Measuring whether it was worth it, honestly. Time saved is the usual claim and the hardest to substantiate, because saved minutes only become value if they are redeployed to something useful. The credible measures are outcome measures: more cases resolved, fewer errors, shorter cycle times, revenue.
Technical view
Build the cost side from inference and infrastructure, integration and maintenance engineering, evaluation and monitoring, licences, human review time and change management. On the benefit side prefer measured outcomes from a controlled comparison against a baseline over self-reported time savings, which are consistently overstated. Include the counterfactual: some of the improvement would have happened anyway, and attributing all of it to the project is the standard error.
Related
Part of AI Terminology Index, section: Product and Adoption.