An AI business case should explain the operating problem, compare viable options, quantify benefits and full costs, identify dependencies and risks, and state what evidence will justify investment. A model is useful when another decision-maker can inspect its assumptions and reproduce the calculation.
Use this public template to prepare a first draft. The example is fictional; it demonstrates a method rather than an expected return from Artebello or a specific technology.
The executive decision page
| Field | What to record |
|---|---|
| Decision requested | The funding or approval needed now, and the limit of that approval |
| Problem | Who experiences it, how often, and which outcome is affected |
| Baseline | Measured volume, time, quality, cost, and current process owner |
| Options | Current process, simpler process/software changes, and the proposed AI approach |
| Proposed scope | The task, user group, information boundary, and explicit exclusions |
| Expected value | Cash effects and operational capacity, reported separately |
| Full cost | Setup, recurring usage, integration, review, training, support, and transition |
| Dependencies | Data, permissions, staffing, vendors, governance, and integration |
| Risks | Consequences of errors, required controls, owners, and residual issues |
| Evidence gate | What the pilot must demonstrate before expansion |
Writing the decision page first helps expose ambiguity. “Improve productivity with AI” does not say which work will change or who will validate the result.
Benefits without double counting
Estimate capacity using measured task volume and net time saved. Net time saved includes the work that remains: review, correction, exception handling, and administration. Classify the result as avoided cash cost, redeployed capacity, faster service, improved quality, or another clearly described outcome.
Only treat capacity as cash saving where an operating plan can actually avoid paid cost. Count additional revenue at contribution after variable costs and distinguish genuinely additional demand from demand moved between channels. Benefits from the same change should not appear in multiple lines unless they are separate effects.
The ROI measurement guide provides a wider value framework. The voice-ordering model shows how an attractive gross saving can become a weak investment after realization and full cost are included.
Worked example: internal knowledge assistance
Assume a fictional team performs 5,000 information-search tasks per month. A pilot proposes a net saving of four minutes per task after verification effort. That would release approximately 333.3 hours per month. At an assumed loaded rate of $24 per hour, its gross capacity value is $8,000.
The base case assumes only 25% of that capacity replaces paid cost, so the model counts $2,000 per month as realizable cash benefit. It reports the remainder as potential redeployment capacity. Operating cost is assumed to be $1,200 per month, and one-time setup and training cost is $9,600.
Monthly net cash benefit is $2,000 − $1,200 = $800. Simple payback is $9,600 ÷ $800 = 12 months. First-year benefit is $24,000 and first-year cost is $24,000, so first-year ROI is 0% under those assumptions.
This example should prompt questions. Can the team remove paid cost, or will it simply search less slowly? Will quality improve? Is the four-minute reduction sustained outside a pilot? Does the operating-cost estimate include ongoing document ownership and monitoring? The investment decision changes when the answers change.
Sensitivity table
Holding the assumed $8,000 capacity value and other costs constant:
| Realization share | Monthly cash benefit | Monthly net cash benefit | Simple payback |
|---|---|---|---|
| 10% | $800 | −$400 | No positive payback |
| 25% | $2,000 | $800 | 12 months |
| 40% | $3,200 | $2,000 | 4.8 months |
Show the downside alongside the base case. Where a project’s value is primarily service quality or resilience, state that rationale and measure it directly rather than disguising it as guaranteed savings.
Pilot decision gates
Define the task set, baseline, comparison period, user roles, and success measures before testing. Include difficult and unsupported requests, not just questions the system is expected to answer well.
Example gates for a knowledge assistant could include verified source references, acceptable task completion time, permission tests, and a documented escalation process. Each gate needs an owner and an agreed threshold appropriate to the use case. A good average result does not cancel a serious permission failure.
Record four possible decisions: continue within scope, redesign and retest, expand after further controls, or stop. This prevents the pilot from becoming an automatic commitment to rollout.
Assumptions register
For each material assumption, record the value, unit, source, date, owner, confidence, and next validation step. Keep observed results separate from forecasts. Update the model when the pilot provides evidence; preserve the original version so the team can compare forecast with actual performance.
Public template and advanced support
Download the editable business-case worksheet. It includes the decision page, assumptions register, cost and benefit lines, and pilot decision record.
The public worksheet is available without an account. A facilitated opportunity review or an organization-specific investment case can be scoped as consulting work. Explore AI opportunity mapping or book a discussion.
Frequently asked questions
What ROI formula should the case use?
State the period explicitly and calculate (period benefit − period cost) ÷ period cost. Include one-time cost in the relevant period. Use the organization’s finance conventions consistently for discounting, tax, and investment approval.
Can we include employee time as value?
Yes, as clearly labelled capacity. A cash-saving claim also needs evidence that the organization can avoid paid cost. Reporting both separately is more useful than applying one optimistic percentage to every benefit.
What should we do before funding?
Review the readiness checklist, inspect the most consequential assumptions, and agree a bounded evidence gate before expansion.