How to measure the ROI of AI

1. Read the baseline. Don’t ask for it.

To know whether AI made something faster or cheaper, you need to know how fast or cheap it was before. Chances are nobody wrote that down.

You don’t need them to. The systems your teams use kept the history for you. Email threads have timestamps. So do support tickets, sales deals and pull requests, going back years. Pick a measure, find when AI arrived for that team, and compare the months before with the months after.

A few measures that work well:

• email: time to first reply;
• support: hours to resolve a ticket, for AI-assisted tickets against the rest;
• sales: win rate and deal size, for reps who use AI heavily against those who don’t;
• engineering: time from a pull request opening to it merging.

Compare like with like. The same team, the same kind of work, over a long enough window that one odd month doesn’t decide it.

2. Count the whole cost

Licences are the number everyone knows, and usually the smallest part. Put in the set-up as well: implementation, any consultants, and the hours spent training people. Then the running cost every month, which covers licences, infrastructure and the time someone spends looking after it.

Count every seat you pay for, not just the ones in use. An idle seat is still a cost, and it is often the easiest one to cut. Our guide to finding unused AI licences covers that.

3. Turn the change into money

Time saved becomes money at the loaded cost of the people whose time it is: salary plus employer costs, per hour. Faster resolution on a support desk might mean fewer contractors. A higher win rate is revenue, if you can tie it to the reps using AI.

Be conservative and write your assumptions down next to the figure. A finance team will trust a modest number they can check far more than a big one they can’t.

4. Do the sum

Net ROI = (total return − total cost) ÷ total cost, over a stated period. Always give the period. A figure without one can mean anything.

Here is the example organisation used on our home page, over 36 months:

Plot the running total month by month as well. It starts below zero, because the set-up is paid before anything comes back. The month it crosses zero is your payback point. In the example it gets there after about seven months.

5. Report it per product

One ROI figure for all your AI hides the interesting part. In the demo data behind our Ask Lumela examples, Amazon Bedrock returns 214% while Microsoft 365 Copilot costs more than it brings back. Averaged together, they tell you nothing. Split out, you know what to expand and what to switch off.

Mistakes worth avoiding

• Using a vendor’s case-study figures instead of your own history.
• Counting the seats people use and forgetting the ones they don’t.
• Measuring too early. Adoption ramps over the first few months, so a snapshot in month two undersells it.
• Quoting a percentage with no time period, or no note of what went into it.

How Lumela does this

Lumela connects to the systems that hold the history, reads the before and after for you, and reports cost and return per product. See what it connects to today.