How do you measure AI ROI?

Leadership buys outcomes, not AI

Leadership does not buy AI. They buy time back, money saved, faster service, and fewer mistakes. If you want to prove AI paid off, speak in those numbers and nothing else.

That means choosing one measure for each deployment, such as hours, cost, response time, or errors, and sticking to it from before to after.

The four numbers every AI deployment needs

Every deployment gets four numbers. Skip one and the result will not survive a second look.

• Before: how long, how much, or how often, measured before you change anything.
• After: the same measure, on a date you set in advance.
• Cost: tools, setup time, and training time.
• Who it affects: how many people, or how many times a month.

Measure the before number first

The before number has to be measured before anything changes. Memory always inflates it. If you ask a team after the fact how long something used to take, you will get a number that makes the new way look better than it is.

If you did not measure first, you cannot prove anything later. Measure on day one.

A worked example (illustrative)

Here is how the math works. These figures are an illustrative example of how to calculate, not a result from a specific client. Use your own measured numbers.

A team spends about 6 hours a week drafting a weekly family newsletter. After setting up a documented AI drafting workflow with human review, it takes about 2 hours. That is 4 hours a week, roughly 200 hours a year, for one workflow.

If the cost is a tool subscription and a day of setup, the readout writes itself.

Rules for honest AI numbers

Your credibility is the asset. One inflated number spends it. Follow these rules:

• Measure the before number before anything changes.
• Count review time, not just drafting time. If a person spends an hour checking the draft, that hour belongs in the after number.
• Count setup and training time in the cost, not just the subscription.
• Measure the after number on the date you set, not on a good day you picked later.
• Report what didn't work, alongside what did.

Vanity metrics to avoid

Some numbers feel like progress but do not tell leadership anything about the business. Skip these:

• Number of prompts sent.
• Number of tools adopted.
• Logins.
• “Satisfaction” without a business number attached.

Turn the four numbers into an annualized result

Leadership thinks in years and budgets. Once you have a before number, an after number, and how often the work happens, multiply the difference out to a year. In the illustrative example above, 4 hours a week becomes roughly 200 hours a year.

Then put the cost next to it. The question leadership is really asking is whether the result is clearly bigger than what it took to get there.

A simple scorecard for one win

Use one scorecard per deployment. Fill in every line:

• Workflow and owner.
• What we measure: hours, cost, response time, or errors.
• Before number and date measured.
• After number and date measured.
• Cost: tools, setup time, training time.
• People or times affected per month.
• Annualized result.
• What didn't work.
• Saved in the playbook? Yes or no, and where.

How to choose what to measure

Pick the measure before you pick the tool. Ask the people who do the work what actually costs them: hours spent, money spent, how long a customer or family waits for an answer, or how often something has to be redone. Choose the one leadership already cares about.

Then decide how you will capture it. A simple time log for two weeks is enough for most drafting work. A count of requests and response times works for anything that comes through an inbox. Write the method down, so someone else could repeat the measurement and get the same kind of number.

Finally, set the date for the after number now, and put it on the calendar. Deciding in advance what “working” means, and when you will check, is what separates proof from a story told after the fact.

Replace adjectives with facts

“Transformative” does not survive a board meeting. “4 hours a week back” does. When you report AI results, take out every adjective and put in the measured number. If you cannot fill in the number, the result is not ready to report.