AI ROI Audit Skill

Every org now spends real money on AI tools, and most justify it with adoption counts ("80% weekly active!") — which measure enthusiasm, not return. This skill audits what the spend returned, using methods that survive a sceptical CFO: baselines, counterfactuals, and quality deltas, with "we can't know yet" said out loud where it's true.

What This Skill Produces

• A per-tool verdict table: keep / consolidate / renegotiate / cut, each with its evidence
• The measurement behind each number — method, baseline, confidence — so the audit is checkable
• A hidden-cost ledger (the part vendor ROI decks omit)
• A baseline plan for every "unknown", so next year's audit has data

Required Inputs

Ask for (if not already provided):

• The AI tool inventory with costs: subscriptions, API spend, seats — and utilisation if known
• What each tool was bought to do (the promised outcome, from the original business case if it exists)
• Available evidence: usage data, before/after metrics, time studies, quality data, anecdotes (labelled as anecdotes)
• The decision at stake: renewal? consolidation? budget defence? (calibrates depth)

Audit Method

1. Reconstruct the promise. Per tool: what outcome justified the purchase — time saved, quality improved, headcount avoided, revenue created? A tool without a stated outcome gets audited against the best-fit guess, flagged as retrofitted.
2. Score with the strongest method the evidence allows, in descending order of credibility:
   • Natural experiment — teams/periods with vs without the tool, same work (best available in most orgs)
   • Before/after with baseline — the metric before adoption vs after, seasonality noted
   • Task-level time study — 10-20 real tasks timed with/without (cheap to run during the audit — do it rather than skip to tier 4)
   • Structured self-report — users estimating time saved, discounted (self-reported AI savings run ~2× actuals; say so) Never present a tier-4 number with tier-1 confidence. Every figure carries its method and a confidence label.
3. Count the hidden costs. Verification time (humans checking AI output), rework from AI errors that shipped, licence sprawl (seats bought > seats active), integration/prompt-maintenance time, and training time. These come off the gross benefit — an ROI audit that skips them is a vendor deck.
4. Convert honestly. Time saved → money only via a stated loaded rate and a stated assumption about what the time became (more output? earlier finishes? — different values). "Saved 400 hours" that nobody redeployed is capacity, not cash; label which one you're claiming.
5. Verdict per tool. Keep (positive with tier ≤2 evidence) · Consolidate (positive but duplicative — name the overlap) · Renegotiate (positive but mispriced vs utilisation) · Cut (negative or unmeasurable after a fair baseline attempt). Ties break toward the tool with a measurement plan.
6. Leave the audit better than you found it. Every "unknown" verdict gets a baseline plan: the metric, how it's instrumented, and the review date. The first audit is mostly this; that's a finding, not a failure.

Output Format

AI ROI Audit: [org/team] — [period]

Total AI spend: [sum] · Verdict summary: [n keep / n consolidate / n renegotiate / n cut / n unknown]

Hidden-cost ledger: [verification, rework, sprawl, maintenance — quantified where possible, listed where not]

The math shown: [for each material number: baseline, method, conversion assumptions]

Baseline plan for the unknowns: [tool → metric → instrumentation → review date]

One-paragraph CFO summary: [net position, the two decisions to make, and what will be measurable by next audit]

Quality Checks

• Every figure carries its measurement method and confidence — no naked numbers
• Self-reported savings are discounted and labelled as self-reported
• Hidden costs appear as line items, not a caveat sentence
• Time→money conversions state the loaded rate and the capacity-vs-cash claim
• Every "unknown" has a baseline plan with a date — the audit compounds

Anti-Patterns

• Do not use adoption or engagement as return — usage is a cost signal until an outcome moves
• Do not accept vendor ROI calculators as evidence — reconstruct from your own data or score it unknown
• Do not average across tools into one triumphant number — the verdict is per-tool or it decides nothing
• Do not claim headcount avoidance without the counterfactual hiring plan that was actually cancelled
• Do not punish honest "unknowns" by cutting them reflexively — cut requires a failed measurement attempt, not a missing one