# AI agents can return multiples of their cost. > Model productivity, avoided AI spend and programme cost. High-volume agent workflows can reach 500–1,500% ROI when the same capability runs repeatedly at scale. ## Change the assumptions. See what must be true. The defaults are illustrative. Replace them with your function’s actual baseline. **Illustrative annual case** — Illustrative model using 48 working weeks. Returned capacity is not cash savings unless the organisation converts it into lower cost, more output or avoided hiring. - People in scope (people) - Hours returned / week (hours) - Loaded cost / hour ($) - Expected active adoption (%) - Capacity converted to economic value (%) - Annual platform + rollout cost ($) - Annual tools / model cost avoided ($) ## Do not scale a forecast. Scale a measured result. - **Baseline** — Measure current volume, elapsed time, manual touches, error or exception rate and tool cost. - **Supervised launch** — Run the workflow with human review and record adoption, quality, model usage and exceptions. - **Expansion decision** — Scale only when the agreed KPI improves and the control model remains acceptable. ## Route work by quality, policy and cost. Use efficient models for routine work, reserve premium models for complex reasoning, set budgets by team or workflow and find duplicated external subscriptions. ## The full cost of the current process. Employee time, delays, errors, AI seats, token usage, integration maintenance, infrastructure and enablement all belong in the baseline. ## Build the ROI case for one real workflow. We will define the volume, current cost, human controls and KPI required for an expansion decision.