The budget review reaches the line for AI. A director asks what this year's spend returned. The answer on the slide comes from the vendor's dashboard: a large number of hours saved. Then the CEO asks the obvious follow-up. Headcount is unchanged, the close still takes the same number of days, and DSO has not moved. Where did the hours go? That exchange is why AI ROI in finance needs measuring the way any other investment is measured: against a baseline, in units the board recognizes, with the full cost on the other side.
Why AI ROI in finance is hard to show
Three things usually get in the way. There was no baseline, so the improvement is an estimate. The benefit was counted in hours, and hours that are not redeployed disappear into the week. And the cost side left out the time people spend reviewing what the AI produced.
The survey picture is mixed, and worth reading carefully. Gartner's June 2026 release, based on a survey of 183 CFOs and senior finance leaders run in May and June 2025, reported that while 84% of finance organizations had implemented AI or planned to, only 7% reported a high or very high impact. KPMG's survey of 1,013 senior finance leaders at large companies, run in March 2026 and published in May, found 71% saying AI is "meeting or exceeding ROI expectations". These figures do not cancel each other. They come from different samples, surveyed about nine months apart, and ask different questions: one about the scale of impact, the other about performance against expectations, which may have been modest. Across all business functions, McKinsey's August 2026 State of AI survey found 37% of respondents attributing at least some EBIT impact to AI use, with about 6% qualifying as high performers that attribute 5% or more of EBIT to it.
None of these tells you your own return. They do suggest that broad claims of impact deserve a measured answer.
Four measures: hours, errors, cycle time and cash
| Measure | How to measure it | Outside reference | How it becomes money |
|---|---|---|---|
| Hours per unit | Logged hours by process ÷ system volume, before and after | None that travels well; use your own baseline | Only if hours are redeployed to named work or replace planned hiring |
| Error and rework rate | Share of items reworked, reversed or corrected | APQC median 95.0% of disbursements first-time error free | Rework hours saved, plus direct losses avoided, such as duplicate payments stopped |
| Cycle time | Invoice receipt to payment; days to close | APQC medians: 15.0 days receipt to payment; 6 days to consolidated monthly statements | Early-payment discounts taken, late fees avoided, earlier management information |
| Cash | DSO, unapplied cash, payments stopped before release | APQC median DSO 38.0 days | Working capital released, valued at your cost of capital |
Hours. Measure hours per unit rather than total hours, so a busy month does not look like a failure. Then say where the released hours went. If the answer is "absorbed", report them as capacity, not return.
Errors and rework. These are often the most credible measures, because they come from the system rather than from time logs. A fall in reversed journals, blocked invoices or misapplied receipts can be counted.
Cycle time. A shorter invoice-to-payment cycle can capture discounts. A shorter close puts information in front of management sooner. The second is real and harder to value, so report it as days rather than forcing a dollar figure onto it.
Cash. This is where AI in collections and cash application shows up, if it shows up anywhere. The arithmetic is simple. Take a company with $120m of annual revenue: one day of DSO is roughly $329,000 of receivables. A sustained reduction of a few days releases cash a board will recognize.
Count the cost side in full
The cost of AI in finance is more than the licence. Include build and integration, data preparation, the controls work needed for audit, ongoing monitoring of accuracy, and changes when the vendor updates its model. Above all, include review time. If people spend hours each week checking AI output, those hours belong in the cost line, not hidden inside the benefit.
Separate AI from everything else
Most AI projects in finance come with process fixes: a cleaned supplier master, a tolerance rule finally written down, a receipt step enforced. Some of the gain belongs to those fixes. Count it, but label it, because it would have been available without the tool.
Where you can, compare like with like. Use one entity or business unit that changed against one that did not, over the same months. Where you cannot, compare the same months year on year and adjust for volume. Report both the result and how confident you are in it.
Reporting it to the board
One page per process is enough. Show the baseline, the current figure, the change, the money effect, the full cost, and whether each number was measured or estimated. Keep three kinds of benefit apart: cash that has been realized, capacity that is available and where it has gone, and risk reduced. Mixing them is how hours-saved figures end up presented as savings.
What an ROI calculation will not tell you
It will undervalue risk reduction. A duplicate payment that was stopped or a fraudulent bank change that was caught leaves no trace in the numbers unless someone records it at the time. Keep a log of exceptions caught, with amounts, from the first day.
It will also say little about the quality of decisions. A faster close and better variance commentary should improve how the business is run, and that rarely converts into a figure anyone trusts. Report it in words, and keep it separate from the money.
Frequently asked questions
How do you measure AI ROI in finance?
Measure hours per unit, error and rework rates, cycle time and cash effects for each process against a baseline taken before the change, then set the money value of those changes against the full cost, including the time people spend reviewing AI output.
What share of finance teams see a return from AI?
Survey answers vary with the question asked. In Gartner's June 2026 release, only 7% of finance organizations reported a high or very high impact, while KPMG's May 2026 survey found 71% of senior finance leaders saying AI was meeting or exceeding ROI expectations.
Do hours saved by AI count as return?
Only when they are redeployed to named work or avoid hiring that would otherwise have happened. Hours that are absorbed into the existing week are capacity, and should be reported as such.
Which costs belong in a finance AI business case?
Licences, build and integration, data preparation, controls and audit work, ongoing accuracy monitoring, model changes, and the time people spend checking AI output.
How do you separate AI's impact from process improvements?
Label process fixes made during the project separately, and where possible compare an entity or unit that adopted the change with one that did not over the same period.
Where CFOLogic fits
When we build an AI-assisted tool or proof of concept with a CFO under CFO Success Partners, the baseline and the measures are agreed before the build starts, so the return can be shown afterwards. The tools sit on top of systems such as SAP, Microsoft Dynamics or QuickBooks Online, read data without changing it, and leave approval with a person.