ROI from AI: How to measure what's hard to measure
"How much will we make on AI?" That question paralyzes decisions. Here's a framework that lets you answer it.
The CFO asks: "What's the ROI going to be?" The CTO answers: "It depends." And the AI project sits idle for another 6 months. This scenario repeats in most companies. According to McKinsey, only 39% of companies see any impact of AI on EBIT. The bottleneck is how you measure it.
Why traditional ROI doesn't work for AI
ROI assumes you know the costs and benefits up front. With AI you know one thing: that you don't know either. Costs escalate (infrastructure, data, experts). Benefits are scattered (time saved here, better quality there).
Trying to calculate ROI before deployment is like calculating the ROI of the internet in 1995. You can, but the numbers will be pulled out of thin air.
Framework: Time-to-Value instead of ROI
Instead of asking "how much will we make," ask "how fast will we see results." That changes the entire perspective.
TIME-TO-VALUE FRAMEWORK
- Week 1-2: Proof of Concept on real data
- Week 3-4: Pilot with 5-10 users
- Month 2: First business metric (time, cost, quality)
- Month 3: Decision: scale, pivot, kill
This is what I do hands-on — advising on AI strategy and building agents that survive the demo.
3 levels of value from AI
Value from AI shows up at three levels. Most companies only see the first.
Level 1: Time savings (easy to measure)
Process X used to take 4 hours, now it takes 30 minutes. Multiply by the hourly rate and you have your savings. Simple, but it's the tip of the iceberg.
Level 2: Quality and scale (moderately hard)
AI lets you do things you couldn't do before: analyze 100% of tickets instead of a 10% sample, respond to customers in 5 minutes instead of 24 hours. The value is real, but it's harder to count.
Level 3: New possibilities (the hardest)
AI opens the door to products and services that were impossible before. 1:1 personalization, demand prediction, autonomous systems. ROI is hard to calculate because you have no baseline.
Metrics that work
Instead of a single ROI number, track a portfolio of metrics:
AI METRICS PORTFOLIO
- Efficiency: Process time, cost per transaction
- Quality: % errors, model accuracy, CSAT
- Adoption: % of users, frequency of use
- Business: Revenue impact, cost avoidance
"You can't manage what you don't measure. But you also can't measure everything. Pick 3-4 metrics that truly show value."
How to talk to the CFO
The CFO doesn't need an exact ROI number. They need:
- • A clear investment cap (the maximum we'll spend on a test)
- • Success criteria (what has to happen for us to scale)
- • A decision point (when we evaluate and decide)
- • An exit strategy (how we back out if it doesn't work)
You're asking for a budget for one concrete experiment, with clear rules of the game. Not a blank check for "AI."
Start with one experiment
Pick one process, give it 90 days and a hard budget cap. Measure what you can and decide on data, not on a forecast in a spreadsheet. You'll only see the real ROI after deployment. It almost always lands different from what you assumed.
