How to measure whether AI is saving your company money

A four-step method any company under 300 people can use to find out if an AI tool is paying for itself: baseline, threshold, pilot, report.

November 13, 2026 · Ceiba Technology Services

Most companies cannot tell you whether their AI spending is working. They can tell you how many licenses they bought and that people seem to like it. Neither of those is a result.

Measuring AI is not difficult. It just has to be set up before the tool is switched on, because afterward nobody remembers how long things used to take.

1. Pick one process, not a tool

“We rolled out an AI assistant” cannot be measured. “Preparing the monthly owner report” can. Choose a process that is frequent, has a clear owner and produces something you can count.

Good first candidates are drafting routine correspondence, summarizing long documents, answering internal policy questions, preparing recurring reports and entering data from documents into a system.

2. Record the baseline

For two to four weeks, record how the process performs today. Use whichever of these apply:

Measure How to capture it
Hours per week Ask the people doing it to log time for the period
Turnaround time Timestamp when a request arrives and when it is finished
Error or rework rate Count items sent back or corrected
Volume Count items completed
Cost per item Hours multiplied by loaded hourly cost, divided by volume

This is the step most companies skip, and it is the only one that cannot be done later.

3. Agree the threshold before the pilot

Decide in advance what counts as success. For example: “Report preparation drops from ten hours a month to four, with no increase in corrections.” Write it down. A threshold set after the results are in is a justification, not a test.

4. Pilot with a small group and measure again

Give the tool to a few people, train them on this specific process and measure the same things for the same length of time. Then do the arithmetic: hours saved, multiplied by cost, minus the full cost of the tool including setup, training and licenses.

If the number clears the threshold, expand. If it does not, change the approach or stop. Stopping a pilot that did not work is a good outcome. It is far cheaper than a company-wide license nobody can justify.

What this looks like in practice

Every AI engagement Ceiba runs follows this sequence. We also do one thing before any of it: review who has access to which files. An AI assistant can read whatever the person using it can read, so permissions get fixed first.

If you would like help choosing the first process and capturing the baseline, talk to us.

Quick answers

How do you calculate the ROI of an AI tool?

Measure one process before the tool is introduced: hours per week, turnaround time, error rate or cost per transaction. Run a pilot, measure the same things again, convert the difference to dollars and compare it with the full cost of the tool, including licenses, setup and training.

How long should an AI pilot run?

Long enough to cover the normal cycle of the process and to get past the novelty period. For a weekly process that usually means several weeks. Set the end date and the success threshold before you start.

Find out what your technology should be doing for you

Start with a technology and automation assessment. We baseline where time and money are going today, and show you what we would fix first.