How to Measure AI ROI: A Practical Guide for CFOs and Finance Teams
Every company seems to be saving thousands of hours with AI.
I would like to know where everyone is keeping them.
Because apparently AI wrote the email, summarized the meeting, analyzed the data, drafted the presentation and saved Brenda six hours on Tuesday.
Fantastic.
What happened Wednesday?
I ask because companies are getting very good at measuring time saved by AI.
I’m less convinced we’re measuring what happens to the saved time.
And if you’re a CFO trying to calculate the ROI of AI, that distinction is about to become fairly important.
AI ROI Is Becoming a Finance Problem
For the last few years, companies have been in the experimental stage with AI.
Try it.
Buy some licenses.
Run pilots.
See what happens.
That period isn’t completely over, but the conversation is changing.
Finance leaders are increasingly being asked to connect AI spending to actual business results. Recent research suggests that’s proving difficult: Protiviti reported in August 2026 that only 35% of finance organizations say they’re effective at measuring AI ROI, even as AI adoption continues expanding across forecasting and other finance work.
You know where this eventually ends.
Someone is going to ask:
What are we getting for this money?
And suddenly “everybody really likes it” becomes less compelling.
I’ve been in Finance long enough to know that enthusiasm has a remarkably short shelf life once somebody asks for an ROI calculation.
Let’s Say AI Saves Your Finance Team 1,000 Hours
Sounds great.
Let’s use it.
You automate reporting, variance commentary, reconciliations, data preparation and some forecast work.
Across the Finance organization, you’ve calculated that AI saves approximately 1,000 hours a year.
What’s the value?
Your first instinct might be:
1,000 hours × employee cost per hour.
There’s your savings.
Maybe.
But I want you to stop there for a second.
Did payroll actually decrease?
Did you eliminate a position?
Avoid a hire?
Increase output?
Shorten the close?
Improve forecast accuracy?
Find revenue?
Reduce working capital?
Make a decision faster?
Or did 1,000 hours simply become available?
Those aren’t the same economic outcome.
And this is where AI ROI gets slippery.
Time Saved Isn’t Automatically Money Saved
This isn’t an argument against productivity.
Productivity matters enormously.
But Finance should be careful about translating every hour saved directly into dollars.
Let’s say AI saves one of your FP&A analysts five hours every week.
Wonderful.
You’re probably still paying the analyst the same salary.
So the company hasn’t necessarily “saved” five hours of compensation.
It has created five hours of capacity.
Now I want to know what happened to it.
Maybe the analyst spends those hours improving the forecast.
Maybe she works with Operations.
Maybe she builds scenarios the team never had time to build.
Maybe she identifies $300,000 of unnecessary spending.
Maybe she leaves at 5:30 instead of 8:00.
I’m not dismissing that last one, by the way.
I enjoy seeing Finance people occasionally experience daylight.
But these are different benefits, and we should measure them differently.
This Is Where I Start Becoming Annoying
If someone tells me an AI tool saved 10,000 hours, I have follow-up questions.
I realize this makes me very popular.
How were the hours calculated?
Measured or estimated?
Compared with what baseline?
Are those recurring hours?
Did employees actually stop doing the work?
What did they do instead?
Did the business need more output?
Did we avoid hiring?
Did quality change?
Did errors decrease?
Did decisions get faster?
If nobody can answer those questions, I’m not saying the savings aren’t real.
I’m saying we don’t know what they’re worth yet.
There’s a difference.
Finance has spent decades explaining this distinction about everything else.
AI doesn’t get diplomatic immunity because everyone is excited about it.
I Would Separate AI ROI Into Four Buckets
If I were sitting with you trying to determine whether your AI investments are actually paying off, I wouldn’t start with one giant ROI number.
I’d separate the benefits.
1. Hard-Dollar Savings
This is the easiest category to defend.
Actual spending went away.
Maybe AI allowed you to:
avoid an additional hire,
reduce outside consulting,
retire software,
eliminate outsourced work,
reduce overtime,
or lower some other identifiable cost.
If you planned to hire three people and only needed two because automation absorbed the additional work, I can work with that.
There’s a before.
There’s an after.
There’s money.
Finance likes money.
It behaves better in spreadsheets than “strategic enablement.”
2. Capacity Created
This is where a lot of AI benefits will probably live.
The company still has the same people.
They’re just spending less time on certain work.
That’s valuable.
But don’t call all of it cost savings.
Call it what it is:
capacity.
Then track where the capacity went.
This is the part I think companies are going to struggle with.
Because if AI saves your Finance team 20% of its time and six months later everyone is still equally busy, I have questions.
Not accusations.
Questions.
For now.
Maybe the team is doing more valuable work.
Great.
Show me.
Maybe demand increased.
Fine.
Show me that too.
Or maybe the organization quietly filled the space with more work because organizations are nature’s most efficient producers of additional work.
Also possible.
3. Better Decisions
This is harder.
It may also be where some of the biggest value sits.
Suppose AI helps FP&A identify a demand slowdown three weeks earlier.
Management reduces inventory purchases.
Cash improves by $4 million.
What was the AI worth?
Now we’re having a more interesting conversation.
Or AI identifies unusual customer behavior.
Finance investigates.
The company changes pricing.
Margins improve.
Or scenario analysis that used to take three days now takes three hours, allowing management to evaluate an acquisition while the decision is still relevant.
Those aren’t labor savings.
They’re decision improvements.
And I’d rather have one meaningful decision improvement than a dashboard proudly informing me we saved 6,742 hours nobody can locate.
4. Risk Reduction
This one will be tempting to ignore because it’s difficult to put into an ROI calculation.
Don’t.
AI may help identify:
anomalies,
fraud,
control issues,
contract problems,
unusual transactions,
forecast risks,
or data problems earlier.
Sometimes the value is preventing something bad.
Finance understands this better than most departments.
Nobody celebrates the problem that didn’t happen.
There’s no company-wide email:
Great news. Nothing terrible happened today because Accounting caught something Tuesday.
Maybe there should be.
I’d read it.
Don’t Start With the AI Tool
Here’s where I think companies get themselves into trouble.
They buy something.
Then they go looking for the ROI.
I would reverse that.
Before approving an AI use case, ask:
What problem are we solving?
Then:
What does the process cost today?
How long does it take?
How many people touch it?
What’s the error rate?
What happens when it’s late?
What business outcome are we trying to improve?
What would we reasonably expect to change?
Now you have a baseline.
This sounds painfully obvious.
Which means companies will skip it constantly.
Six months later someone will ask whether the AI initiative worked and Finance will discover the only measurement anyone kept was the software invoice.
I’ve seen this movie with plenty of technologies that did not contain artificial intelligence.
The ending doesn’t improve because the software can now summarize it.
The Baseline Matters More Than the ROI Formula
Suppose you automate a monthly process.
Before AI:
120 hours.
After AI:
40 hours.
That’s useful.
But let’s keep going.
Did quality stay the same?
Did the cycle get faster?
Were errors reduced?
What happened to the other 80 hours?
If those employees used the capacity to take on work that previously required another hire, now you have a stronger financial argument.
If they used it to produce 11 additional reports nobody asked for, we have accidentally used artificial intelligence to increase reporting.
This would be very on-brand for Finance.
The technology is new.
Our ability to create unnecessary work remains undefeated.
Be Careful With Avoided Headcount
I like avoided-cost calculations.
I also don’t trust them automatically.
You’ll hear:
AI saved us three hires.
Okay.
Were we actually going to hire three people?
Was there an approved headcount plan?
Was workload growing?
Would those roles genuinely have been necessary without the automation?
Or did somebody multiply hours saved by 2,080 and discover three imaginary employees?
There are plenty of perfectly legitimate avoided hires.
Just document the logic.
If you’re going to put $360,000 of avoided compensation into an AI business case, I’d like to meet the three people we’re apparently not hiring.
Metaphorically.
Please don’t schedule that meeting.
Measure AI at the Use-Case Level First
I would resist trying to calculate one company-wide “AI ROI” number too early.
AI is too broad.
The economics of AI-assisted invoice processing are different from forecasting.
Forecasting is different from coding.
Coding is different from customer service.
Customer service is different from marketing content.
Instead, start with individual use cases.
For each one, track:
Cost
What does the technology actually cost?
Include licenses, implementation, integration, data work, training and ongoing support where relevant.
Baseline
What did the process look like before?
Change
What became faster, cheaper or better?
Outcome
What business result followed?
Confidence
How much of that result can you reasonably attribute to AI?
That last question matters.
If revenue increased 12%, please don’t give the chatbot all 12%.
Sales may object.
I Would Also Track Whether Anyone Uses It
This seems embarrassingly basic.
It isn’t.
Your company buys 500 AI licenses.
Three months later, 147 people use them regularly.
You do not have a 500-person AI deployment.
You have 147 users and 353 tiny monthly donations to a software company.
Before discussing sophisticated ROI models, I’d want:
active users,
frequency of use,
use cases,
adoption by function,
and whether the tool has actually replaced part of an existing workflow.
Buying technology and using technology remain separate corporate activities.
We’ve had decades to work on this.
Progress has been mixed.
What About AI That Doesn’t Produce Immediate ROI?
This is where Finance needs some judgment.
Not every worthwhile investment produces a measurable return next quarter.
Sometimes companies need to experiment.
Sometimes the capability itself matters.
Sometimes learning has value.
Sometimes waiting has a cost.
I’m fine with that.
Just call it what it is.
If you’re making a strategic investment because you believe the organization needs to develop AI capability, say so.
Don’t torture a spreadsheet until an uncertain experiment produces a 347% ROI.
Excel has been through enough.
You can make a strategic bet without pretending you know exactly what the return will be.
Finance’s job isn’t to eliminate uncertainty.
It’s to make the uncertainty visible enough that someone can make an intelligent decision anyway.
CFOs Are Going to Get Pulled Deeper Into This
This isn’t only an IT problem anymore.
Deloitte’s 2026 CFO Signals research found 19% of surveyed North American CFOs said they had the greatest responsibility for AI governance at their organizations, while 44% reported using AI in financial planning and budgeting. The same research found CFOs wrestling with cost transparency and the tension between deploying AI quickly and managing its risks.
That makes sense to me.
Eventually AI touches:
capital allocation,
headcount,
productivity,
technology spending,
risk,
planning,
and operating performance.
Finance was always going to end up in the conversation.
We usually arrive around the time somebody wants to know what something costs.
Here’s the AI ROI Scorecard I’d Actually Use
If you’re a CFO or FP&A leader looking at your AI investments, go pull your biggest five use cases.
Not 47.
Five.
For each one, answer:
What did this cost us?
What process did it change?
What was the baseline?
How much time did it actually save?
What happened to that capacity?
Did we avoid a real cost?
Did quality improve?
Did a decision happen faster?
Did revenue, margin, cash or risk change?
How confident are we that AI caused the improvement?
If you can’t answer all of them yet, that’s fine.
You’re learning.
But now you know what needs measuring.
That’s considerably more useful than announcing that AI saved 38,000 hours and hoping nobody from Finance asks what happened to them.
Unfortunately, I am from Finance.
The Question Isn’t Whether AI Saves Time
I’m fairly convinced it will.
In plenty of places, it already does.
The more interesting question is what companies do with the time afterward.
If your FP&A team spends less time preparing reports and more time helping management make decisions, that’s valuable.
If you can grow without adding as much overhead, that’s valuable.
If you catch problems earlier, that’s valuable.
If decisions improve, that’s valuable.
If employees spend fewer nights doing work a machine can handle perfectly well, I think that’s valuable too.
We just shouldn’t pretend all of those things are the same kind of return.
AI is going to make a lot of work faster.
Finance’s job is to figure out whether faster eventually became better.
Because somewhere right now, a company has apparently saved 100,000 hours with AI.
I’d still like to know where they put them.
