If you read a LinkedIn post that says "we implemented AI and productivity went up 73%", be suspicious. The 73% usually comes from extrapolating a 4-week pilot with 6 users. That isn't ROI. That's marketing.
The 5 typical lies in presented ROI
- "We increased productivity by X%" — without defining which productivity or how it was measured
- "Saved Y hours per week" — without subtracting the team time spent maintaining the system
- "5× ROI" — without counting the opportunity cost of what didn't get done
- "Paid back the investment in 3 months" — without amortizing the recurring cost you'll have forever
- "Improved CSAT by 30 points" — comparing against a bad week, not the 6-month average
How real AI ROI is measured
Honest baseline
Measure the prior metric for at least 8 weeks (not 2). Adjust for seasonality if your sector has peaks. Document the current cost of the process including the hidden cost (management, errors, rework).
Total cost of ownership (TCO) over 24 months
- Setup: implementation + team onboarding + integrations
- Recurring: model licenses, infra, monitoring
- Operation: time from your team maintaining the system (it's not zero)
- Continuous improvement: 10-20% of annual cost so it doesn't go stale
Honest benefit
- Time recovered × real fully-loaded hourly cost (not salary, company cost)
- Errors avoided × average cost of the error
- Additional volume processed without scaling headcount
- Attributable revenue (with tracking, not estimation)
- Subtract the time your team spends maintaining the system
What does NOT enter the ROI (but matters)
There's value that doesn't fit in a spreadsheet but is real: your team no longer burning out on repetitive work, being able to sell a service that used to be impossible, your competition taking longer to match your service level. That doesn't go in the board deck — you feel it in the mood, in employee churn, in how fast you close new clients.
That's why for the past few months we've kept a public calculator: /roi. Three inputs, an honest range (low-high) and the assumptions visible so the math can be audited. If the numbers work for you, perfect. If they don't, also perfect — better to know before you sign.