Support Business Case Creation

Whose Document Is This?

The business case is the document an investment committee approves, defers, or kills. On most artificial intelligence (AI) initiatives it is owned by the sponsor or by finance, not by you — which is why ECO Domain II, Task 9 is worded "support business case creation." You supply the evidence, translate the technology into language a finance reviewer can test, and act as the reality check that stops the organization funding a benefit the model cannot deliver. Task 5 (Determine return on investment) produces the numbers; the business case packages them into an investment decision with options, risks, and an ask.

Gathering Financial Data and Projected Benefits

Nothing you hand to the case should be a figure you invented. Every input needs a named source who will defend it in the room:

• Baseline operational metrics — from the process owner: current volume, cycle time, error rate, rework rate for the decision you are improving. Without a baseline there is no measurable benefit, only an assertion.
• Unit economics — from finance: fully loaded cost per hour, cost per claim, cost of a false accept.
• Benefit logic — from you and the subject-matter expert (SME), as one causal chain: the model flags X% of cases, the reviewer handles them Y minutes faster, Z hours are released per year, valued at the finance unit rate.
• Cost quotes — from vendors, cloud pricing calculators, and your own Phase 1 sizing. Date them; license and compute pricing moves.
• Precedent — results from the same AI pattern elsewhere. A predictive-maintenance case is more credible citing a comparable deployment than an unsourced industry statistic.

Working with Finance on Costs and Projections

Finance owns the cost model, but their standard IT template misses what makes AI different: cost does not stop at go-live. Walk them through the lines they do not normally carry — data acquisition, licensing and remediation; annotation and labeling, frequently the largest single item in Phase 3 Data Preparation; bursty training compute versus recurring, volume-driven inference compute; the machine learning operations (MLOps) platform, monitoring and drift detection; periodic retraining, because a model is not one-and-done; human-in-the-loop review staffing; change management and documentation; and decommissioning cost.

Then ask finance two questions before any numbers are built. What hurdle rate (or discount rate) must the case clear, and over what horizon? And does the committee accept capacity released and cost avoided as benefit, or only cash that leaves a budget line? If only cash counts, soft benefits move to a strategic annex instead of into the net present value (NPV) — far better than a case that clears the hurdle on paper and then fails its post-implementation review.

Structuring the Executive Narrative

Executives approve stories with numbers in them, not spreadsheets. Six moves, in order:

1. The problem in business terms. Not "we will deploy an anomaly-detection model" but "8% of card transactions are held for manual review, 71% of those turn out legitimate, costing 14,000 analyst hours a year."
2. The decision being improved. Name it: who makes it today, how often, on what information, and how well. An AI project that cannot name a decision is usually a reporting project.
3. The option set, including the do-nothing option. Baseline, a process or rules fix without AI, buy, build. A single-option case reads as advocacy and committees discount it.
4. The ask. Money, people, elapsed time — and specifically what the first tranche buys.
5. The risks, in the same units as the benefits: data availability, adoption, regulatory exposure, performance uncertainty.
6. The gate structure. Stage the ask across CPMAI phases: fund Business Understanding through Data Preparation with a data-sufficiency go/no-go before the build is funded. A bounded first spend with a defined kill point is the most persuasive thing you can offer a nervous committee.

Providing Technical Validation

Your distinctive contribution is catching what a finance reviewer cannot. Three failure classes recur.

Claims the model cannot support. A correlational model does not explain why customers churn, a population-level accuracy figure does not guarantee any individual outcome, and a batch-scored design does not deliver a real-time promise. Strike causal and guarantee language out of the case.

Unrealistic accuracy assumptions. Watch for benefit arithmetic built on near-perfect accuracy, on accuracy holding equally across every segment, or on the best point of the performance curve rather than the operating threshold the business will run at. Ask which error costs more — a false positive or a false negative — and confirm the benefit model prices them differently.

Double-counted benefits. The 6,000 technician hours in your case may already sit inside an approved automation program's case. Reconcile against other in-flight initiatives and present the net incremental benefit.

Reviewing and Refining the Case

The case is a living artifact, not a one-time submission. Estimates firm up phase by phase: after Phase 2 Data Understanding you know acquisition and remediation cost; after Phase 3 you know annotation effort; after Phase 5 Model Evaluation you know real performance and therefore the real benefit rate. Keep one controlled copy with a dated version number, a change log, and an assumption register listing each assumption with its owner, confidence and review date. When an assumption breaks, the case goes back for re-approval — it is not quietly edited.

When the Sponsor Asks You to Move the Number

The exam tests this directly. A sponsor sees the case land just under the hurdle rate and asks you to raise projected accuracy from 82% to 92% so it clears. You neither restate a number you cannot defend nor turn the moment into an integrity confrontation. Present the sensitivity range and the assumptions: benefit at conservative, expected and optimistic accuracy, what would have to be true for the high case, and what evidence would settle it. Then offer the move that usually works — a smaller, phase-gated ask that funds the data work and comes back with a real number. The committee keeps the decision, and every figure with your name on it stays defensible.