Baselines and ROI you can defend

A value case is a forecast with a measurement plan attached. You measure the current process before anything changes, because once the agent ships the old process is gone and the baseline cannot be recovered honestly. You write a value hypothesis that could be proved false. You build the model from drivers, each with a source: volume, adoption, minutes saved, errors avoided, minus every cost of owning the thing. Then you discount it, find the payback month, and ask which assumption the answer depends on most. That assumption is what you measure next, and it is what benefits tracking watches after go-live.

• Baseline first or never. A baseline measured after go-live is a reconstruction, and finance will treat it as one.
• Time saved is not cash saved. Freed hours become money only through fewer hires, less overtime or contractor spend, or redeployment to measured work. Apply a realisation rate that finance agrees.
• The human in the loop is a cost. Review and approval minutes come off the gross saving.
• The base case is not the expected case. When most ranges are skewed against you, the median outcome sits well below the base case.
• The tornado tells you what to measure. The widest bar is where the next week of evidence-gathering should go.