Underwriting an AI enablement budget means doing what a credit officer does with a loan: state every assumption, apply a haircut to each one, and show what the deal looks like when the assumptions are wrong. Most business cases do the opposite. They pick the best published multiple, apply it to headcount, and arrive at a number nobody in the room believes.
The honest version is less flattering and harder to refuse: four inputs, two of them on the cost side that vendor proposals never mention, ending in a payback month rather than a percentage.
What underwriting an AI enablement budget means
A business case argues for a decision. An underwriting memo prices a risk. Your CFO reads dozens of the first kind every quarter and almost none of the second, and the second is the one that gets funded.
Three habits are worth borrowing. Every number carries a named source, so the debate is about the source rather than your optimism. Every assumption carries a stated discount, which flips the burden: your reviewer has to argue the haircut is too small rather than that the forecast is too big. And the downside case is written down, in your own hand, before anyone asks.
Then start from a figure nobody will dispute. Run 120 Microsoft 365 Copilot seats at the published $30 per user per month on an annual commitment and that is about $43,000 a year, on top of base licenses whose prices Microsoft raised in July 2026. A program protecting that spend is a small fraction of it, and the license invoice is a number finance already has in a system they trust.
The four inputs
Resist the dashboard. The model that gets signed has four inputs and fits on one page.
- People who will finish, not people invited. Count the ones whose managers have cleared the time.
- Hours of their attention the program consumes, pre-work and follow-up sessions included.
- Loaded hourly cost of those people, from payroll rather than from a benchmark.
- Hours per week the trained work gets back, after the haircuts two sections down.
Everything else is commentary. Role count, vendor, format and session design decide whether the program works, and they reach this model only through those four numbers. What drives the quote itself sits in our breakdown of what an AI enablement program costs and how the number is built.
The cost base an AI enablement quote leaves out
Published ROI frameworks model the value side in forensic detail and treat total cost as the invoice. Your team’s hours are the larger line, and they belong in the model.
Use payroll. The Bureau of Labor Statistics puts wages and salaries at 70.0 percent of employer compensation cost for private industry workers as of June 2026, with benefits making up the other 30.0 percent. A $90,000 salary is therefore roughly $128,600 of employer cost, about $62 an hour across a 2,080-hour year. Most ROI templates use a wage rate and understate the cost base by nearly a third.
A cost base for a 220-person company sending 60 people across three roles through a 10-hour program:
| Line | Who pays it | Amount |
|---|---|---|
| Program invoice | External | $11,000 |
| Attendee time, 600 person-hours at $62 | Payroll | $37,200 |
| Internal coordination and document prep, 25 hours | Payroll | $1,550 |
| Total cost base | $49,750 |
The invoice is 22 percent of the commitment. Walking in with the $11,000 and no mention of the other $38,750 is the fastest way to lose a CFO who will spot it in ten seconds. Put it in yourself. Vendors who split their own quote into facilitation, setup and reinforcement make this easier to assemble, which is one reason to insist on a proposal with real line items before you model anything.
Where AI enablement ROI comes from
Where the return comes from, ranked by how well each source survives review.
Hours returned on recurring work. The strongest, because the task existed before the program and has a known frequency. A weekly client summary that took 90 minutes and now takes 35 is auditable: the artifact exists both ways and someone can time it.
Cycle time on work that waits on a person. Better than hours saved when the queue is the constraint. If proposals sat three days waiting on a draft and now sit one, the gain shows up in a system of record rather than in a survey.
Error reduction and headcount avoidance. Leave both out. Error reduction is unprovable unless you already count errors, and headcount avoidance reads to a CFO as a promise not to fill a role, which you cannot make for other managers. Offer them as unmodeled upside.
Missing from that list: adoption. Seats activated, prompts sent and weekly active users are useful for running a program and worthless as return. McKinsey’s 2026 global AI survey, fielded between May 4 and June 8 2026 across 1,719 respondents, found 80 percent saying AI had improved their own productivity while only 37 percent reported a positive contribution to EBIT, a share flat against 2025. Six percent qualified as high performers attributing 5 percent or more of EBIT impact to AI. That gap between felt productivity and booked results is the central fact of this exercise, and it is why your reviewer is right to discount.
What closes that gap is the argument in our guide to moving from license rollout to measurable team capability.
The haircut finance will apply, and should
Three discounts, in sequence, on any hours-saved figure. Apply them yourself and the model stops reading as a pitch:
Retention. Some people stop using the method once sessions end. Assume 30 percent fall away by 90 days unless reinforcement is in scope and you have evidence it holds.
Self-report. Halve anything that comes from asking people how much time they saved. People are poor at this in a consistent direction.
Redeployment. An hour saved is worth nothing until it is spent on something that matters. With a visible backlog, count most of it. Without one, count half or less, and say so. This is the discount vendor models never apply and the one a CFO reaches for first.
Then there is the number you should decline to use. Search for the return on AI spend and you meet $3.70 back per dollar invested, with top-tier users at $10. That figure comes from a 2024 IDC study commissioned by Microsoft. It may well be defensible, and it is still a software vendor’s number about software spend. Quoting it to finance costs more credibility than it buys. In the programs we run, the budgets that clear are the ones where the sponsor brought the discounts rather than the multiples.
Three scenarios and a payback month
Same 60 finishers, same $49,750 cost base, three sets of assumptions. Hours per week are per trained person across a 46-week working year.
| Downside | Base | Upside | |
|---|---|---|---|
| Hours returned per week | 0.5 | 1.5 | 3.0 |
| Still using at 90 days | 60% | 70% | 80% |
| Kept after self-report haircut | 50% | 50% | 60% |
| Recovered hours redeployed | 40% | 50% | 60% |
| Realized hours per year | 166 | 725 | 2,385 |
| Value at $62/hour | $10,300 | $44,900 | $147,800 |
| Months to payback | 58 | 13 | 4 |
Read the downside column honestly: on hours alone, the program does not pay back. That is the true shape of the decision, and hiding it gets business cases killed on the second question. The case for signing is that the base case clears a year, the upside is large, and the cash at risk in the downside is the $11,000 invoice, roughly three months of the license bill the company already pays.
That table asks a reviewer for a probability judgment rather than for belief. Most finance leaders are glad to have that conversation; most training proposals never start it.
The biggest lever is not the vendor’s price
Attendee time is 75 percent of the cost base, so the variable that moves payback furthest is not what you negotiate off the invoice.
Take the same program from 10 hours of attention to 6. Internal time drops from $37,200 to $22,320, the cost base falls to about $34,900, and base-case payback moves from 13 months to 9. Negotiate $3,000 off the invoice instead and payback moves from 13 months to 12.5.
So the question to put to every vendor is not “can you do it cheaper”. It is “what is the fewest hours of my team’s attention in which you can change how they work”. That question also sorts providers who build sessions around your documents from providers reading a deck, because only the first kind can answer it. How different pricing structures hide or expose that number is the subject of our piece on per-seat, per-cohort and retainer models.
Set the baseline before the first session
Every model above collapses without a before-reading, and a baseline captured after training begins is worthless. Two weeks is enough.
Pick three recurring tasks, one per role, that happen at least weekly and produce a file or a record. For each, capture the time a person spends, the number produced per week, and one sample of the output. An hour per role, and it separates a post-program claim from a measurement.
Two rules. Never baseline a task that is about to change for other reasons; you will never untangle the attribution. And never baseline with a survey if a timestamp exists anywhere in your systems, because a ticket queue, a CRM stage or a version history beats self-reporting every time.
Which budget line this belongs on
This decides more approvals than the arithmetic does. Three options, in descending order of how easily they clear.
Against software spend. The strongest position. You are protecting a recurring license commitment you cannot cancel, and the ask is a fraction of it. Census Bureau survey data puts AI use at 32 percent of firms with 100 to 249 employees and 37 percent of firms with 250 or more, so both the spend and the question of whether it is working are normal.
Inside the existing learning budget. Also easy, because it is a reallocation rather than an increase. ATD’s 2026 State of the Industry report put average direct learning spend at $846 per employee for 2025. A program reaching 60 people at $11,000 is about $183 a head, well inside a line that already exists.
As a new line item. Hardest, and usually avoidable. If you end up here, check whether you are asking for a transformation program rather than training, in which case the number and the approver are both different.
When the number comes back thin
Sometimes the base case does not clear. Three honest responses.
Shrink the scope. One role, 12 people, one workflow, one real measurement. A small program that produces a defensible number buys the budget for the larger one far better than a large program that produces a story.
Change the target. If the only workflows you can baseline are low-frequency, the hours are not there. Find the weekly task, not the quarterly one.
Or decline. A program with no measurable task behind it will not acquire one by being approved, and that answer is worth more to your credibility than a funded program that fails in March. It rests on the same argument as treating license spend and capability as separate problems.
Bring the four inputs and we will build the model with you on a free AI-readiness call, including the scenario your CFO will push back on first.
Frequently asked questions
What is a realistic ROI for AI enablement?
Base-case payback between 9 and 15 months, measured on recurring task hours after the three haircuts. Published figures of 3x or 4x come from vendor-sponsored studies that count only the invoice as cost. Put attendee hours in the denominator at a loaded rate and the honest first-year multiple is nearer 1x, which is still a good decision.
How long before the program pays back?
Nine to 15 months in the base case, four to six if the trained work is high-frequency and the team has a backlog to absorb the recovered hours. If your model shows payback inside 90 days, a haircut is missing.
How do I calculate hours saved with no time-tracking system?
Use artifacts and timestamps rather than asking people. Count how many of a given document, ticket or report each person produces per week, then time three before the program and three after. Where a system of record already stamps stage changes, use that and skip the timing exercise.
What should I use as the baseline?
Three recurring tasks, one per role, each producing a file or a record at least weekly, captured over two weeks before the first session: time per instance, instances per week, one output sample.
Why will my CFO discount the productivity numbers I show?
Because the gap between people feeling faster and the P&L moving is well documented. McKinsey’s 2026 survey found 80 percent reporting better individual productivity against 37 percent reporting any positive EBIT contribution. Your reviewer has read something like that, or has watched it happen. Applying the discount yourself removes the objection before it is raised.
Should this come out of the L&D budget or the software budget?
Software, if you have license spend to point at. You are protecting a recurring commitment rather than requesting a new initiative. The learning budget is the fallback and still easier than a new line, since it is a reallocation inside a number already approved.
What share of AI spend should go to enablement?
A planning range of 15 to 25 percent of annual license and tooling spend, which for a company paying $43,000 in seats lands between $6,500 and $11,000. Treat it as a sanity check on a quote rather than a rule; the real driver is how many roles need their own material.
What do I do if the pilot shows no measurable gain?
Find out which of the four inputs broke before concluding the program failed. Usually it is redeployment: people saved the time and refilled it with the same low-value work, which is a management problem rather than a training one.
Key takeaways
- Underwrite rather than pitch: name every source, discount every assumption yourself, and write the downside case before anyone asks.
- Four inputs only. Finishers, hours of their attention, loaded hourly cost from payroll, and hours per week returned after haircuts.
- Attendee time is usually three quarters of the real cost base. At BLS-implied loaded rates, 600 person-hours is $37,200 against an $11,000 invoice.
- Apply three haircuts in order: retention at 90 days, half of anything self-reported, and a redeployment discount on hours with nowhere useful to go.
- Shortening the program beats discounting the invoice. Cutting 10 hours of attention to 6 moves base-case payback from 13 months to 9; a $3,000 discount moves it half a month.
- Decline the $3.70-per-dollar figure, which comes from a 2024 IDC study commissioned by Microsoft, and point the ask at license spend you cannot cancel rather than at a new budget line.
Count your finishers, pull a loaded rate from payroll, and baseline three tasks. A free AI-readiness call turns those numbers into the scenario table above in about half an hour.
Arthur Wandzel
