Approving an AI enablement program is a small decision wearing a large decision’s clothes. The quote is somewhere between $2,000 and $15,000 for a cross-industry workshop. The exposure is 40 people giving up half a day, the manager hours that have to follow, and the renewal argument you will have in nine months whichever way this goes.
We sit on the vendor side of that meeting often, and the pattern holds: most buyers ask four or five questions a vendor can answer with a slide, and skip the two that would change the proposal. Finance is right to be wary. Deloitte surveyed 1,854 senior executives across 14 countries in late 2025 and found only 6% reported payback inside a year, against 38% who expected generative AI returns within a year of investing. Most who got there took two to four years.
A training spend is not a platform build, though, and the questions that protect you on a seven-figure system are the wrong ones here. Below is the list we would want asked of us.
The ten questions, and the answers that should stop you
Most lists like this stop at the questions. The questions are the easy part. What makes one useful in a live meeting is knowing which answer means yes and which means the proposal is not ready.
| Ask | The answer that should stop you | A sound answer |
|---|---|---|
| What is the fully loaded cost, including our people’s hours? | “The investment is $9,000” | A figure roughly double the invoice, with participant and manager hours priced in |
| Which recurring piece of work is this meant to change? | “General AI fluency across the team” | One or two named deliverables, with the people who produce them |
| Who timed that work before we start, and how? | “We’ll establish a baseline afterwards” | A named person, a method, and a date before kickoff |
| Which mechanism produces the return: cost, revenue, risk, or capacity? | “Productivity” | One primary mechanism, named, with the second-order ones labeled as upside |
| Where does your percentage come from? | A case study with no sample size or window | A named study, its population, its window, and why it transfers here |
| Who owns the measurement after you leave? | “We’ll provide a dashboard” | An internal owner with the metric in their objectives |
| What are the pass and fail conditions? | “We’ll review how it went” | Written thresholds, agreed before the money moves |
| When do we review, and with what data? | “In due course” | A dated review with the data source named |
| What does phase two cost if this works? | Silence, or a number that appears later | A quoted range, disclosed now |
| What have you seen fail in a company our size? | “Our approach works” | Two specific failures and what the vendor changed |
Five of those are worth expanding, because they are the ones that reliably get a weak answer.
The cost the invoice does not show
The invoice is the smaller half of what you are approving. Price the hours: the Bureau of Labor Statistics puts employer costs for private-industry workers at $46.89 per hour worked in June 2026, so a half-day session for 40 people is 160 hours of paid attention, about $7,500 before anyone opens a laptop.
| Line | A worked example | Often missing from the proposal |
|---|---|---|
| Vendor invoice | $9,000 for two half-day sessions | No |
| Participant time | 160 hours, about $7,500 | Yes |
| Manager reinforcement | 6 managers, 1 hour a week for 8 weeks, about $2,250 | Yes |
| Internal coordination | 20 hours of someone’s second job, about $940 | Yes |
| Licenses for people who did not have them | Varies | Sometimes |
Call it $19,700 against a $9,000 quote. Not an argument against the spend, just the number to set against the forecast benefit. Judging quotes on the invoice line alone is how two offers that look 30% apart turn out identical, and the difference between cost per head and cost per hour saved is where that comparison usually goes wrong.
Training magazine’s 2025 industry report, surveyed across US organizations of 100 or more employees, put annual per-learner training spend at $782 at midsize companies and $468 at large ones. A $225-a-head workshop is a thin slice of a budget you already approve every year. The question is not whether you can afford it, but whether this one will land.
Who measured the current process, and when
This question separates a proposal from a pitch, and almost nobody asks it before signing.
If nobody timed the Monday pipeline report, the monthly board pack, or the first draft of a proposal before training starts, no claim made later can be checked. You will be handed a survey in which people say they feel faster. That is not nothing, and it is not evidence. People misjudge their own gains badly: a controlled study of experienced open-source developers found them 19% slower with AI access while estimating afterward that they had been 20% faster.
The fix costs almost nothing. Before the sessions, pick two recurring deliverables and record how long they take and who produces them. A week of light timesheet discipline from four or five people is enough. Our measurement guide for proving capability moved sets out the method, and sequence beats sophistication here: a rough baseline taken before is worth more than a precise one reconstructed after.
When a vendor resists, it is usually not evasion. Most have never been asked. The ones who agree on the spot are telling you how they expect to be judged.
Which financial mechanism this runs through
“Productivity” is not a mechanism. Hours recovered become money through exactly one of four routes, and which one you pick changes who has to act after the training.
- Cost out. A vacancy left unfilled, fewer contractor hours, an outsourced line cut. The only route finance books directly, and it needs a named budget line.
- Capacity in. Freed hours go to backed-up work that already has a value. The honest answer for most companies of 50 to 500 people, and it depends on a manager redirecting the time, not on the training.
- Revenue. Faster proposal turnaround, more outreach, shorter cycles. Real, slower to show, and contaminated by everything else sales did that quarter.
- Risk. Fewer errors in a controlled process, less reliance on the one person who knows how something works. Hard to price, easy to defend after an incident.
Make the sponsor choose one as primary. If they choose capacity, the follow-up question is which backed-up work absorbs it, and that is a conversation with an operations leader, not a vendor. Capability that nobody redirects shows up in a survey and never in a result, which is the whole argument in why buying licenses is not the same as building capability.
The gap between felt gains and booked gains is now well documented. McKinsey’s State of AI survey found around 80% of respondents saying AI had improved their own productivity while 6% credit it with 5% or more of company EBIT, and Gartner reported in October 2026 that 44% of CEOs struggle to demonstrate AI’s financial returns. The mechanism question is how you avoid joining them.
Where the vendor’s percentage came from
Three moves break open almost any productivity claim, and they take about ninety seconds each.
Ask for the population. LSE research with Protiviti found trained employees reporting 11 hours a week saved against 5 hours for those who had not been trained. Directionally useful, self-reported, and measured on people who opted into a study. National panel data lands lower: generative AI accounted for about 5.4% of total work hours, roughly 2.2 hours a week on a 40-hour schedule. Both are true. They describe different groups.
Ask for the window. A Basware study of 400 CFOs reported $1.36 returned per dollar invested, which reads as a strong result until you notice it covers three years and finance technology rather than workforce training. A return quoted without its window is a return you cannot compare to anything.
Ask what the number would look like if it failed. Any claim that cannot come out low was not measured. When a vendor cites a 30% improvement with no spread, no sample size, and no failure case, you are looking at marketing copy with a decimal point.
Then do the arithmetic on your own numbers, small and conservative. Take that $19,700 loaded cost. If 10 of the 40 people recover one hour a week, that is 520 hours a year, about $24,000 at the BLS rate. Payback inside a year on a tenth of the group getting a modest result. Keep every input visible so each one can be argued with, which is the point of a finance-ready model your team cannot dismiss, and read the output as a hypothesis, not a promise. The payback windows worth expecting from AI enablement sit well inside the two-to-four-year band Deloitte found for enterprise AI work, because the ticket is small and the signal arrives fast.
Who owns the number in six months
Ask for a name. Not a function, not a committee, a person whose objectives contain the metric.
In a company of 50 to 500 people there is no Chief AI Officer, so this has landed on someone as a second job: a COO, a chief of staff, a head of operations. Fine. Put the metric in their goals and give them two hours a month. An unowned metric decays faster than the skill does, and training research has long put the share of training delivered that is never applied on the job at around 45%. Missing follow-through is most of that gap.
One test we use with buyers: ask who presents the result, to whom, on what date. If the answer is the vendor, you bought a session. If it is your own sponsor presenting to you in March, you bought a program.
Three conditions worth more than a better business case
The instinct when a proposal feels thin is to send it back for more forecasting. More forecast on an unmeasured baseline produces a more confident wrong number. Approve instead, and attach three conditions.
- Baseline before kickoff. Two named deliverables, timed, in writing, before the first session. No baseline, no start date.
- A named owner with the metric in their objectives. Sponsor, not vendor. Two hours a month, written down.
- A dated review with pre-agreed thresholds. Pick the date now, name the data source, and write the pass and fail numbers before anyone is invested in the answer. The rules for evaluating a pilot before you scale it cover how to set thresholds that survive contact with a mixed result.
Those three cost nothing and turn a hopeful spend into a testable one. Set the thresholds against what adoption looks like at six and twelve months instead of a vendor’s target, and by month six you will know whether you are funding a habit or a memory.
Two questions that waste the meeting
“What is the industry benchmark ROI for AI training?” There is no credible one. Published figures come from groups unlike yours, measured on self-report, over windows chosen by whoever quotes them. Asking invites a number nobody can defend and pulls the meeting away from your own two deliverables, the only data that will settle this.
“Will people get a certificate?” The easiest request in the world for a vendor to grant, and it tells you nothing about whether the work changed. A certificate of attendance has never made a board pack quicker to produce.
Ask instead what the provider has seen fail at your size, and listen for whether the examples are specific.
Frequently asked questions
What should a CFO ask first about an AI enablement program?
Which recurring piece of work this is meant to change, and who timed it before the program starts. Everything else rests on those two answers, and a proposal without them is asking you to fund a hope.
How much should we expect to pay per employee?
Market pricing for hands-on workshops runs about $2,000 to $15,000 per engagement, which for a group of 30 to 50 lands anywhere from $40 to $500 a head depending on scope and session count. Set that against the $782 a year midsize companies already spend per learner, then double it for your people’s hours.
What payback period is reasonable for AI enablement?
Six to twelve months on a conservative model, far shorter than the two to four years Deloitte found for enterprise AI work generally. The ticket is small, the cost is mostly hours, and the signal arrives within a quarter. If a vendor needs three years to pay back a workshop, the model is wrong.
How do I check a vendor’s productivity claim?
Ask for the population, the window, and what the figure would look like in a failure case. Self-reported time savings run high: a controlled study of experienced developers found them 19% slower with AI access while believing they had been 20% faster.
Our team already has licenses. Does that change the questions?
It sharpens them. You have already paid for access, so the only question left is whether anyone’s work changed, and your admin console holds half the baseline you need.
Should finance own the measurement, or the sponsor?
The sponsor owns it, finance sets the standard for what counts. Finance owning the collection turns a capability question into an audit, and the sponsor stops caring about a number someone else reports.
What should I attach to the approval if I say yes?
A baseline taken before kickoff, a named owner with the metric in their objectives, and a dated review with written pass and fail thresholds. No extra forecasting needed, and you can defend the decision whichever way it goes.
When should I say no?
When nobody will name the deliverable that has to change, when the sponsor cannot say which of cost, capacity, revenue, or risk the return runs through, or when the vendor will not agree to a baseline. Each one means the program has no way to be judged.
Is a pilot always the safer answer?
Not when the pilot is a smaller version of the same unmeasured thing. No baseline and no thresholds buys a delay, not information. A full program with the three conditions attached beats a vague pilot, and our piece on diagnosing an enablement push that did not work shows how the two failure modes diverge later.
Key takeaways
- The invoice is roughly half the cost. Price participant and manager hours at employer cost, about $46.89 an hour, before comparing any quote to any benefit.
- No baseline, no business case. A rough measurement of two deliverables taken before the sessions beats a precise one reconstructed afterwards.
- Make the sponsor name one mechanism: cost out, capacity in, revenue, or risk. “Productivity” means nobody has decided who acts next.
- Interrogate every vendor percentage on population, window, and failure case. Self-reported gains run well above measured ones.
- Approve with three conditions rather than demanding a stronger forecast. Baseline, named owner, dated review with written thresholds.
- Expect payback inside a year on conservative inputs. If the model needs three years for a workshop, the model is wrong.
If a proposal is already sitting on your desk, book a free AI readiness call. We will tell you which parts of it are measurable and which parts are decoration.
Arthur Wandzel
