AI ROI Worksheet: Calculate Real Payback Before You Sign a Contract
Every AI vendor will pitch you a return on investment. Almost none of them will share the assumptions, the timeline, or the costs they leave out. This worksheet does the opposite. It walks through six categories of value an AI implementation can deliver, the five cost categories most teams underestimate, and a simple payback-period formula that survives a CFO's first pass. Use it before signing any AI contract — or, if you've already signed, use it to baseline where you are and what you should expect.
How to Use This Worksheet
Work through Sections 2–7 in order. For each value or cost line, write down (a) the dollar figure and (b) the confidence level (High/Medium/Low). At the end, you'll have a year-1 payback number AND an honesty score — the percentage of your inputs you marked High confidence. Worksheets that come out to 'positive ROI' with 80% Low confidence inputs are how AI projects fail. Aim for 60%+ High confidence before signing a contract.
Section 2: Value Category 1 — Labor Cost Displaced
**Formula:** (Hours per week saved × number of people × loaded hourly cost) × 52 = annual savings
**Inputs to fill in:**
• Hours saved per person per week: ______ (confidence: H/M/L)
• Number of people: ______
• Loaded hourly cost (salary + benefits + overhead, usually 1.3–1.4× wages): $______
• Annual savings: $______
**Worked example:** AI customer-service triage saves 8 hours/week for each of 5 reps. Loaded cost is $35/hour. Annual savings = 8 × 5 × 35 × 52 = $72,800.
**Honesty check:** If your 'hours saved' number came from the vendor demo and not from a 30-day shadow trial, mark this Low confidence. Vendors consistently overstate hours saved by 2–3x. If those hours never actually get freed up (people fill them with other work), it's not a real cash saving — just a productivity gain.
Section 3: Value Category 2 — Error Reduction
**Formula:** (Current error rate − projected error rate) × volume × cost per error = annual savings
**Inputs to fill in:**
• Current error rate (%): ______ (confidence: H/M/L)
• Projected error rate with AI (%): ______ (confidence: H/M/L)
• Annual transaction volume: ______
• Cost per error (rework time + refunds + customer churn impact): $______
• Annual savings: $______
**Worked example:** Invoice processing — current error rate 3%, projected 0.5%, annual volume 12,000 invoices, average cost per error (re-processing + customer escalation) = $45. Annual savings = (0.03 − 0.005) × 12,000 × 45 = $13,500.
**Honesty check:** Vendors quote error rates from their best customers. Use your CURRENT measured error rate — not their case study. If you don't measure your error rate today, mark Low confidence and budget the first 60 days of the pilot for measurement before claiming savings.
Section 4: Value Category 3 — Revenue Lift
**Formula:** Multiple sub-formulas depending on lift type:
• Conversion lift: incremental conversion rate × visitors × average order value
• Upsell/cross-sell lift: incremental % × customer base × ACV
• Speed-to-quote: faster response × close rate uplift × pipeline value
**Inputs to fill in:**
• Lift category: ______
• Baseline metric: ______
• Projected uplift: ______ (confidence: H/M/L)
• Annual revenue impact: $______
**Worked example:** AI-powered product recommendations on an ecommerce site lift average order value from $84 to $96. Annual orders: 18,500. Annual revenue lift = ($96 − $84) × 18,500 = $222,000. Apply your gross margin (e.g., 35%) for true contribution: $77,700.
**Honesty check:** Revenue lift claims are the most inflated category in AI ROI calculations. Always multiply by your gross margin — not gross revenue. Always derate by 30–50% for the first 12 months while the system tunes. Always require a measured A/B test before counting lift as 'real.'
Section 5: Value Category 4 — Risk and Compliance
**Formula:** (Probability of incident × cost of incident) before AI − (probability × cost) after AI = annual risk-adjusted value
**Inputs to fill in:**
• Type of risk (fraud, fines, churn, security): ______
• Current annual cost of incidents in this category: $______ (confidence: H/M/L)
• Projected reduction with AI: ______% (confidence: H/M/L)
• Annual risk-adjusted value: $______
**Worked example:** Fraud detection AI projected to cut card-not-present fraud losses from $48,000/year to $14,000/year. Annual risk-adjusted value = $34,000.
**Honesty check:** This is the right place for 'soft' benefits, but quantify them. 'Better compliance posture' is not a number. 'Reduced expected regulatory fine from $X to $Y based on [specific framework]' is.
Section 6: Value Category 5 — Customer Experience
**Formula:** Two-step — (a) baseline retention/satisfaction impact, (b) translate to dollars.
**Inputs to fill in:**
• Current churn rate (annualized): ______%
• Projected churn reduction with AI: ______ percentage points (confidence: H/M/L)
• Average customer lifetime value: $______
• Number of customers: ______
• Annual retention value: $______
**Worked example:** Faster customer-service AI projected to reduce annual churn from 14% to 11% (3 percentage points). 4,200 customers × $1,800 LTV × 0.03 = $226,800 in retention value over the typical customer lifetime, recognized annually as ~$45,000 (assuming 5-year LTV horizon).
**Honesty check:** Almost every AI vendor will promise CX gains. Don't claim retention value unless (a) you measure churn today, (b) you can isolate AI's contribution from other variables, (c) you have at least 90 days of post-launch data, or you've allowed for a 50%+ derating.
Section 7: Value Category 6 — Capacity Expansion
**Formula:** (New work product enabled × revenue per unit) − (additional capacity cost) = annual capacity value
**Use this when** AI doesn't displace existing labor but instead lets your team do work you couldn't do before — handling more leads, supporting more customers, producing more content, opening more markets.
**Inputs to fill in:**
• Additional output enabled per year: ______
• Revenue per unit of new output: $______
• Annual capacity value: $______
**Honesty check:** This category is real but easy to fake. The output has to be (a) something customers actually pay for, (b) something your team couldn't have done with hiring instead, and (c) something your sales/marketing pipeline can actually deliver to. If you can't fill the new capacity, it's not value.
Section 8: Cost Side — The Five Categories Most Teams Underestimate
Total Cost of Ownership for Year 1 includes:
• **License/subscription:** $______ (vendor will quote this readily)
• **Implementation services:** $______ (often 30–100% of license cost for non-trivial deployments)
• **API/usage costs:** $______ (variable; can dwarf license costs at scale — get a usage forecast and a cap)
• **Internal team time:** $______ (the project lead's 25%+ time, your IT lead's 10–15%, training time for end users — count this in dollars)
• **Integration costs:** $______ (custom connectors, data pipeline work, identity integration — almost always more than the vendor estimates)
• **Change management:** $______ (training, documentation, the productivity dip during rollout)
• **Tail-year costs:** $______ (year 2+ tuning, retraining, vendor renewal increases — at least 60% of year 1 typical)
**Annual TCO (Year 1):** $______
**Annual TCO (Years 2–3 average):** $______
**Honesty check:** If your TCO is just 'license + implementation,' you're missing 40–60% of the real cost. Most failed AI ROI calculations are not failed math — they're missing cost lines.
Section 9: Putting It Together — Payback and Sensitivity
**Total annual value (sum of Sections 2–7):** $______
**Total annual cost (Section 8):** $______
**Net annual value:** $______
**Payback period (months):** Year 1 total cost ÷ (annual value ÷ 12) = ______
**Sensitivity check — run these three scenarios:**
• **Base case** (your numbers as written): payback in ______ months
• **Conservative case** (cut every Low confidence value input by 50%, increase cost by 25%): payback in ______ months
• **Pessimistic case** (cut all value by 50%, increase cost by 50%): payback in ______ months
**Decision framework:**
• Payback < 12 months in base case AND < 24 months in conservative case → go.
• Payback 12–24 months in base case, ≥36 months in conservative case → pilot only, with tight measurement before scaling.
• Payback > 24 months in base case → reconsider the project, the vendor, or the scope.
**Honesty score:** % of inputs you marked High confidence: ______. If under 60%, do another 30 days of measurement before signing.
Frequently Asked Questions
Frequently Asked Questions
Counting hours saved as cash savings when those hours don't actually reduce headcount or unlock measurable new output. If your AI saves 200 hours a month across 10 people but you don't reduce headcount and you don't measure their new output, that's a productivity claim, not a P&L impact. Be honest with yourself — your CFO will be.
About 20 minutes if you already know your baseline metrics. If you don't know your current error rate, churn rate, or volume — that's the real result. You can't calculate ROI on a project you haven't baselined.
No. Use it internally to pressure-test their claims. If the vendor's numbers don't survive your sensitivity check (conservative case payback > 24 months), don't sign. You're the one paying the bill — you do the math.
Trust the nervousness — it's usually pointing at a Low confidence input you flagged but didn't fully internalize. Re-do the conservative case with that specific input cut in half, and see if the project still pencils out. If it does, ship. If it doesn't, the worksheet is helping you make the right call.