How to prove AI ROI to a skeptical board
Boards aren't anti-AI. They're anti-unfalsifiable-number. To prove AI ROI, run a controlled before-and-after on one workflow and bring back a single defensible figure.
In short
To prove AI ROI, run a controlled before-and-after on one workflow, isolate the AI variable, and bring finance one baselined number.
- The gap most companies have isn't value, it's provable value: most see productivity gains but can't measure the return.
- A modest, defensible number beats a large, unfalsifiable one.
- Skepticism is a filter for rigor, not an objection to evidence.
The gap isn't value, it's proof
Most boards aren't anti-AI; they're anti-unfalsifiable-number. The problem leaders bring them usually isn't that AI created no value, it's that they can't prove the value they're sure they're getting. IBM's research captures it: roughly 79% of executives already see productivity gains from AI, but only about 29% can confidently measure the return. So when you walk in to prove AI ROI and the room is skeptical, the skepticism is rational, because most numbers presented to boards are program-level, un-baselined, and self-graded. An RGP survey of finance chiefs found only 14% had observed a clear, measurable impact from their AI investments, and that nearly half said the CFO owns making AI deliver measurable value. The job, then, isn't to argue harder; it's to bring a number that survives the doubt.
That's a method, not a louder slide. Here's the one that holds up.
Why the usual ROI deck fails the board
The standard AI ROI deck fails for predictable reasons. It claims a return for the whole program rather than a workflow, so there's nothing concrete to verify. It has no before-state, so the after-number floats free. And it's self-graded, with the team that ran the project also scoring it. On top of that, boards know the timing math: Deloitte found only about 6% of companies saw AI payback within a year, with typical payback running two to four years against the seven to twelve months they expect from ordinary tech, and McKinsey found only a small minority report any measurable EBIT impact at all. So a vague, near-term ROI claim collides with everything the board already knows. This piece is about designing the demonstration and presenting it; for which metrics to choose and how to value softer gains, our guide to measuring AI ROI is the companion.
A method to prove AI ROI on one workflow
Six steps from a single workflow to a number a skeptical board will accept. The rigor is in steps two and three.
- 1
Pick one workflow, not the program
Choose a single repeatable, high-volume workflow with a number the board already tracks: days to close, cost per invoice, tickets resolved per rep. One workflow is defensible; 'our AI initiative' is not.
- 2
Baseline the before-state for two to four weeks
Document the current number before anything changes. If you don't measure the before, you can never prove the after, and two to four weeks smooths out normal weekly variation.
- 3
Isolate the AI variable
Change only the AI step. Hold the team, volume, season, and tooling constant, and where you can, run a control: a matched group or period that keeps the old process, so the difference is attributable to the AI and not to a busy month.
- 4
Measure the same number, the same way
Re-measure the identical metric over a comparable window, usually 30 to 60 days, using the same definition and instrument. No moving the goalposts between before and after.
- 5
Convert to one defensible figure
Reduce it to a single number finance respects, like a payback period or a unit cost. Boards trust numbers that are baselined, unit-based, and traceable, measured against a documented before-state and evidenced by data, not asserted.
- 6
Present the before and after as one slide
Show where you started, where you are now, the delta, and why it's attributable. State the control and what you held constant, and name the source of every number. One workflow, one number, one mechanism of attribution.
A number survives a skeptical board exactly when it's baselined, unit-based, and traceable. The skepticism is a filter for rigor, not an objection to evidence.
"You can't isolate the AI variable"
The sharpest counterargument is that you can't really isolate the AI variable, so any ROI claim is attribution games, crediting AI for gains that would have happened anyway. The honest version of this is correct: without a counterfactual, self-reported attribution routinely overstates how much the AI actually caused. That's exactly why the method demands a control and a single changed step. You don't claim the whole transformation; you claim one isolated workflow with the comparison shown. The second objection, that payback runs years so any near-term number is noise, misreads what you're proving. The demonstration isn't the enterprise return; it's proof of mechanism on one workflow, the unit economics that justify funding the scale-up. And the third, that boards won't believe any number, has a clean answer: they won't believe a placement-only, un-baselined, self-graded number, and they're right not to. They will believe a baselined, unit-based, traceable one with the control stated up front. Proving it on one workflow is also the on-ramp to wider adoption, because the proof is what earns the budget. Pick the workflow and the owner by role, and run the whole thing as part of a real AI transformation.
Common questions
How do you prove AI ROI to a board?
Run a controlled before-and-after on one workflow: baseline the current number, change only the AI step while holding everything else constant, re-measure the same metric, and convert it to one defensible figure like payback period or unit cost. Present the before, after, delta, and the control on a single slide. A modest, attributable number beats a large, unfalsifiable one.
What evidence convinces a CFO that AI is paying off?
A number that's baselined, unit-based, and traceable: measured against a documented before-state, expressed per unit, and evidenced by data rather than asserted. CFOs distrust program-level, self-graded ROI claims for good reason, since only a small share of companies can confidently measure AI return. Show one workflow, one figure, and the attribution.
How long does it take to show AI ROI?
On a single workflow, usually 30 to 60 days after a two-to-four-week baseline, which is when time-savings and automation wins show up. That's proof of mechanism, not enterprise payback, which Deloitte found typically runs two to four years. The point of the short demonstration is to earn budget to scale, not to forecast the total return.
Can you really isolate the AI variable?
Imperfectly, but well enough if you change only the AI step and run a control: a matched group or period that keeps the old process. You don't claim the whole transformation, you claim one isolated workflow with the comparison shown. That's what makes a modest number defensible where a sweeping one isn't.
Prove it on one workflow, then scale
Candova helps your team put AI to work on a real workflow and measure the result, so you walk into the board with a number that holds up.
Power users save 10+ hours a week. Learn how.
The practical AI habits behind it, one a week.
Written by
Rich Hornstein
CFO & General Counsel of Candova
Rich is a CPA and an attorney with more than 25 years in finance and law at high-growth technology companies. He led Quotient Technology (formerly Coupons.com) through its roughly billion-dollar IPO as both CFO and General Counsel, and held finance and legal leadership roles at companies including McAfee and LogLogic before joining Study.com and Candova.