Back to Blog
Metrics5. November 202611 min

How to Validate AI Time Savings: The ROI Interview That Turns Estimated Hours Into Confirmed Hours

Learn how to validate AI time savings with a structured ROI interview that confirms hours with the people doing the work, so your CFO accepts the number.

Why CFOs Reject AI Time Savings Claims

Most AI business cases die in the same meeting. A programme lead presents a slide saying that a thousand employees save forty minutes a day, the finance director asks where the forty minutes came from, and the honest answer is that it came from a survey, a vendor benchmark, or a spreadsheet assumption. The conversation ends there, not because the CFO is hostile to AI, but because the number has no chain of custody.

Finance teams are trained to ask a simple question of any claimed benefit: who confirmed it, and how? A self-reported estimate fails that test for predictable reasons. People overstate savings on tasks they enjoy automating, they forget the time spent reviewing and correcting AI output, and they rarely subtract the hours that quietly moved to another task rather than disappearing. Vendor benchmarks fail for a different reason: they describe a generic user doing a generic task, not your analyst reconciling your ledgers.

The result is a credibility gap. The AI team believes it has delivered value, the CFO believes it has delivered anecdotes, and the budget for the next phase is quietly trimmed. Closing that gap does not require more data. It requires a different kind of evidence, one where the hours are confirmed with the person doing the work, tied to a named task, and recorded in a form that finance can audit. That method is the ROI interview, and it is the subject of this article.

Estimated Hours Versus Confirmed Hours

The distinction matters, so it is worth defining precisely. An estimated hour is a projection. It is derived from task descriptions, role benchmarks, or an assumption about how much of a process a model can handle. Estimates are useful and necessary at the opportunity stage, because you cannot interview anyone about work that has not yet changed. They tell you where to look and how large the prize might be.

A confirmed hour is an estimate that has been tested against reality by the person who performs the task. The performer states how long the task took before AI, how long it takes now including review and rework, how often it occurs, and what happens to the freed time. The confirmation is recorded with a name, a date, and a task definition. It is still a human statement, but it is a specific, attributable, challengeable one, which is exactly what finance can work with.

There is a third category that CFOs care about most: the measured saving, where system data such as ticket cycle times, throughput counts, or tool telemetry corroborates what the interview recorded. The most defensible ROI narrative moves each claim up this ladder, from estimated to confirmed to measured. Fronterio treats this as a loop rather than a one-off exercise. A savings estimate from a scan is the starting point, the interview converts it to a confirmed figure, and ongoing usage and outcome data harden it into a measured saving. Each rung changes how much weight a sceptical reader can place on the number.

Anatomy of a Structured ROI Interview

A good ROI interview is short, scripted, and repeatable. It should take fifteen to twenty minutes per workflow, be conducted with the person who actually performs the task rather than their manager, and follow the same question set every time so that results are comparable across teams. Ad hoc conversations produce ad hoc numbers.

The core questions fall into five groups. First, task definition: what exactly is the task, what triggers it, and what is the output? Second, the baseline: how long did the task take before AI, and how do you know, whether from a timesheet, a ticket history, or memory of a typical week? Third, the current state: how long does it take now, counting prompting, checking, correcting, and any approvals that the AI output now requires? Fourth, frequency and volume: how many times per week or month does this happen, and has that changed since adoption? Fifth, the destination of the freed time: was it redeployed to higher-value work, absorbed by growing volume, or simply not used?

That last question is the one most programmes skip, and it is the one finance will ask. Hours saved are not the same as capacity released, and capacity released is not the same as cost avoided. Recording the destination of the time lets you classify each saving honestly. Interviewers should also capture a confidence rating and one piece of corroborating evidence, such as a ticket queue, a report, or a before-and-after document, so the claim can be spot-checked later.

Running the Interview Without Leading the Witness

The quality of the data depends on the quality of the interviewer. People want to be helpful, and if they sense that a large saving is the desired answer, they will provide one. The interviewer's job is to be mildly sceptical and entirely neutral, and several techniques help.

Ask for ranges before point estimates. Asking how long the task took before and after, then asking for the best and worst case, exposes how much uncertainty sits behind a number. Anchor on a recent, specific instance rather than a general feeling: walk through last Tuesday's report rather than discussing reports in the abstract. Ask what the person does when the AI output is wrong, because the answer reveals the real review overhead. And ask whether the saving holds for colleagues, since an enthusiastic power user is rarely representative of the median employee.

It also helps to separate the interviewer from the sponsor. If the programme lead who needs the ROI to look good conducts the interviews, a CFO will discount the results. Having a finance business partner sit in on a sample, or having a champion from another team conduct the interviews, adds independence at little cost. Fronterio's champion network is often used for exactly this purpose, because champions know the work but have no stake in a particular team's number.

Finally, record the interview as structured fields, not as free-text notes. Free text cannot be aggregated, filtered, or audited, and it will not survive a finance review.

Turning Interview Data Into a CFO-Ready Number

Once the interviews are complete, the arithmetic is straightforward but the discipline is in the adjustments. Start with the confirmed hours per occurrence, multiply by confirmed frequency, and subtract the review and rework time the interviewee reported. Then apply a realisation factor that reflects the destination of the freed time. If a team reports that half of the saved hours were redeployed to measurable work and half were absorbed, only the first half counts toward capacity value, and only a fraction of that counts toward cash.

Present the result in tiers rather than as a single headline. A typical CFO-ready summary shows gross confirmed hours, net confirmed hours after review overhead, the portion redeployed to named priorities, and the portion that translates to avoided hiring, reduced contractor spend, or retired licence cost. Each tier is smaller than the last, and showing that shrinkage openly is what builds trust. A programme that volunteers its own haircut is believed far more readily than one that makes finance find it.

Value the hours at a loaded cost rate agreed in advance with finance, not one chosen by the AI team. Disclose the sample size and the coverage: eighty confirmed interviews across six workflows is a stronger statement than a company-wide extrapolation from ten enthusiasts. Where a measured source exists, such as cycle time in a ticketing system, show it next to the interview figure and explain any variance. Agreement between the two is persuasive, and disagreement, honestly explained, is nearly as persuasive.

Closing the Loop From Estimate to Measured Saving

The interview is not the end of the process. Its real value appears when it is embedded in a loop that keeps improving the evidence. The savings scan identifies where AI is likely to be worth money and produces estimated hours. The interview tests those estimates with practitioners and produces confirmed hours. Usage telemetry, workflow data, and outcome metrics then provide the measured layer, and the gap between estimate and confirmed becomes a calibration signal for the next scan.

That calibration is underrated. If interviews consistently confirm only sixty percent of estimated hours in a particular function, you have learned something concrete about how to discount future estimates in that function, and your next business case will be more accurate before anyone is interviewed. Over several cycles the estimates become more credible because they are continually corrected by confirmed data, and the CFO sees a programme whose forecasts improve rather than a programme that keeps asking for belief.

Inside Fronterio, each opportunity carries its status on the ladder, so a board pack can show how much of the claimed value is estimated, how much is confirmed, and how much is measured. Re-interviewing at ninety days matters as well, because savings decay when workflows change, when models are swapped, or when novelty fades. A saving that was confirmed once and never revisited is still a claim; a saving confirmed, measured, and re-confirmed is an asset that finance can carry in a forecast.

Common Failure Modes and How to Avoid Them

Several mistakes recur across organisations that attempt this method. The first is interviewing only volunteers. Enthusiasts self-select, and a sample drawn from them will overstate the average saving. Draw the sample from a roster of licence holders or workflow owners, and include people with low usage, whose answers are often the most informative about friction.

The second is double counting. The same hour can appear in the writing assistant saving, the meeting summariser saving, and the email drafting saving if each tool is assessed in isolation. Anchor every interview to a task rather than a tool, and reconcile the tasks against a person's actual working week so that total claimed savings never exceed plausible capacity.

The third is ignoring quality effects. Some AI uses do not save time but improve accuracy, reduce rework, or shorten response times to customers. Forcing these into an hours metric understates them, so record them as separate benefit types with their own evidence. The fourth is neglecting governance. If an interviewee discloses use of an unapproved tool or sensitive data, that is a finding in its own right, and under the EU AI Act's Article 4 duty to ensure AI literacy among staff and the deployer duties in Article 26, it belongs in your governance record rather than being quietly ignored. Run the interview as an ROI instrument, but treat what it reveals as intelligence for the whole AI estate.

Frequently asked questions

How do you validate AI time savings?

Validate AI time savings by moving each claim from estimated to confirmed to measured. Interview the person who performs the task using a standard script, recording the baseline time, the current time including review and rework, the frequency, and where the freed time went. Corroborate with system data such as ticket cycle times or tool telemetry where available. Then apply agreed cost rates and a realisation factor so the final figure reflects what finance can actually recognise, not just gross hours.

Why don't CFOs trust self-reported productivity gains from AI?

Self-reported gains lack a chain of custody. Employees tend to overstate savings on tasks they like automating, forget review and correction time, and rarely account for hours that simply shifted to other work. Surveys also suffer from selection bias, because enthusiastic users respond more often. CFOs accept claims that are attributable to a named task, challengeable, and corroborated by independent data, which is why structured interviews and measured sources carry far more weight.

What is the difference between estimated and confirmed hours saved?

Estimated hours are projections based on task descriptions, benchmarks, or assumptions, and are useful for prioritising opportunities. Confirmed hours have been tested with the person doing the work, who states the before and after time, frequency, and review overhead. Confirmed figures are recorded with a name, date, and task definition. A further step, measured savings, adds system data that corroborates the interview and is the strongest evidence a finance team will accept.

How many employees should you interview to validate AI ROI?

There is no universal number, but aim for coverage across every workflow you plan to claim value from, with enough interviews per workflow to see the spread. Many teams start with eight to twelve practitioners per workflow, drawn from a roster rather than volunteers, and include low users. Disclose the sample size alongside the result. A smaller, representative, independently sampled set is more persuasive than a large self-selected survey.

Are hours saved by AI the same as cost savings?

No. Hours saved become capacity, and capacity only becomes cash if it is converted into avoided hiring, reduced contractor spend, retired licences, or measurable redeployed output. If freed time is absorbed by rising volume or simply unused, the financial benefit is lower or zero. Present hours in tiers, gross, net of review time, redeployed, and cash-equivalent, so the CFO can see exactly how much is recognised.

How often should AI time savings be re-validated?

Re-validate at around ninety days after the first confirmation, then at least twice a year or whenever a workflow, model, or tool changes. Savings decay as novelty fades, processes shift, or vendors update models. Regular re-confirmation also calibrates your estimates, showing how much of a projected saving typically survives contact with real work, which improves the credibility of every future business case.

Ready to get started?

Fronterio helps you implement everything discussed in this article, with built-in tools, automation, and guidance.