The mental model that unlocks everything: a token is a unit of production cost, like a kilowatt hour or a billable minute. Treated as overhead (one big bill, no attribution), AI spend is uncontrollable and every finance review becomes a fight about the total. Treated as cost of goods (attributed per workflow), it becomes ordinary unit economics: this workflow costs $0.40 per processed ticket and replaces $6 of labor; that one costs $30 per report and saves an afternoon.
Attribution requires tagging usage by workflow and team at the API layer, which every serious provider supports. If you cannot see spend per workflow, you do not have a cost problem, you have a metering problem, and it is fixable in a week.
Attribute value at the workflow level, never the token level. Three honest measures, in descending rigor: displaced cost (hours saved times loaded rate, the workhorse metric), throughput (same team, more output, measured in units), and quality delta (fewer errors, faster cycle time, measured against the pre-AI baseline you instrumented). A workflow with a token bill and no baseline is unpriceable, which is why measurement precedes automation in every mature program.
Full ROI scope also counts the costs that are not tokens: the build, the harness, the evaluation time, the review time. A workflow that saves $10k monthly in labor but consumes $8k of senior review is a different investment than its token bill suggests.
One page, monthly: each production workflow with tokens spent, cost per unit of output, value per unit (displaced cost, throughput, or quality delta), and trend. Workflows without a value line are pilots; pilots get ninety days to earn one.
The failure you actually fear is not high usage; it is runaway usage: a looping agent, a misconfigured batch job, a prompt-injected session burning spend on garbage. So cap the failure modes precisely. Per-session ceilings that halt and escalate rather than silently fail. Per-workflow monthly budgets set at three times observed baseline, alerting at one and a half. Anomaly alerts on rate-of-change, because runaway sessions announce themselves in minutes, not months. Employee-level limits high enough that no one doing real work ever encounters them.
The distinction to defend in the budget meeting: limits exist to catch malfunctions, not to ration a productivity tool. Rationing a tool that returns multiples on labor cost is the only genuinely expensive option on the table.
Run AI spend like an investment portfolio, monthly, on one page: top ten workflows by spend, each with cost per unit, value measure, and trend. Kill or fix the ones underwater. Feed the winners. Investigate anomalies. Twenty minutes, same discipline as ad spend, and it permanently ends the "AI is expensive" conversation by replacing it with line items.
Meter spend per workflow and team, cap failure modes (per-session ceilings, anomaly alerts, budget alarms), and review a one-page Token P&L monthly. Attribution plus tripwires beats blanket limits.
Set them high enough that productive work never hits them, and rely on session-level caps and anomaly detection to catch malfunctions. Limits that ration normal usage cost more in lost productivity than they save in tokens.
Per workflow: displaced labor cost, throughput gain, or quality improvement against an instrumented baseline, minus the full cost side (tokens, build, evaluation, review time). Portfolio-level ROI is the sum of workflow lines, not a feeling about the bill.
Bring one AI question from this article. We will tell you what we would do, whether or not you hire us.
Book a 15-Min Chat →