How to Calculate the ROI of Enterprise AI Agents

Enterprise AI agent ROI is not the model vendor’s price list. It is the change in a business flow you already measure, minus the cost of a production-grade runtime. If you cannot name the flow, you do not have a business case — you have a lab.

The equation

Annual value ≈ (hours recovered × loaded cost) + (errors avoided × cost per error) + (cycle-time value you can evidence). Annual cost ≈ platform + integration + evaluation + HITL time + on-call. ROI = (value − cost) / cost. Payback is months until cumulative value covers cost.

Worked example (illustrative)

1. Flow: integration exceptions between CRM and ERP, 40 hours/week of specialist time, loaded cost $90/hour → $187k/year.
2. Agent + factory: recover 50% of that time after 90 days ($94k/year run-rate) and cut repeat errors by $40k/year in rework you already track.
3. Cost: $180k first-year (integration, AgentsAI operating cost, evals, 0.3 FTE HITL). Year-1 value ~$100k (ramp) → revisit scope. Year-2 run-rate can turn positive if the flow was real.
4. Decision: if the baseline hours were theater, stop. If they are in the GL, continue and tighten allowlists.

Notice what is missing: “productivity up 10x” and token spend as the only line. Those slides do not survive a controller. Time-to-value — idea to a metric on a P&L or risk register — is the metric we use in factory conversations.