Cost modeling

AI ROI theater is everywhere: a pilot's license fee compared against a hand-waved productivity number, presented to two decimal places. Your credibility as the strategy owner rests on modeling costs the way a CFO would — five lines, none skipped:

• Build / setup — configuration, integration, prompt-and-workflow design, testing. For buy-and-configure initiatives this is weeks; for real builds, quarters. Rule of thumb from deployment reality: integration costs 1–3× the license/build cost in year one, and it's the line vendors' ROI calculators omit.
• Run — licenses, per-use model costs (tokens — the chunks of text a model bills by, a few characters each), infrastructure. Usually the smallest line, which surprises boards. Per-unit economics matter more than totals: 'each triaged exception costs ~$0.04 against 11 minutes of agent time' is a sentence that ends debates.
• Data readiness — the line most models omit and most timelines die on. If the initiative needs clean history or organized documents and you have neither, the cleanup is a real cost of this initiative (charge it here, even if the resulting data asset serves five later ones — then say so, as an argument, not an omission).
• People — the owner's time, the SMEs who review outputs during rollout, training hours across the affected team. Change capacity is a cost; pretending it's free is how initiatives ship and then shelf-ware.
• Maintenance — prompts drift, vendors change models, integrated systems update, volumes grow. Budget 15–25% of build cost annually, forever. An AI initiative is a product you now own, not a project you finished.

Three honesty rules

• Model the ramp, not the plateau. Month one runs at half accuracy-of-trust and double supervision. Cost curves front-load; benefit curves back-load. A model showing value from week two is describing a demo.
• Ranges, not points. Every line gets low/likely/high. False precision reads as inexperience to exactly the audience you need — a range with reasoning reads as command.
• Count the option value honestly — in words, not dollars. The first initiative builds capabilities (data plumbing, eval habits, a trained team) that cheapen the next five. Real, important, and not a number you invent — it's a sentence in the recommendation, and it belongs to the portfolio (next lessons), not to any single business case.