3.2.3. Calculating AI ROI
💡 First Principle: Return on investment compares everything the initiative returns with everything it costs, over a defined period. AI business cases go wrong in two predictable ways: counting benefits that never become money or capacity (time "saved" that is not redeployed) and leaving out costs that sit outside the AI bill (data work, integration, change management, governance, ongoing operations). A comprehensive framework fixes both.
The basic formula:
ROI = (Total benefits − Total costs) ÷ Total costs
Benefit categories named in the exam guide:
- Time savings: hours saved × fully loaded labor rate, counted only when time is redeployed to valuable work or avoids hiring. Saved time that disappears into slack is not a financial return.
- Cost reduction: lower cost per transaction, reduced outsourcing, fewer errors and less rework, lower losses (such as fraud).
- Revenue growth: higher conversion, larger basket size, better retention, new products or services.
- Productivity gains: more output per person or per asset; capacity to grow without proportional headcount.
Cost categories that comprehensive frameworks include (total cost of ownership):
- Usage and infrastructure: tokens, instances, provisioned capacity, storage, and data transfer.
- Licenses and subscriptions: per-seat products, Marketplace subscriptions.
- Data work: acquisition, preparation, quality improvement, labeling, pipelines.
- Build and integration: internal team or partner time, connecting to existing systems.
- Change management and training: communication, user training, process redesign.
- Governance and risk: reviews, testing, monitoring, guardrails, audits, legal.
- Ongoing operations: monitoring, retraining, support, model version upgrades.
A worked example. An insurer deploys AI-assisted claims summarization for 200 adjusters.
| Item | Annual amount |
|---|---|
| Time saved: 45 minutes per adjuster per day × 220 days × 200 adjusters = 33,000 hours | |
| Value of time redeployed to handling more claims (avoids hiring 12 adjusters at $85,000 loaded cost) | $1,020,000 |
| Reduced claim leakage from more consistent reviews | $180,000 |
| Total annual benefits | $1,200,000 |
| Generative AI usage (tokens) and platform costs | $210,000 |
| Integration, data preparation, and build (amortized over 3 years) | $150,000 |
| Training, change management, and process redesign (amortized) | $60,000 |
| Monitoring, governance, evaluation, and support | $130,000 |
| Total annual costs | $550,000 |
| ROI = (1,200,000 − 550,000) ÷ 550,000 | ≈ 118% |
Notice that usage cost is less than 40% of total cost. A business case that counted only usage would have claimed an ROI above 470%, which a CFO would rightly distrust. Notice also that the time-savings benefit was converted into avoided hiring; if the insurer had no growth to absorb, the honest benefit would be smaller unless the freed time was put to other measurable use.
Two complementary measures are often used: payback period (how many months until cumulative benefits exceed cumulative costs) and sensitivity analysis (how ROI changes if adoption is 60% rather than 100%, or token prices change). AI benefits depend heavily on adoption, so presenting a range is more credible than a single number.
⚠️ Exam Trap: Options that calculate ROI using only model or API costs, or that count all theoretical time saved as cash savings, are distractors. The correct approach uses a comprehensive framework: benefits across time savings, cost reduction, revenue, and productivity; costs across usage, data, integration, change management, governance, and operations.
Reflection Question: A business case claims a 900% ROI for an AI email assistant, based on minutes saved per employee multiplied by average salary, with token costs as the only expense. Identify at least three weaknesses and how you would correct them.
