The P&L of an AI Deployment
What an AI system actually costs. Revenue impact. Token spend. Engineering. Audit costs. The format your CFO reads in 90 seconds and signs off on, or doesn't.
Most AI business cases are written as a savings story: headcount avoided, hours saved, a single line that justifies the program. The savings are usually real and usually one-time. The P&L below is the format that survives the second quarter, because it accounts for the lines the savings story leaves out.
Revenue impact comes first — not the model's accuracy, but the change in revenue attributable to the deployment, measured against a counterfactual. If the team cannot state what revenue moved and against what baseline, the number on the slide is an estimate wearing a decimal point.
Token and inference spend is the line that grows with success. A pilot priced at a flat monthly figure becomes a variable cost that scales with usage, and the unit that matters is cost per resolved task, not cost per call. A deployment that gets cheaper per call and more expensive per outcome is losing money in a way the monthly invoice hides.
Engineering and audit cost close the picture. The build is a capital line; the operating layer is a recurring one. Evaluation harnesses, monitoring, incident response, and the audit evidence the EU AI Act and ISO 42001 require are not overhead to be trimmed — they are the cost of keeping the deployment legible to the people who sign off on it. Strip them and the savings reappear as risk.
The format is one page and the CFO reads it in ninety seconds: revenue delta, token spend, engineering run-rate, audit cost, net. A deployment that cannot be expressed this way is not ready for the board, regardless of how good the demo was. The discipline is the point — the P&L is the artifact that turns an AI project into an AI operation.