Time saved is not EBITDA

Time saved is the most common way to describe an AI result, and it is often a real operational change. It is not, by itself, a line in the P&L. Hours freed are capacity. Earnings move when something in the accounts moves: overtime or contractor bills go down, an approved hire is not made, volume rises with the same staff, or a cost that would otherwise have been spent is actually avoided. Until finance can show which of those happened, “time saved” should stay in the operations column.

This note is about that translation: how to measure the financial impact of AI when the first result you can see is hours, and how to keep a capacity figure from being reported as EBITDA. It is a finance question, not a marketing one, and it is not a method for crediting campaigns or channels.

A saved hour is not yet a cash saving

Managerial accounting already makes the distinction. A favorable labor-efficiency variance means fewer hours were used than the standard allowed for the output produced. It does not automatically mean the company paid less cash. The same distinction applies when a team reports savings from AI.

A 2026 guest post on CFO.com, written by an accounting faculty member, walks through the point with a simple accounts-payable illustration and then separates four kinds of claim. The argument does not depend on any vendor survey. If average handling time on 1,000 invoices falls, hours are released. What happens next determines the financial result. If salaries, overtime, and contractor bills are unchanged, the immediate benefit is time available for other work, not a reduction in spend.

A benefits schedule should keep four claims apart:

A realized spending reduction, shown in an overtime payment, a contractor invoice, or another cash outflow that actually declined, after checking that the work did not simply move to another team.
Avoided future spending, such as a deferred hire, which needs an approved staffing plan, a credible workload forecast, and a date when the cost would otherwise have arisen. Hours spread across several people do not automatically eliminate a role.
Released capacity, stated as hours made available and the work assigned to them. If demand has not grown and no useful work has been reassigned, the capacity is available, but it is not yet a delivered financial outcome.
Quality, standing on its own: errors, rework, or unresolved exceptions, measured directly, not buried inside a salary-based estimate.

The same hour should not be counted twice. If released time leads to lower overtime, do not also count the full salary value of those hours as a second saving. If a software bill rose in the same period, deduct the incremental running cost before reporting a net cash benefit. Capacity, quality, and a future avoided hire can all be real. They are not the same as EBITDA already earned.

What a productivity study can and cannot show

A published study can document a productivity change without documenting a P&L change. That is not a weakness in the research. It is a reminder of what the metric is.

The Quarterly Journal of Economics published “Generative AI at Work” in 2025. The authors studied the staggered introduction of a generative AI conversational assistant using data from 5,172 customer-support agents at one firm. Access to the assistant increased worker productivity, measured as issues resolved per hour, by 15% on average, with substantial differences across workers. The abstract describes a productivity result, not a measured reduction in payroll, contractor spend, or earnings.

That is exactly the translation problem a portfolio company faces with copilots and drafting tools. Issues resolved per hour, invoices coded per hour, or hours no longer spent on a first draft can be observed, and sometimes measured carefully. They do not, on their own, tell you whether the wage bill changed, whether hiring plans changed, or whether the freed time was used on work that shows up in revenue or margin. The study is also one firm and one workflow: evidence that a productivity gain can be real and still leave the earnings question open, not a benchmark to copy into a board pack.

The reporting gap is about outcomes, not activity

Finance teams know that seeing activity is not the same as seeing a return.

A 2026 CloudZero survey of 260 senior finance leaders, 135 of them CFOs, reports that only 22% can tie AI spend to business outcomes, and that 60% admit they are already spending more on AI than they can justify. Eighty-seven percent say they need to close the outcome gap within the year. When the survey asked which unit-cost metric mattered most, cost per dollar of revenue led at 54%, and cost per employee hour was close behind at 49%. Respondents could pick several metrics. Wanting a cost per hour is not the same as having converted hours into earnings. CloudZero is a software vendor, and the figures are from its own survey. The report states that the percentages describe associations in a single point-in-time survey, not proven cause and effect.

A 2026 Coupa survey of 600 global finance leaders reports that 76% say difficulty quantifying AI ROI is hindering further investment. Coupa is also a software vendor, and the public page does not state fieldwork dates. Neither survey says how often PE-backed companies record time savings versus margin impact. They do show that hours are easier to count than earnings, which is why they show up first and get over-read.

Keep the two claims from being mixed

Confidence labels exist so that a capacity result and a financial result are not written as if they were the same number.

If you observed a change in handling time or output per hour against a stated baseline, you can often label that operational result measured. The financial translation is a separate line. It is measured only when the accounts, or an equivalent finance record, show the cost out, the avoided spend, or the volume-driven contribution, and when you can explain other changes in the business. If you can only model the translation — for example, by multiplying hours by a loaded wage — label it estimated and say what is assumed. If the translation is a business-case figure that has not been tested, it is projected. If you cannot show where the time went, the earnings claim is not yet attributable, even when the hours themselves are real.

Do not move a figure up a label to make the pack easier to present. A measured reduction in overtime of a modest size is more useful than a large salary-equivalent “saving” that never left the operations report.

Do not refuse to report a genuine capacity gain because it has not yet become cash. Report the hours, the work assigned to them, and the quality measures. “Fund capacity when capacity is the objective,” in CFO.com’s wording. Just do not call it EBITDA.

What finance should write down

Put cost and result in one ledger, and give the hours and the earnings their own cells.

For each initiative, record:

Cost for the period, all in: subscription, usage, implementation, and the internal time to run and review the work.
The operational result: the unit of work, the “before” and “after,” the source system, and the comparison you used.
Where the freed time went, if you can tell: overtime down, contractors down, a hire not made, backlog down, additional volume, more review, or simply absorbed into the week.
The financial translation, if any, with its label.
Quality and rework, as their own measures.
A keep, fix, or stop recommendation that does not depend on dressing the hours up as earnings.

If the honest line is “hours were released; we cannot yet show a P&L effect,” that is a decision input. It may still support a keep, especially if quality held and the cost is modest. It does not support an earnings claim in a board pack or an LP letter.

A hypothetical accounts-payable example

The following is a worked illustration only. It is not a client result, and the figures are invented to show the method.

Suppose a portfolio company turned on an invoice-coding tool nine months ago. Average handling time on a clean invoice fell. Finance can see the subscription and the implementation cost. Someone has already multiplied the released hours by a loaded labor rate and called the product an EBITDA benefit.

A more careful pass would split the claim. The change in handling time, compared with the same season a year earlier and with the months just before go-live, and checked against sites that had not yet rolled out, can be labeled measured as an operational result, with the usual limits: this workflow, this period, these sites. Exception rates and rework are checked so that speed is not being bought with errors.

Then the hours. Some of the released time went to exception review, which is work, not a saving. Overtime and contractor bills did not move. No hire was cancelled against an approved plan. Volume was stable, so the capacity was not used to absorb growth. The remaining hours were absorbed into the week. In that state, there is no measured cash saving to report. The salary-equivalent figure stays off the earnings line, or appears only as an estimated capacity value, clearly boxed as incomplete. A keep, fix, or stop call can still be made: keep the tool if the operational improvement is real and the cost is justified as capacity and quality; fix the exception load; and do not take an EBITDA number to the board until finance can show where the rest of the time went.

None of those numbers should be copied out of this illustration. Only the split should be: measure the hours, then ask what happened to them, then label the financial translation separately.

The number to take into the room

The portfolio-company CFO is the one who has to defend the translation. The operating partner can ask for the same split at each company, so that a fund-level conversation does not add hours in one place to margin in another and call the sum a return.

A readable report says what the tools cost, what changed in the work, what, if anything, changed in the accounts, and how sure finance is about each step. Time saved can be part of that report. It just cannot be the earnings line unless the accounts already say so.

If you want a structured pass at this for one portfolio company, the AI Value Baseline is the fixed-scope version of the same work: a handful of initiatives already in use, a reconstructed baseline, and a labeled ledger that keeps hours and earnings from being mixed.

Sources

Figures were checked against the live pages on November 10, 2026.

CFO.com, “A saved hour from AI is not yet a cash saving,” October 20, 2026, a guest post by an accounting faculty member at Shenandoah University. The article draws the managerial-accounting distinction between a labor-efficiency improvement and a cash saving, and it sets out a four-part benefits schedule: realized spending reduction, avoided future spending, released capacity, and quality. A survey statistic from a professional-services firm that appears in the same article is not used here. https://www.cfo.com/news/a-saved-hour-from-ai-is-not-yet-a-cash-saving/833361/
Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, “Generative AI at Work,” The Quarterly Journal of Economics 140, no. 2 (May 2025): 889–942. The published abstract, confirmed on the Stanford Graduate School of Business publication page on November 10, 2026, states that the authors studied 5,172 customer-support agents and found a 15% average increase in issues resolved per hour. https://doi.org/10.1093/qje/qjae044 and https://www.gsb.stanford.edu/faculty-research/publications/generative-ai-work. An earlier NBER working-paper abstract still shows slightly different figures (5,179 agents and a 14% average); this note uses the published journal version.
CloudZero, “Finding the ROI of AI: The Finance Perspective,” 2026 survey of 260 senior finance leaders, 135 of them CFOs. The report states that 22% can tie AI spend to outcomes, 60% are already spending more on AI than they can justify, and 87% say they must tie spend to outcomes within the year. Asked which unit-cost metric mattered most, 54% chose cost per dollar of revenue and 49% chose cost per employee hour. Respondents could pick several metrics. The report states that figures describe associations in a single point-in-time survey, not proven cause and effect. https://www.cloudzero.com/finance-needs-ai-roi-2026-survey-report/
Coupa, “The Strategic CFO Report 2026,” a survey of 600 global finance leaders. The public page reports that 76% say difficulty quantifying AI ROI is hindering further investment. Fieldwork dates were not shown on the page. https://www.coupa.com/resources/the-strategic-cfo-report/