Measuring AI ROI

Most organizations measure AI ROI wrong in one of two directions: they're either wildly over-optimistic (counting theoretical maximum savings rather than actual ones) or they dismiss AI entirely because they can't assign a number to "better thinking."

The goal of this guide is a defensible, conservative number you can present to a CFO or board. That means: real inputs, conservative assumptions, and a clear methodology.

AI ROI at its simplest:

ROI = (Benefits - Costs) / Costs × 100%

The challenge is measuring benefits accurately. There are three types:

Type 1: Time savings. An AI does something that previously took human time. This is the easiest to quantify.

• Measure: hours saved per week × hourly cost of the person × 52
• Example: 3 hours/week × $75/hour × 52 = $11,700/year

Type 2: Quality improvement. AI output is better than what you had before, leading to downstream value.

• Harder to measure. Use a proxy: error rate reduction, customer satisfaction improvement, rework rate.
• Example: Reducing support ticket rework from 15% to 5% — quantify the cost per rework, multiply by volume.

Type 3: Revenue generation. AI enables something you couldn't do before — faster outreach, better personalization, more content.

• Hardest to attribute. Use controlled experiments if possible. Otherwise, use conservative estimates and label them as such.

Many ROI calculations undercount costs. Include all of these:

Direct costs:

• API usage fees (price per token × volume)
• Subscription fees for AI tools
• Infrastructure costs for any self-hosted models

Indirect costs:

• Engineering time to build and maintain the workflow
• Time spent writing and refining prompts
• Human review time for AI outputs that require checking
• Training and onboarding time for team members

Risk costs:

• Error rate × cost per error × volume
• Security/compliance review time

A complete cost picture makes your ROI calculation more credible, not less. CFOs trust numbers that don't look cherry-picked.

You can't manage what you don't measure. Set up logging from day one:

1. Track volume: How many tasks ran through your AI workflow per week/month?
2. Track errors: What percentage required human intervention or correction?
3. Track time: How long does the workflow take end-to-end, including any human touchpoints?
4. Track cost: Record API costs per run. Most providers expose this in their dashboard or API response.

With these four numbers, you can calculate weekly ROI on any workflow. Review monthly for the first quarter, then quarterly once the workflow is stable.

Build a simple spreadsheet: volume × time saved per task × hourly cost = gross benefit. API cost + maintenance time cost = total cost. Divide and multiply.

When presenting AI ROI to executives, three principles apply:

Be conservative. If you're uncertain whether a benefit is real, leave it out. Present the floor, not the ceiling. Conservative numbers that hold up in scrutiny are worth more than optimistic numbers that get challenged.

Show the methodology. Don't just present the final number. Show the formula, the inputs, and the assumptions. This builds credibility and allows leadership to adjust assumptions they disagree with rather than rejecting the whole analysis.

Separate proven from projected. Distinguish between ROI you've already achieved (measured) and ROI you expect to achieve (projected). Label projected ROI clearly and state what assumptions it depends on.

A good ROI presentation has four slides: the opportunity, the methodology, the current results (if any), and the projection with assumptions labeled.