow can be rational if it materially improves a high-consequence decision.
Cost control therefore cannot be separated from system design. Who can initiate work, which model is selected, what tools the system may use, how long it can run, and when a human must approve continuation are financial controls as much as technical ones.
03 / Second-Order Effects
The management accounting of work will change.
Budget ownership becomes ambiguous.
AI may sit inside every department while being procured through technology, operations, finance, or individual teams. Without a deliberate structure, nobody owns the complete cost and everyone believes someone else is governing it.
Unit economics reach into the workflow.
Businesses will need to understand AI cost per useful outcome: per reconciled account, qualified lead, resolved case, released feature, reviewed contract, or completed decision. The denominator matters more than the raw compute line.
Model routing becomes a financial policy.
Not every task needs the most capable model. Routing by risk, complexity, latency, privacy, and cost can preserve quality while preventing premium intelligence from becoming the default for low-value work.
Autonomy creates contingent spend.
An agent that can create work for another agent can also create cost recursively. Rate limits, budgets, run ceilings, approvals, and stop conditions become the equivalent of delegated purchasing authority.
04 / BlackBoxx Take
The ledger needs to know who—or what—made the decision to spend.
Karl’s analysis
Finance should not begin by treating every token as a problem. It should begin by making the flow legible: intent, actor, authority, model, tools, cost, output, reviewer, and outcome. Once those are connected, the organization can make rational tradeoffs instead of imposing blunt limits.
The CFO will not govern AI by approving a software invoice. The CFO will govern the operating system that creates the invoice.
A useful control model has three layers. First, visibility: know where consumption occurs. Second, authority: define who and what may incur it. Third, value: connect spend to an outcome with enough context to judge whether it was worthwhile.
This is not surveillance for its own sake. The objective is economic clarity. Good governance should let valuable work move faster because the boundaries are known.
05 / What I’m Watching
Whether AI cost management grows into operating governance.
- Cost attribution by workflow, customer, product, and business outcome
- Budgets and approval limits that apply to agents as well as employees
- Model routing policies that account for risk and quality, not price alone
- Duplicate subscriptions and invisible consumption embedded in SaaS products
- Audit trails that connect autonomous actions to a responsible human owner
- Gross-margin effects where AI is part of the delivered product
- Whether finance dashboards measure useful work or merely count activity
Sources and boundary of analysis
The finance-and-AI theme was raised publicly by Karl on LinkedIn on August 6, 2026 and expanded here through BlackBoxx operating-system work. This governance framework does not assert vendor market share, aggregate AI spending, or guaranteed productivity outcomes. It is intended to separate measurable economics from hype.