Search for an AI training business case template and you get a document for a decision you already made. Thirteen of the eighteen results we swept are templates for buying AI software. In the one with the most detailed cost model, training is a single line called change management.
Your situation is the other one. The licenses are bought, the invoice recurs, and someone has asked why the thing already approved needs a second check. That is a utilization case, and it fits on one page. Below is the artifact: seven blocks, a word budget each, and the sentence that belongs in every field.
Why this page is not an AI project case
An AI project case argues that a capability should exist. An AI enablement case argues that a capability you already pay for is not being used, and that the gap is a training gap rather than a software one.
That decides what evidence you may cite. Most published AI productivity research measures the tool against no tool, so it belongs to the software line you already approved. Borrowing it bills the vendor’s gain to your training budget, and the invoice date gives that away in seconds.
What you can cite is the gap itself. MIT’s Project NANDA, in The GenAI Divide: State of AI in Business 2025, found 95 percent of generative AI pilots showed no measurable profit-and-loss impact against $30 to $40 billion of enterprise spend. The report puts the cause in learning and workflow rather than model quality. Read it as direction rather than a population estimate; the method mixes 150 interviews, a 350-person survey and 300 public deployments. Direction is enough, because the case sits in the distance between a paid seat and a changed habit, the argument in our piece on why buying licenses is not the same as building capability.
One page is the brief, not the style
The pages in this corner open on the one-page promise and then prescribe nine, the best-structured of them listing seven sections of one to two pages each.
Hold the limit, because the limit does the editing. Four hundred words is what an approver reads while the meeting is starting, and what survives is what they were going to weigh anyway.
Send it where the reviewer can edit the cells. A PDF invites comments; a live sheet invites a counter-number, and a reviewer who types their own figure in has joined the model.
The seven blocks at a glance
| # | Block | Words | What it has to do |
|---|---|---|---|
| 1 | Current state | 60 | Name what you pay for and what you cannot show |
| 2 | The ask | 40 | Scope, group, duration, dollar range |
| 3 | Cost base | 60 | Invoice plus participant time at loaded rates |
| 4 | Value paths | 90 | One to three paths, each with an owner |
| 5 | Pass threshold | 50 | The number that counts, written in advance |
| 6 | Decision on a miss | 40 | What changes if the threshold is not met |
| 7 | Signature line | 30 | Owner, version, date, re-review date |
That is 370 words and a page. Blocks 5 and 6 almost never appear in a draft, and they are the two that turn a pitch into a commitment. A pitch gets funded once.
Blocks 1 and 2: the state and the ask
Open on the state, not a problem. A problem statement invites debate over whether the problem is real; a state is a fact from a system of record and goes straight to the remedy. Figures in the worked examples below are illustrative, not benchmarks.
Current state. We hold 180 assistant seats at an annual cost of $X. Admin console shows 61 percent activated and 24 percent weekly active. No team reports a changed output. The last session was vendor-led, ninety minutes, no follow-up. We cannot say what the seats changed.
Three features make that paragraph work. Every number is a count from a system rather than a survey, and the failure sits with the enablement history rather than the staff. The last sentence frames the gap as something nobody can answer, which a finance reviewer treats as their own problem.
Then the ask, kept narrow enough to be hard to refuse.
Ask. Two hands-on sessions for the 22-person client-services team, applied to our own proposal and summary documents, plus four weeks of structured practice. Market range for a program this size is $4,000 to $9,000. Sessions cover assistant fluency only: no tool selection, no policy work.
A market range rather than a single quote tells a reviewer you shopped. Naming what the program excludes stops the scope drift that arrives as a question you cannot answer.
Block 3: cost at loaded rates
Put participant time in the cost base at loaded cost, from your own payroll. It is usually the largest line, and omitting it is the cheapest way to lose a reviewer’s trust on everything else.
The Bureau of Labor Statistics puts total employer compensation for private industry workers at $46.89 per hour worked as of June 2026: $32.82 in wages (70.0 percent) and $14.07 in benefits (30.0 percent). Wages alone understate the cost of time by about 30 percent. Use your own figure and say which figure it is.
| Line | Who pays | Amount |
|---|---|---|
| Program invoice | External | $6,500 |
| Participant time, 22 people x 8 hours at $46.89 | Payroll | $8,252 |
| Document prep and coordination, 20 hours | Payroll | $938 |
| Total commitment | $15,690 |
The invoice is 41 percent of the commitment. Showing the other 59 percent unprompted buys the benefit of the doubt in block 4.
Give the figure a scale. ATD’s 2026 State of the Industry report, drawing on 2025 data from 340 organizations, put average direct learning spend at $846 per employee, down from $1,254. At 22 people the program above is $713 a head, inside what finance accepts for all workplace learning combined. Where the money comes from is a separate argument, and the four approved sources are in our budget defense playbook.
Block 4: the three paths that count as value
Hours saved are not money saved, and this is where most cases quietly fail. Headcount times hours times a rate produces a number, not a return.
LinkedIn’s 2025 Workplace Learning Report, built on 937 L&D and HR professionals and 679 learners, puts the test in the reviewer’s own words: how will this initiative help you to make money, save money, or mitigate risk? Three paths, and each needs a named owner.
Save money. An approved budget line gets smaller: contractor hours, overtime, a seat count trimmed at renewal. Owner is whoever controls that line.
Make money. More units of something counted today: proposals out, matters opened, claims cleared. Owner is whoever reports that count.
Mitigate risk. Fewer corrections on work already going out. BetterUp Labs and Stanford’s Social Media Lab, in Harvard Business Review in September 2025, surveyed 1,150 full-time US employees and found 40 percent had received AI-generated work poor enough to need redoing within the past month. Untrained use has a bill, and it lands on reviewers rather than the training line.
The rule that makes the block credible: a path with no named owner and no date scores zero, and you write the zero down. One funded path and two zeros reads better than three paths of hopeful arithmetic, because the reviewer can see the test was run.
The same report notes that engagement and retention are still the most common measures of learning impact. Neither answers the three-way test, so keep both out of this block. The AI training ROI arithmetic behind the block belongs in a separate sheet, which we work through in building a model finance cannot dismiss.
Blocks 5 and 6: the threshold and the attached decision
Write the pass mark before the program runs. It is the one field separating a measurement plan from a story, and almost nobody includes it.
Pass threshold. By day 90: 70 percent of the 22 trained staff using an assistant weekly on a named task, and median proposal turnaround down from 4.5 days to 3.0 on the 40 most recent, measured from the same CRM stamp as the baseline.
Two properties matter. The metric already exists in a system, so nobody builds reporting to settle the question. And the baseline is taken after the licenses go live and before the first session, which keeps the software’s gain inside the baseline where it belongs.
The threshold above pairs a reach number with an output number on purpose. Measuring AI adoption alone tells you practice is happening and nothing about whether the work improved. A defensible day-90 pair sits in our read on adoption benchmarks at six and twelve months, and the full protocol in our measurement guide for proving capability moved.
Then attach the decision, in one sentence.
If missed. We do not extend to the second team, we cut the renewal of 40 unused seats, and we take the diagnosis rather than a repeat session.
A reviewer funds a case faster when a miss has a stated consequence, because the downside is bounded on paper. A case with no attached decision asks for trust twice: now, and again when the number comes in soft.
Jack Phillips’ ROI method runs five levels: reaction, learning, application, business impact, and return. The chain breaks at the lowest unmet level, so satisfaction scores, which are level one, can never sit in block 5. Most AI enablement metrics stall at level three, and that is the floor your threshold has to clear.
Block 7: the signature line
Thirty words, and most drafts leave out at least two of them.
- Owner. One name against the whole page, not a department.
- Version and date. An undated case is an opinion.
- An owner per assumption. Conversion, decay, adoption: a name per row, in the sheet.
- Re-review date. Day 90 for whether capability moved, day 180 for whether day 90 held.
The re-review date is what makes the second program easier to fund than the first: it books the meeting where the result gets read in public.
Three sentences to cut before you send it
Almost every draft carries these, and each costs more than it earns.
- The headline industry statistic. A percentage from a vendor report invites the reviewer to audit the citation instead of the proposal. Your own console numbers in block 1 are stronger, because they are about you.
- The tool list. Naming four assistants makes the case look like a software request, which routes it to IT and restarts the clock. Describe the work, not the logos.
- Anything containing “transformation.” It signals a program with no end state, and a reviewer prices open-ended commitments at a premium. Two sessions and a named metric is a thing that finishes.
Drop the three-scenario model too. Conservative, base and upside is standard in software cases and works against you here: the upside column becomes the only one anyone discusses.
Routing: who reads it before the meeting
The document’s job in the room is to be understood already, which depends on who saw it first.
| Reader | When | Must not be surprised by |
|---|---|---|
| Manager of the team being trained | First | Hours their staff spend offline |
| Owner of the value-path budget line | Next | Their line named in block 4 |
| Finance partner or controller | Before submission | The loaded rate, and time in the base |
| IT or security | Only if block 1 names a system | That this is not a procurement request |
| The approver | Last | Nothing at all |
Pre-clearing the value path matters most. A reviewer who hears their own budget line named for the first time in the meeting will defend it, and the case stalls over a point you conceded on paper.
If block 1 is thin because nobody has read the current state, run a short baseline first; scope and market pricing are in our piece on what a capability audit covers and what it costs.
What the second meeting argues about
The real test of this page is day 90, when the threshold from block 5 is read against what happened. A well-built case makes that meeting an argument about two or three cells. A pitch makes it an argument about whether the program was ever a good idea.
Bring three things: the same baseline source, the threshold exactly as written, and the decision from block 6 either executed or waived with a stated reason. Waiving it quietly is how a second program stops being fundable.
If you want the inputs tested before a reviewer tests them, that is what a free AI-readiness call is for: an hour on your console numbers, your comparison case, and whether the threshold you picked is readable in ninety days.
Frequently asked questions
How long should an AI enablement business case be?
One page, around 400 words, with everything else in an appendix. What fits is what an approver weighs anyway, and a nine-page case signals an open-ended commitment.
Can I use a generic AI business case template?
Only for the shape. Most justify buying AI software and treat training as a change-management cost line, so their value section argues the tool against no tool. Your licenses are already paid for, which voids that comparison.
What goes in the value section if we have no data yet?
Named paths with named owners, and a zero against any path you cannot staff. One funded path beats three estimated ones, because the reviewer can see the test was applied.
How do I price participant time?
At loaded cost from your own payroll. BLS figures for June 2026 put private-industry compensation at $46.89 per hour worked, 30.0 percent of it benefits, so a wage-only figure understates time by a third.
Should I include a cost of doing nothing?
Yes, if you can count it. Seats paid for and unused is a real line on your own invoice. Survey-based risk framing is weaker, though the finding that 40 percent of US employees received AI work needing redoing within a month sizes the review burden.
Do I need a control group to get this approved?
No, but you need a comparison case. Train one group this month and the next the following month, measure both in both months, and the untrained group is your comparison at no cost. Isolating the program’s effect is a required step in the Phillips method.
What if the approver asks for a three-year model?
Give one in the appendix and keep the page as is. A three-year view on a two-session program invites forecasts you cannot defend. The stronger answer is a short first phase with a threshold and a decision attached.
What is the most common reason these get rejected?
Provenance. A number with no stated source in block 1 or 3 moves the meeting onto your credibility, and nothing later in the page recovers it. Every cell should name where it came from.
Key takeaways
- This is a utilization case, not an adoption case. The licenses are paid for, so tool-versus-no-tool research belongs to the software line, not the training ask.
- Seven blocks, roughly 400 words. Blocks 5 and 6, the pass threshold and the decision on a miss, turn a pitch into a commitment.
- Put participant time in the cost base at loaded cost. Above, a $6,500 invoice is 41 percent of a $15,690 commitment.
- Value has three paths: save money, make money, mitigate risk. A path with no named owner scores zero, and you write the zero down.
- Cut the headline statistic, the tool list, and the word transformation, then pre-clear the value-path budget line before the meeting. The approver should be the last to read it and the only one with no surprises.
Arthur Wandzel
