Nobody in finance dismisses a training model because the arithmetic is wrong. The arithmetic is one line and everyone gets it right. They dismiss it because every input came from the people who want the program approved, and because nothing in it says what would have happened anyway.
So build the thing they test. Three columns do most of the work: where each number came from, what the comparison case is, and which budget line the benefit lands on. An AI training ROI calculator gives you none of the three.
What finance is actually testing
A reviewer reads an AI enablement model with five questions in mind, and only the last is about the number.
- Provenance. Which cells came from a system and which came from a survey.
- Comparison. What the untrained version of this quarter looks like.
- Conversion. Which budget line gets smaller or which output gets larger, and who owns it.
- Symmetry. Whether the model can produce a negative, and under what assumption.
- Ownership. Whose name sits beside each assumption, and when it is re-read.
Clear the first four and the fifth is a formality. Fail the first and the rest of the meeting is about your credibility rather than the program. In the programs we run, the model that survives review is almost never the one with the biggest number.
Why the calculators fail the first question
Six of the nine results on page one for this search are input forms. They compute the same product of headcount, hours saved and an hourly rate, so the contest is over which defaults ship inside them. The defaults are the problem.
AI CERTs offers three productivity presets, topping out at 30 percent for “AI transformation of workflows,” applied across a default 2,080-hour year. Run the division: that is 624 hours per employee per year, 12 hours out of every 40-hour week.
Put it beside the only nationally representative measurement of the same thing. Alexander Bick, Adam Blandin and David Deming, writing for the Federal Reserve Bank of St. Louis in February 2025, found generative AI users reported saving 5.4 percent of their work hours, about 2.2 hours a week, or 112 hours a year. The preset runs 5.6 times that, and what it exceeds is a self-report, already a ceiling rather than a result.
Tailabs is cleaner and still unsourced: its defaults turn a $26,400 program into $106,145 of “annual capacity value.” OneRange does real work, asking for a loaded rate and an untrained comparison group, yet even it argues that cutting the hours input by 75 percent still clears 5x. A test that cannot fail is not a test.
The arithmetic was never the hard part. The inputs do all the work, and the inputs are what gets audited. The wider case against the format sits in our piece on where ROI calculators go wrong on AI projects.
Rank every input by how you got it
Add a column called Source and sort every cell into one of four tiers, because a reviewer will do it anyway and would rather you went first.
| Tier | Where the number came from | What a reviewer does with it |
|---|---|---|
| 1 | A timestamp or count in a system of record: ticket queue, CRM stage, version history | Accepts it |
| 2 | An artifact counted by someone other than the trained person: drafts, files reviewed, proposals sent | Accepts it with a date range |
| 3 | A manager’s observation of one named recurring task | Accepts the direction, not the size |
| 4 | A participant’s estimate of time saved | Discounts hard, or strikes the row |
One rule governs the column: no cell may sit at tier 4 if a tier 1 number exists anywhere in your systems. Usually one does. A weekly client summary has a file history; a proposal has a CRM stage change. Reaching for a survey when a timestamp exists is the fastest way a model loses the room.
The case for discounting tier 4 is uncomfortable and worth citing anyway. In METR’s randomized trial, 16 experienced open-source developers worked 246 tasks in repositories they knew well. They forecast a 24 percent speedup, estimated 20 percent afterward, and measured 19 percent slower. METR has since called that slowdown outdated for current tools, so read it narrowly: where perception and reality were both measured, they differed by 39 points, in the flattering direction.
Price the cost side from your own payroll
Use loaded cost, not wages. The Bureau of Labor Statistics puts total compensation for private industry workers at $46.89 per hour worked as of June 2026: $32.82 of wages (70.0 percent) and $14.07 of benefits (30.0 percent). Value participant time at a wage rate, as AI CERTs asks you to, and you understate it by about 30 percent. Use your own payroll figure and say so in the cell.
Then put participant hours in the model, because they are usually the biggest line. A 35-person group through a six-hour session plus two hours of pre-work and follow-up, at that loaded average:
| Line | Who pays it | Amount |
|---|---|---|
| Program invoice | External | $9,000 |
| Participant time, 280 person-hours at $46.89 | Payroll | $13,129 |
| Internal coordination and document prep, 30 hours | Payroll | $1,407 |
| Total cost base | $23,536 |
The invoice is 38 percent of the commitment. Presenting the $9,000 and omitting the other $14,500 is a 20-second discovery for anyone who has priced a project, and it costs you the benefit of the doubt everywhere else. Participant time is the line we most often see missing from a draft model, usually because the vendor’s quote never mentioned it.
It also changes which quote is cheaper, since a shorter session with the same effect wins on the line that dominates, the argument in our breakdown of cost per head against cost per hour saved. For scale, ATD’s 2026 State of the Industry report put direct learning spend at $846 per employee for 2025.
Hours saved are not money saved
This is the step the arithmetic hides, and where most AI enablement business cases quietly fail. Multiplying hours by a rate produces a number, not a return. The return exists only if one of three things happens, and each has a different owner.
Cost avoided. A line already in the budget gets smaller: contractor hours, overtime, a renewal left untaken. Owner: whoever controls that line. The only path finance treats as cash without argument.
Capacity redeployed. The hours go to named work with an owner and a date. Not “more strategic work.” The migration that keeps slipping, the backlog of client reviews. Owner: the manager who reassigns the time.
Throughput. More units of a counted thing at the same cost: proposals out, claims processed, matters opened. Owner: whoever reports the count today.
The rule that makes the model honest: a path with no named owner and no date scores zero. Write the zero down. One funded path and two zeros persuades better than three paths of hopeful arithmetic, because the reviewer can see you ran the test rather than arguing around it.
The same St. Louis Fed work shows how far claims shrink on the way up: 5.4 percent of hours saved among users became 1.4 percent across all workers and an implied 1.1 percent rise in aggregate productivity. Much of that is adoption rather than failed conversion, so it is no coefficient to borrow, but your model moves a personal claim up two levels and it will thin on the way.
The comparison case you can afford
Finance does not need a clinical trial. It needs to see you considered the other explanation. The Phillips ROI Methodology, the standard with a literature behind it, makes that a required step: isolate the program’s effect with a control group, a trend line, forecasting, or estimates adjusted for credibility.
At 50 to 500 people the affordable version is a staggered start. Train one group in month one and the next in month two, measure both in both months, and the untrained group is your comparison at no extra cost. Brynjolfsson, Li and Raymond read a staggered rollout that way across 5,172 support agents in the Quarterly Journal of Economics in May 2025, finding issues resolved per hour up 15 percent and novices gaining most. Where staggering is impossible, a trend break on a metric you already track carries weight, provided you stop short of claiming cause.
Then handle the confound no calculator mentions. Your licenses are already paid for, so the comparison case is the same licenses with an untrained team, not an AI-free company. That disqualifies most of the headline research as evidence for your training line. Noy and Zhang’s randomized study in Science in 2023, with 444 professionals on writing tasks, found time down 0.8 standard deviations and quality up 0.4. That gain belongs to the tool you already bought.
The consequence is a date. Baseline after the licenses are live and before training starts, because baselining earlier bills the software’s gain to the training budget, and anyone can spot it from the invoice date. The protocol is in our measurement guide for proving capability moved.
Carry a rework line
A model that cannot produce a negative is not a model. Three cells fix that, and two have published figures behind them.
Error cost. In the field experiment Fabrizio Dell’Acqua and colleagues ran with 758 BCG consultants in 2023, tasks inside the tool’s capability range came back 12.2 percent more numerous and 25.1 percent faster. Outside it, consultants using the model were 19 percentage points less likely to reach the correct answer. Untrained use is how work ends up outside the range, so the cell argues for the program. Model it as review hours on a named class of output.
Rework from poor output. BetterUp Labs and Stanford’s Social Media Lab, writing in Harvard Business Review in September 2025, surveyed 1,150 full-time US employees, and 40 percent had received AI-generated work poor enough to need redoing in the past month. The authors put the drag at roughly $9 million a year at a 10,000-person company, about $900 per employee. At that rate a 35-person group loses near $31,500 a year, more than the whole cost base above. Untrained use costs more than fixing it, which hands a reviewer a reason to fund the program that does not rest on your hours estimate.
Decay. Capability fades when nothing reinforces it, so running a day-30 gain flat across 46 weeks claims something you have not measured. Pick a number, label it an assumption, and let finance argue with that one cell instead of doubting all of them.
Report the break-even, not the return
Lead with the question a reviewer asks third: how little has to be true for this to pay. Take the $23,536 cost base, 35 people, a 46-week year and the same loaded rate, then solve for the hours per person per week the program has to return.
| Share of saved hours that converts to cash | Hours per person per week needed to clear year one |
|---|---|
| 100 percent | 0.31 (19 minutes) |
| 50 percent | 0.62 (37 minutes) |
| 30 percent | 1.04 (1 hour 2 minutes) |
| 15 percent | 2.08 (2 hours 5 minutes) |
Read the bottom row against the national self-report. At 15 percent conversion, year one clears only if your team reaches 2.08 hours a week, 95 percent of the 2.2 hours US users claim. That row is the honest edge of the case, and showing it makes the rows above it believable.
Add payback in weeks on the full cost base, since a percentage with no clock cannot be ranked against the other requests on a reviewer’s desk. What counts as defensible is in our discussion of the payback period to expect from enablement work; the approval side sits in our guide to underwriting an enablement budget a CFO will approve.
The page that gets signed
The model is a spreadsheet. What finance signs is the one page in front of it, carrying five things.
- A version and a date. An undated model is an opinion.
- An owner per assumption. Conversion rate, decay, adoption, hours returned: a named person per row, not a department.
- A pre-registered threshold. The number that counts as a pass, written before the program runs. Almost nobody does this, and it separates measurement from storytelling.
- A re-review date. Day 90 for capability, day 180 for whether day 90 was real.
- The decision that changes if the number misses. Scope cut, different group, no renewal. A model with no attached decision is a report, and reports do not get funded twice.
Build it where your reviewer can open and edit it, with the Source column visible. The point is not winning the first meeting; it is that the second one, six months later, argues over the same cells.
For a second pair of eyes on the inputs before you take it in, that is what a free AI-readiness call is for. What makes the hour worth it is the gap between a paid license and a changed habit, the case in our argument that buying licenses is not the same as building capability.
Frequently asked questions
Should I use an AI training ROI calculator at all?
Use one to sketch, never to submit. The defaults decide the answer before you type: one popular calculator’s top preset works out to 624 hours per employee per year, 5.6 times what US generative AI users self-report saving.
How should I value an hour saved?
At loaded cost from your own payroll, and only for hours that landed somewhere. BLS data for June 2026 puts private-industry compensation at $46.89 an hour, 30.0 percent of it benefits, so wage-only models understate time by about a third.
Do we really need a control group?
You need a comparison case, which is cheaper. Train one group in month one and another in month two, measure both in both months, and the untrained group is your comparison for free. Failing that, a trend break on a metric you already track works if you claim association rather than cause.
How do I separate the training’s effect from the tool’s effect?
By when you baseline. Take the before-reading after the licenses are live and before the first session, so the software’s gain sits inside the baseline and what follows belongs to the program.
What discount should I apply to self-reported time savings?
Enough that you would defend it unprompted. Where perception and measurement were both captured, developers estimated a 20 percent speedup and measured 19 percent slower. Better still, do not need the discount: find the timestamp.
How long before the model holds real numbers?
Two weeks of baseline, 90 days for a capability reading, 180 days before you know whether the 90-day figure held. Anything financial inside the first month is arithmetic on an assumption.
What about the adoption numbers in our admin console?
Use them to run the program, never to price it. Seats activated and weekly active users tell you whether practice is happening and nothing about whether output improved. Pair each with a scored output or leave it out.
What if the model does not clear?
Shrink it until it does, or decline. One role, a dozen people, one weekly task with a timestamp behind it. A small program with a defensible number buys the budget for a larger one better than a large program with a story.
Key takeaways
- Finance rejects training models on provenance, not arithmetic. Add a Source column and tier every input before anyone asks.
- Never use a self-report where a system timestamp exists. Where both were measured, perception and reality differed by 39 points.
- Participant time at loaded cost usually exceeds the invoice: above, a $9,000 quote is 38 percent of a $23,536 cost base.
- Hours become money three ways only, and a path with no named owner scores zero.
- Your licenses are already paid for, so the comparison case is the same licenses untrained.
- Report the break-even hours per person per week rather than the return. It is what a reviewer can argue with.
Start with the Source column. Fill it in for the three cells you trust least, and what you carry into the room is a different document from the one a calculator hands you.
Dirk Jan van Veen, PhD
