Perspective · ROI
Measuring AI ROI in the first 90 days.
"What's the ROI?" is the right question to ask about AI — and the one most pilots can't answer, because nobody set it up to be measured. After watching plenty of projects, I've come to believe the ROI conversation isn't hard. It's just usually had too late. Here's how to measure it honestly in 90 days, before you've spent real money.
Decide the number before you build
The most common ROI mistake isn't bad math — it's measuring after the fact. If you don't write down what success looks like before you start, you'll end up arguing about vibes: "it feels faster." Pick one or two concrete metrics up front, and capture today's value as your baseline. Without a baseline, you can't prove anything moved, and a pilot you can't prove is a pilot you can't scale.
The metrics that actually matter
You don't need a finance degree. For most operational pilots, ROI comes down to some mix of:
- Time saved — hours per week a team gets back, valued at a loaded labour rate.
- Cycle time — how fast a request, exception, or document moves from start to done.
- Accuracy / rework — error rates and the cost of fixing mistakes downstream.
- Service level — response time, on-time rate, or customer satisfaction.
- Cost avoided — overtime, temp staff, or penalties you no longer pay.
Pick the one or two that map to the workflow you're piloting. More than that and you'll drown in measurement instead of doing the work.
The baseline trap
Here's where honesty matters. It's tempting to measure the baseline loosely and the pilot generously — that's how you get ROI numbers nobody believes. Measure both the same way. If "time spent" before the pilot was a guess, make the "after" a guess too, or measure both properly. And count the full cost of the pilot: the tool, the build time, and the human review it still requires. Honest ROI that's modest beats impressive ROI that falls apart under a CFO's questions.
Don't forget the costs that hide
A real ROI picture includes the unglamorous line items: the time your team spends reviewing AI output, the cost of the tool at the volume you'll actually run, and the ongoing maintenance. AI that "saves" ten hours but creates eight hours of checking isn't a win. The good news: if you're measuring properly, you'll catch that in 90 days instead of a year.
The honest go/no-go
At 90 days you should be able to say, plainly, whether the metric moved enough to justify scaling. Three outcomes are all valid: scale it (the number moved, expand), fix it (promising but not there, iterate once more), or stop it (it didn't pay back — and learning that for the cost of a pilot is a win, not a failure). The companies that win with AI aren't the ones that never kill a project; they're the ones that kill the wrong ones cheaply and double down on the right ones.
Why 90 days
Ninety days is long enough to get real signal and short enough that you haven't bet the year. It forces a narrow scope, which is exactly what makes a pilot measurable in the first place. If a project can't show movement on its metric in a quarter, that's usually a sign it was scoped too big — a problem I see constantly, and one I wrote about in why most AI pilots never reach production.
Want an ROI you can defend?
An AI Opportunity Assessment scores your workflows, sets the baseline and the target metric up front, and hands you a 30/60/90-day plan built to prove (or disprove) ROI — in about two weeks.