Before step one: the mistake this fixes
Most AI business cases get written to justify a technology purchase and then die in finance, because finance can see that the benefit line is a percentage applied to a headcount number somebody guessed. The version that gets funded prices one process, states what it costs to run that process today, and asks for the capacity to change it. Six steps, roughly two weeks of work, and each one produces something you can put on a page.
Step 1. Pick one process that already has a number attached
Not a capability. A process, with an owner outside IT, that somebody already reports on monthly. Invoice exceptions, claims triage, KYC refresh, tier one ticket resolution, contract review. If it has a monthly number in an existing report, half your baseline work is done and the number won't be argued about later. Output of this step: one named process, one named executive owner, and a link to the report that already tracks it. Skip anything where the benefit is described as productivity across the organisation. That case can't be measured, so it can't be defended at the first review.
Step 2. Compute the current cost per completed outcome
Take the volume and the fully loaded time. An illustration: 12,000 invoice exceptions a year at 18 minutes of handling each is 3,600 hours, and at a loaded rate of 60 per hour that's 216,000 a year in handling alone. Add the cost of the errors that get through, plus the rework and the downstream cost of a late payment. Output of this step: one number, cost per completed outcome, with the arithmetic visible. Every later claim gets expressed as a change to this number, which keeps the argument on one axis. Measure it before you build anything. A baseline collected after the pilot starts is worthless, and teams discover this exactly once.
Step 3. Size the change to the process, not the technology
Walk the process step by step and mark each one: removed, shortened, unchanged, or new. AI adds steps as well as removing them, and honest cases show the additions. Review of AI output is a new step. Exception handling for the cases the model routes wrongly is a new step. Model spend is a new line. Then state the target: cost per completed outcome moves from X to Y, and here is the step by step reason. A case that says the same number improves by 30 percent without naming which steps changed is a wish. Output of this step: a two column before and after of the process, with each changed step attributed.
Step 4. Price the delivery path, not the proof of concept
This is where most cases underfund themselves into failure. A prototype costs a few weeks. Getting the same thing into production costs integration into the system of record, an exception path, monitoring, an owner, and a change process for the prompts and models. Teams routinely budget the first and get surprised by the rest, then stall at 80 percent done with no money left to finish. Ask for the delivery capacity: the people, the integration work, and the run cost for twelve months. Name what production means for this process in one sentence, so everyone agrees when it's reached. Output of this step: a funded path from where you are to a workflow carrying real volume, with a date.
Step 5. Put the control cost in the base case, not in a risk appendix
Logging, access control, human review, model and prompt versioning, and the evidence an auditor will ask for are part of running the workflow. Putting them in an appendix invites someone to cut them, and cutting them is what turns a working pilot into something legal won't let you scale. There's a compliance line to state plainly and then move past. If the process falls under the EU AI Act as high risk, Article 12 requires automatic event recording across the system's lifetime and Article 14 requires human oversight an operator can actually exercise, so those aren't optional line items. Article 4, applicable since 2 February 2025, obliges deployers to ensure sufficient AI literacy among the staff operating the system, which means training is a budget line and not a memo. Article 99 sets the ceiling for the most serious breaches at EUR 35 million or 7 percent of worldwide annual turnover. The better argument is commercial rather than legal. The controls are what let you take the workflow from one team to twenty without re-litigating the risk each time. Output of this step: control costs sitting in the main cost table.
Step 6. Write the decision, not the vision
End the document with the sentence the committee is actually approving. Something like: fund a team of five for two quarters to take invoice exception handling from 216,000 a year in handling cost to a target of 120,000, in production, with the audit trail in place, reviewed at the end of quarter one against cost per completed outcome. That sentence gets voted on. A slide about becoming an AI led organisation gets a follow up meeting. Output of this step: one decision, one number, one date, one owner. One stance to close on, and reasonable people disagree with it: don't ask for a platform in your first case. Ask for one workflow into production. The platform argument wins easily once a business unit has a number it likes and wants ten more of them, and it loses badly when it arrives before the evidence. Operators who have taken AI to production at 300,000 organization scale tend to sequence it that way. Difinity partners with regulated teams to find the use cases worth funding, ship them, and put the control in place so the second one is faster than the first.