How to Measure the Return on AI Before You Launch
We have shipped systems where we can no longer prove the saving, because we did not time the work before it changed. The step that prevents this takes ten minutes and expires the day you launch.
· 5 min read
Patrick Laughlin, Founder, Chief AI Officer
Naval Academy, class of 2011. Decorated naval aviator, 10+ years. President, 2023 Inc. 5000 company.

Here is a failure of ours. It is the most useful thing I have to give on this subject.
A licensed apparel company. In a discovery interview, the accounts receivable specialist described collections: pull the aging report, work down it customer by customer, send the past-due notices. About four hours a week. She said it plainly, unprompted, and it went in my notes the same day.
That is a clean baseline, from the person who does the work.
Then we built the replacement. By the time it reached testing, three things had happened. The task had moved to a different person, who works it daily rather than in one weekly sweep. A new receivables hire had started and was being trained on the automated process. And we had never recorded a before for the person who now owns it.
So the honest position is this. There is a baseline. There will never be a clean comparison. Four hours a week for one person on a weekly sweep is not comparable to twenty minutes a day for a different person on a daily one. I can tell you exactly what we built. I cannot tell you what it saved.
Not because it saved nothing. Because the measurement had a window, and the window closed while we were shipping.
It has happened to us more than once
A grant consulting firm was producing its application packets by mail merge; its best week by hand was about a hundred and fifty applications. We built a packet generator that produces a complete packet in one click, and it has been in client testing since mid July. We have not yet timed a real run through it against the old way. Until we do, I will not publish a number, and there is a perfectly good number waiting to be collected.
At a national wealth management firm with several hundred advisors, the advisor pilot had a careful defect protocol, which produces bug counts. Bug counts are useful and they are not a business case, so together with the firm we added a time measurement and a baseline before anyone got access; the pilot post has the detail.
The ten-minute step
Before anything ships, for each workflow, write five lines.
The task, in the words of the person who does it, not in the words of the scope document.
Who does it, by name.
How often, and how long, timed once with a stopwatch rather than recalled.
What is actually being counted: minutes, errors, days, dollars. Pick one primary measure and let the rest be secondary.
The date you will measure again, and the name of the person who will read the number.
That is the whole thing. It takes ten minutes per workflow and it is the difference between a business case and an anecdote.
Why timed, not remembered
Recollections drift, and they drift in the direction of whatever the person is feeling that week.
Ask someone how long a monthly task takes, then ask again three weeks later, and two honest answers can differ by an hour. Neither is timed. If you build a business case on a recollection you have built it on a range that moves depending on the day, and the first person who checks will find that out.
One stopwatch reading is not statistics. It is still far better than a memory, and it takes one afternoon.
What the vendor calculator leaves out
Every AI vendor has one, and they all have the same three holes.
They assume everybody uses it. Assume sixty percent in the first year and you will still probably be generous.
They price only the license. Your people's hours in discovery, testing and training are real costs and they land on your payroll, not the invoice.
They count time saved as though the hour comes back in cash. It does not. An hour freed becomes an hour of something else, and unless you can name the something else, the saving never appears in a financial statement. The honest question is not how much time will we save. It is what will this person do with the hour instead, and does that show up in a figure the finance team already tracks.
Numbers that survive a reference call
A competing quote you did not have to pay. A vendor line item you cancelled. A hire you had budgeted and did not make. Days sales outstanding before and after, taken from the accounting system rather than from anyone's impression. Error rates. Cycle time from a timestamp, not from a feeling.
Counted things, not felt things. Every one of those can be produced by someone other than you, which is the test.
Then keep measuring
We measure adoption weekly after a launch and report it, including the weeks it goes down. A system with no usage number attached to it is not in production. It is installed.
If nobody is reading the number, it is not a number.


