What are you building the case for?
This becomes the title of your case. Be specific enough that someone else in the building would recognize it on sight.
Transactions, documents, cases — whatever the countable unit is here. A round number is fine.
A few quick ones — these decide whether AI is even the right tool for this work, before anything gets priced.
Pick every one that's actually true here. The case at the end only shows what you select.
What this tool actually tests.
How do you check an AI initiative's output before trusting it?
With a check that fails differently than the model does. Asking the model if it's sure, or having a second model grade the first one's output, tests nothing — both share the same blind spot as the thing being checked. A real check is a deterministic recomputation, a fact the model didn't produce, or the same claim recovered a genuinely different way. Anything else is ceremony, not verification.
What should you own in an AI vendor contract?
The evaluation sets, labelled exceptions, and failure taxonomy your team generates during the engagement — named explicitly as your deliverables, not folded into generic "customer data" language. That's the artefact that's actually expensive to rebuild if you switch vendors. Also worth a direct answer from any vendor: name the data they have that you structurally cannot obtain yourself. If the answer is "we've learned from our customers," that's not a data asset, it's a description of what they're building out of your process.
How do you calculate ROI for an AI project honestly?
By counting what actually stopped happening, not what got faster. A step that runs quicker with a model attached and a step that gets deleted entirely produce the same demo and a completely different return. The honest number prices the deleted steps only, discounts the estimate because self-reported time typically runs high, and states plainly when the answer is zero rather than manufacturing one.
Is this AI initiative even the right problem for AI?
Only if there isn't one correct, computable answer sitting under it already. Work with a known formula, an optimization, or a physical control loop is a deterministic problem wearing an AI initiative's clothing — decades of control theory already solves it for less money and fewer surprises. AI earns its place on work that's genuinely judgment-based: extracting structure from messy documents, reasoning over incomplete information, work where there isn't only one right answer.