How to Measure AI ROI: Pick One Number Before You Buy the Tool

You measure AI ROI by naming one number per workflow before the trial starts, then checking that number at day 30. Not tools adopted. Not hours “saved.” One number, chosen in advance, attached to a workflow you can point at. Everything else is a story you tell yourself after the invoice clears.

This is the Outcomes layer of the Ordered Clarity system, and it asks: what worked, and what compounds? Weak Outcomes create fake productivity.

Why most AI ROI math is fiction

The standard calculation goes: the tool saves me two hours a week, my time is worth $100/hour, the tool costs $20/month, therefore 40x return. Three things are wrong with it.

The baseline was never recorded. “Two hours” is a guess made after you already liked the tool. Nobody times the old workflow first.

Saved time isn’t recovered time. An hour freed at 2pm on a Tuesday becomes an hour of something else — often an hour of evaluating more AI tools. Time only converts to money when it gets reassigned to work that bills or ships.

The subscription isn’t the cost. The real cost is the subscription plus the setup, plus the learning curve, plus the review pass on the output, plus the switching cost when the tool dies. A $20/month tool with a 6-hour setup and a 20% rework rate is not a $20/month tool.

The four numbers

Pick the one that matches the workflow. One workflow, one number, written down before you start.

Workflow type	The number	Where it lives
Producing things (content, code, assets)	Units shipped per week, and rework rate	Your own tracker; PostHog if it’s in a product
Selling things	Revenue per channel, net of ad spend	Triple Whale for stores, your books otherwise
Being found	Citations and mentions in AI answers; rankings	Profound, Semrush
Serving people	Time to resolution, and the themes in the complaints	Dovetail

Rework rate is the one people skip and the one that decides most verdicts. A tool that doubles your drafts and triples your edits lost. Count what you shipped, then count what you had to fix.

The 30-day check

1. Write the number down first. Before the trial. A sentence: “Today we publish 3 posts a week and rewrite about 1 in 3.” If you can’t state the current number, you are not ready to buy the tool — you are ready to measure the workflow.
2. Change one thing. One new tool at a time. Two at once and neither verdict is real.
3. Log the real cost. Subscription, setup hours, review time. Actual hours, not remembered ones.
4. Check at 30 days. The number moved, or it didn’t.
5. Make the call. Keep, cut, or downgrade the plan. A tool that didn’t move its number is not “still bedding in.” It’s a subscription.

Start with the free instruments

Measurement shouldn’t cost more than the thing it measures. Three of the useful ones are free:

• Microsoft Clarity — $0, permanently. Heatmaps and session recordings. If AI is writing your landing pages, this shows you whether anyone reads past the fold.
• PostHog — generous free tier, then usage-based from $0.00005/event. Events, funnels, and session replay for anything product-shaped.
• Dovetail — free tier for clustering transcripts, tickets, and survey answers into themes with citations back to the actual customer sentence.

Add paid measurement only when a channel is big enough that being wrong about it costs more than the tool. Profound at $99/month makes sense when AI answer engines are a real acquisition channel, not before. Triple Whale starts at $219/month and earns it only if you’re buying traffic.

What not to measure

Tools adopted. A bigger stack is not progress. A clearer stack beats a bigger stack.

Prompts written. Activity, not outcome.

Vendor-reported time savings. The dashboard telling you the tool saved you 14 hours is marketing owned by the party being evaluated.

Everything at once. Four numbers exist in the table above. You need one per workflow, not all four.

The honest version of the math

Take one workflow. Write the current number. Add the tool. Log the real cost for 30 days. Compare.

Most of the time the answer is smaller than the pitch and still worth it. Sometimes the number doesn’t move and you cancel — which is also a return, because you stopped paying for it. The point isn’t a heroic multiple. The point is that you know, and you can say why.

Run this on your three most expensive AI subscriptions this month. Then audit the rest of the stack.

Stop collecting AI tools. Start building the machine.