The opportunity is there. Now the question changes: does the math work? Without an objective answer, the project competes with everything else in the budget queue—and loses to whatever has a number attached. This module teaches you how to build that number honestly.
AI ROI equation — each component needs a real number, not an optimistic estimate.
🧮 ROI components
ROI isn’t just “we saved X hours.” It’s a complete account that includes what comes in (revenue) and what goes out (build cost, ongoing operating cost, and the risk factor that discounts the estimate based on the probability of success).
📐 The formula in practice
Year 1 ROI = (120 − 50 − 20) × 0,70 = R$35k—positive, but not exciting.
- • Reduction in operating costs
- • Revenue generated or protected
- • Errors avoided × cost of an error
- • Time freed up × hourly cost
- • Development and configuration
- • Cloud APIs and infrastructure
- • Maintenance and monitoring
- • Ongoing human supervision
💡 Practical tip
Always ask the client what the process you’re going to automate currently costs—in time and money. That’s the baseline for the gains. Without it, you’re estimating in the air.
💸 Total Cost of AI
The most costly mistake in AI projects is calculating only development costs and forgetting the costs recurring and growing that come later: per-token APIs, scaled infrastructure, maintenance when the model changes, and the human oversight that no AI system can eliminate completely.
Build cost (one-time)
Development, pipeline configuration, system integration, team training, pilot costs.
Monthly operating cost (recurring)
LLM APIs, vector database, storage, computing, drift monitoring, human review of outputs.
Maintenance cost (periodic)
Updating prompts when the model changes versions, recalibrating when data drifts, bug fixes, documentation.
⚠️ Warning
Projects that ignore the cost of human oversight often end up in the red by month 6. AI doesn’t run itself in production—someone always needs to review, correct, and maintain it.
📉 Baseline: Measure “Before”
The baseline is the snapshot of the process before any changes. Without it, any improvement is just a perception. With it, you have the comparator that turns “we think it improved” into “it improved X% compared with the documented baseline.”
✓ With a baseline
- ✓Before: 4h/day of manual triage work
- ✓Afterward: 40 min/day with human review
- ✓83% reduction—a defensible figure
- ✓Foundation for the next contract
✗ Without a baseline
- ✗"It got much faster"
- ✗"The team loved the solution"
- ✗Without a number for the CFO to approve expansion
- ✗Without evidence for the next pitch
💡 Practical tip
Treat baseline data collection as the project’s first deliverable—before any development. If the client doesn’t have the data, helping them measure the current process is a valuable service in itself.
⏳ Payback and time horizon
Payback is when the investment is recouped. The horizon is the limit of how long the calculation remains valid — AI changes quickly, and solutions have shorter useful life than traditional software. An AI project with a 36-month payback period probably won’t make it there intact.
📊 Calculation example
💡 Rule of thumb
For AI consulting projects, a payback period of over 18 months is a warning sign. The context changes, the model changes, the team changes. Prioritize cases that pay off within 12 months.
🎯 Business metrics vs. model metrics
The model has 94% accuracy. The decision-maker doesn’t know what that means for them — and shouldn’t need to. The consultant translates model metrics expressed in terms of business outcomes.
| Model metric | What it means | Business metric |
|---|---|---|
| 94% accuracy | 6 out of 100 predictions are wrong | Average cost of each error × 6 = R$X/month avoided |
| p95 latency < 2s | 95% of calls in <2s | Service SLA met; NPS protected |
| Recall 89% | 11% of relevant cases undetected | R$Y in revenue at risk from missed opportunities |
💡 Practical tip
Prepare two slide decks: one technical (for the team implementing it) and one business-focused (for those who approve and renew). Never use the technical deck in the executive meeting.
🧾 One-page business case
The one-page business case is the artifact that makes approval possible. It fits on one page because the decision-maker reads it in two minutes — and the a size constraint forces clarity: if it doesn’t fit, the thinking is still unclear.
📄 Structure of the one-page business case
What is costing money or holding back growth. Include the baseline number.
What will be done, in plain language. No technical jargon.
Expected benefit in R$, payback in months, and how it will be measured.
Total cost of build + 12 months of operation. Don’t underestimate it.
Top 2-3 risks and how they’ll be mitigated. Don’t sweep them under the rug.
⚠️ Attention
An optimistic business case that hides costs and risks wins the first contract — and destroys the relationship when the project meets reality. Be conservative about gains and honest about costs. Trust is the asset you’re building.
🎒 Module summary
Next module:
4.2 — Roadmap by waves: sequencing and prioritizing