Does this AI use case actually pencil?

How to get a number you can defend

An ROI model is only as honest as the inputs, and four of them are routinely set too generously. Getting these right is the difference between a business case that survives a finance review and one that doesn't.

Deflection rate — start lower than feels right

The share of contacts fully resolved without a human is the input people inflate most. A well-built system on a clean knowledge base handles a meaningful portion of tier-1 volume, not all of it. Model the conservative case first; if the project only works at an optimistic rate, it doesn't work.

Loaded cost, not salary

Use the fully loaded hourly cost — benefits, tooling, management overhead, and the cost of recruiting and training replacements. Using base salary alone understates the saving substantially, which is one of the few places people are accidentally too pessimistic.

Handle time should be the median, not the average

A handful of two-hour escalations drag the mean well above what a typical contact actually costs. Those long tickets are also the least automatable, so an average built from them flatters the model twice over.

Subtract what it costs to run

Inference is an ongoing cost, and so is the human time spent reviewing escalations and maintaining the eval suite. Net saving is what matters. The cost estimator gives you the running side of that equation.

If the case only clears the bar under the best-case assumptions on every input, that is a signal to scope smaller rather than to proceed — and it's cheaper to learn now than after the build.