AI business cases fail when they lead with technology. Finance does not fund embeddings. Finance funds margin, retention, cost takeout, or risk reduction. I start from the business line the sponsor already cares about and work backward to the model.
I use three buckets: revenue or retention uplift, cost or time saved, and risk avoided. Every AI use case must map to at least one bucket with a measurable proxy. If the proxy is soft, I label it soft and do not pretend it is hard ROI.
Pilot math is different from scale math. I show both. A copilot that saves ten minutes per rep per day is interesting in a pilot and material at ten thousand seats. I also show inference cost at scale, because nothing kills an AI program faster than unit economics that work in demo and collapse in production.
The other half of the case is option value. Some AI bets unlock future products rather than immediate P and L. I name that explicitly instead of hiding it inside inflated benefit lines. Sponsors respect honesty. They do not respect spreadsheets that only go up.
When I ran adoption programs at Atlassian, the story that landed with leadership paired usage quality with downstream engagement metrics, not vanity counts. The business case stayed alive because I could show movement in numbers that mattered to the P and L owner, not just to the AI team.