AI ROI Business Case

Use this when you need to justify an AI investment with numbers a CFO will trust: value (time saved, revenue lift, quality, risk reduction), full cost including the per-interaction run cost at expected volume, and an honest payback against doing nothing. This is AI-specific because the line that sinks most AI cases is run cost at scale, not build cost. For a generic, non-AI business case, use /business-case-builder instead.

Related skills: Borrows the business-case structure from /business-case-builder. Pull the per-interaction run cost from /ai-cost-model. Place the investment inside a sequenced plan with /ai-transformation-roadmap. For pricing and margin context that strengthens the case, use /unit-economics-analyzer.

The hard part most teams miss

In 2026, enterprises fund measurable outcomes, not pilots. The model is rarely the constraint. Run cost and adoption decide whether the case holds. Three things teams get wrong:

AI ROI is overclaimed on value and underclaimed on run cost. The seductive number is hours saved times a loaded rate. The number that kills the case six months in is the per-interaction cost times real volume. Tokens are cheap once and expensive a million times. Model the run cost at expected scale before you celebrate the value, because that line moves from a rounding error to the dominant cost as usage grows.
Value must be measured against what people actually do today, not a vendor slide. "Saves 10 hours a week" is meaningless without the baseline: how long the task takes now, who does it, how often, and what they do with the time freed. If the freed time does not convert to revenue or headcount avoided, it is comfort, not return. Measure against the real baseline or do not claim the value.
Doing nothing is a real option, and for a given use case it often wins. The do-nothing column is not a formality. Some AI investments do not clear their own run cost plus change-management drag. When that is true, say so plainly. A credible "not yet, here is the threshold that would change the answer" builds more trust than a forced yes.

Process

Step 1: Gather inputs

Ask the user:

What is the AI investment? {{INVESTMENT}} (the feature or capability, one or two sentences, the job not the architecture)
What do people do today? {{BASELINE}} (the current process: who, how long, how often, what it costs in time or money or errors. This is the baseline everything is measured against.)
What value do you expect? {{EXPECTED_VALUE}} (time saved, revenue lift, quality improvement, risk reduced, with rough magnitudes)
What is the build cost? {{BUILD_COST}} (engineering, design, data work, integration, in dollars or people-time)
What is the run cost per interaction and the expected volume? {{RUN_COST}} {{VOLUME}} (cost per call or per token at the chosen model, and daily or monthly interactions with a growth trajectory. If unknown, run /ai-cost-model first.)
What is the adoption assumption? {{ADOPTION}} (what fraction of eligible users or tasks actually route through the AI. Unadopted features deliver zero value at full run cost.)
Who approves, and over what horizon? {{APPROVER}} {{HORIZON}} (who signs, what they care about, and the period the case is judged over, typically 12 to 24 months)

Step 2: Establish the baseline

Convert the current process into a defensible number. This is the denominator for every value claim.

State the task, its frequency, and the loaded cost of who does it today. Use [ASSUMPTION: fully loaded cost of $X/hour] and flag it for validation.
Show the current annual cost of the task as it runs now: frequency x time-per-task x loaded-rate.
If the baseline is quality or risk rather than time, state the current error rate or incident cost in dollars.
If you cannot establish a baseline, mark it and treat the value as qualitative, not financial.

Step 3: Quantify value against the baseline

For each value driver, show the math from the baseline, not from a target.

Time saved: (baseline time per task - AI-assisted time per task) x frequency x loaded-rate x adoption. Apply the adoption fraction. Only count freed time that converts to revenue or avoided headcount, and say which.
Revenue lift: tie to a measured conversion, retention, or throughput change. State the mechanism, not just the percentage.
Quality improvement: convert reduced errors or rework to dollars (errors avoided x cost per error).
Risk reduction: value as avoided cost (incident probability x incident cost), labeled as avoided, not earned.

Keep value that cannot be quantified in a separate qualitative list. Do not invent numbers for it.

Step 4: Model the full cost

Build, run, maintain, and change. The run cost is the one teams forget to scale.

Build: one-time engineering, data, and integration cost.
Run (per interaction, at volume): cost-per-interaction x interactions-per-year. Show this at current volume and at the projected end-of-horizon volume, because this line grows with success. Pull the per-interaction figure from /ai-cost-model.
Maintenance: prompt and model upkeep, eval maintenance, monitoring, re-tuning as models change. [ASSUMPTION: maintenance at 15-20% of build cost per year] unless the user has a figure.
Change management: training, workflow redesign, the productivity dip during rollout. This is real and routinely omitted.

Step 5: Risk-adjust and compare to doing nothing

Risk-adjust the value: multiply the headline value by a confidence factor per driver (High 0.9, Medium 0.6, Low 0.3) so the ROI reflects what is likely, not the best case. Show the factor used.
Compute payback on risk-adjusted net benefit.
Build the do-nothing column: the cost of the baseline continuing unchanged over the horizon. The case only clears if the AI investment beats that, net of its own run and change cost.

Step 6: Output the business case

# AI ROI Business Case: (investment name)

**Approver:** (who) | **Horizon:** (months) | **Adoption assumed:** (X%)

## Executive summary
(3-4 sentences: what is proposed, what it costs to build and run, what it
returns risk-adjusted, and whether to proceed. The only section some will read.)

## Baseline (what happens today)
Task: (description) | Frequency: (N/year) | Done by: (role at $X/hr loaded)
Current annual cost: N x time x rate = $BASELINE/year
[ASSUMPTION / REVIEW markers as needed]

## Value drivers (measured against baseline)
| Driver | Mechanism | Gross value/yr | Confidence | Risk-adj value/yr |
|---|---|---|---|---|
| Time saved | (freed time -> revenue/headcount) | $X | High/Med/Low | $X x factor |
| Revenue lift | (measured conversion/retention) | $X | ... | $X |
| Quality | (errors avoided x cost/error) | $X | ... | $X |
| Risk reduction | (prob x incident cost, avoided) | $X | ... | $X |
| **Total** | | **$X** | | **$X** |
Show the math: e.g. time saved = (2.0 - 0.5 hr) x 4,000 tasks x $80 x 0.7 adoption = $336,000

## Cost model
| Cost | One-time | Year 1 | End-of-horizon |
|---|---|---|---|
| Build | $X | - | - |
| Run (per-interaction x volume) | - | cost x vol = $X | cost x grown-vol = $X |
| Maintenance | - | $X | $X |
| Change management | $X | $X | - |
| **Total** | **$X** | **$X** | **$X** |
Run-cost note: at (vol now) the run cost is $X; at (vol at horizon) it is $X.
This line grows with adoption and is the one to watch.

## Risk-adjusted ROI and payback
ROI = (risk-adj value - total cost) / total cost x 100 = X%
Show it: ($336,000 - $210,000) / $210,000 x 100 = 60%
Payback = total investment / monthly risk-adj net benefit = X months
Show it: $250,000 / $10,500 = ~24 months

## Do-nothing comparison
Cost of baseline continuing over horizon: $X
AI net position over horizon: $X
Verdict: (the AI case clears do-nothing by $X / does NOT clear it; here is the
threshold that would change the answer: ___)

## Recommendation
Decision: (Proceed / Proceed with conditions / Defer / Do not proceed)
Rationale: (2-3 sentences tying the risk-adjusted numbers to the decision)
Conditions / watch items: (run-cost ceiling, adoption milestone, kill criteria)

## Qualitative value (not in the numbers)
- (benefit, who values it, why it matters)

Step 7: Review