A quote that reads $420 a head tells you what a vendor charges. It tells you almost nothing about what you are buying, because heads are the wrong denominator. The one that answers the question your CFO will ask is hours of work returned.
Convert each quote to a cost per hour saved before you rank them. On three real quotes for the same 120 people, the ranking usually flips: the lowest per-head price lands last, and the quote priced 33% higher per head wins by a wide margin.
Why cost per head survives on every quote
Every AI enablement proposal leads with a per-head price for a plain reason: both sides can compute it on the day you sign. Headcount is a known number. Hours saved will not be known for months, and only then if somebody bothered to measure.
So the market settled on the figure that is available rather than the figure that decides anything.
The published bands show how little that figure sorts options. Pertama Partners put per-person program cost at $2,000 to $3,500 for a 50-person team, $1,200 to $2,000 at 500 people, and $800 to $1,200 at 5,000, in a November 2025 budget guide. Training Magazine’s 2025 Training Industry Report puts all corporate training at $874 per learner, rising to $1,091 at small companies and falling to $468 at large ones.
Those bands are honest. They are also no help in a bake-off, because every quote you receive will sit inside them. Three vendors can land within $60 a head of each other and sell three products that share nothing but a room.
What a per-head number does tell you
It tells you which business the vendor is in.
A pricing benchmark out this year by Daria Danilina at AI Enablement Insider looked at 26 firms that sell training. Of the six training specialists, four charge £1,000 or more per learner. Of the full-stack firms that also sell build work, none do, and two sit at £100 to £300 a head.
She notes the sample skews UK and sub-50 staff, and that the per-learner cut covers 11 firms, so treat it as a pattern and not as a market figure.
The pattern is still worth knowing. A low per-head figure often means the session is a door-opener for an automation project. A high one usually means teaching is the product.
Neither model is wrong, and the two produce different Mondays.
Format moves the number just as hard. Coursera for Business publishes $399 per license per year for 2 to 499 licenses. A half-day or full-day live session for 10 to 25 people runs $3,000 to $12,000, per Pertama’s same guide.
Divide the second by the people in the room and you can land within $50 a head of the first while buying two different things. Which of those two structures favors your team size is a separate question, and the per-head versus flat-fee comparison works through it.
The measured size of the hours-saved denominator
Before you divide by hours saved, look at what the measured number is. Not the vendor claim. The nationally representative one.
Bick, Blandin and Deming surveyed US workers for The Rapid Adoption of Generative AI, released as NBER working paper 32966 and published in Management Science in 2026. They found between 1 and 5 percent of all work hours assisted by generative AI, and reported time savings equal to 1.4 percent of total work hours.
The authors’ ongoing tracker put the August 2026 reading at 2.1 percent of total work hours saved. In that wave, 40.5 percent of employed respondents used the tools for work, and 5.3 percent of all work hours were spent in them.
Among people who use the tools, the figure is larger. Writing for the St. Louis Fed in February 2025, the same authors recorded an average saving of 5.4 percent of work hours for users, which is 2.2 hours a week on a 40-hour week.
Now the part that should change how you read a quote. Of daily users, 33.5 percent said the tools saved them four hours or more that week. Of people who used them a single day, 11.5 percent said the same.
Read that spread twice. The return is concentrated in people who reach for these tools most days, so your quote comparison is a bet on how many staff a program converts into daily users. That is a question about design and follow-through, not about price.
The five inputs, and where each one comes from
Five numbers turn any quote into a cost per hour saved. Three of them you own.
1. Total cost. Vendor invoice plus your own payroll. The Bureau of Labor Statistics put total compensation at $78.85 an hour for management, professional and related occupations in June 2026, counting benefits, with management, business and financial roles at $89.14 and professional roles at $74.42. Use the band that matches who sits in the room. No vendor prices this line and it is often the larger half.
2. People reached. Not people invited. People who finish the thing you are paying for.
3. Keep rate. The share of those people still using the tools weekly at the end of your horizon. Measuring adoption at a single point is cruder than it sounds. It is still the input that decides the comparison, and the one nobody will guarantee.
4. Hours saved per week, per person who keeps going. Start at the measured 2.2 and adjust down for shallow formats. Mark that adjustment as your assumption, out loud, because it is not a finding.
5. Horizon. Fifty weeks for a one-year view. Anything longer is a forecast about a product category that reshapes itself twice a year.
Then:
cost per hour saved = total cost / (reached x keep rate x hours per week x weeks)Three quotes for 120 people, one denominator
Three proposals of different shapes for the same 120-person company, priced at market rates, over a 50-week horizon. Internal hours are costed at $78.85.
| A. Self-serve licenses | B. Six live one-day cohorts | C. Cohorts plus 8 weeks of follow-through | |
|---|---|---|---|
| Vendor invoice | 120 x $399 = $47,880 | 6 x $7,500 = $45,000 | $60,000 |
| Cost per head | $399 | $375 | $500 |
| Your hours | 144 (only the engaged) | 880 (7 each + coordination) | 1,260 (7 each + 3 of follow-up + coordination) |
| Your payroll cost | $11,354 | $69,388 | $99,351 |
| Total cost | $59,234 | $114,388 | $159,351 |
| Weekly users at week 50 | 18 (15%) | 42 (35%) | 66 (55%) |
| Hours saved per user, per week | 1.5 | 2.2 | 2.2 |
| Hours returned over 50 weeks | 1,350 | 4,620 | 7,260 |
| Cost per hour saved | $43.88 | $24.76 | $21.95 |
On price per head, B wins and C looks 33% overpriced. On cost per hour saved, C wins, B is close behind, and the license-only option costs twice as much as either. The ranking inverts because the denominator changed, and nothing else did.
Two things carry that result. Self-serve licenses reach the people who were already going to be fine, which is why the keep rate sits low and the hours per user sit lower. And payroll swamps the invoice in any format that puts people in a room, which is the line every quote leaves out.
Resist one temptation here. At $21.95 an hour against a loaded rate of $78.85, column C looks like it returns $3.59 for every dollar. It does not, not yet.
An hour saved becomes money only when it gets spent on work somebody wanted done.
The number that moves the answer is not the price
Run the sensitivity check before you talk price.
Hold C’s price and drop its keep rate from 55% to 30%. Hours returned fall to 3,960, and the cost per hour saved climbs to $40.24, behind B and barely ahead of the licenses. Now lift A’s keep rate to 35% while adding the extra hours those people spend learning: 3,150 hours returned at a total cost of $74,374, or $23.61 an hour, and the self-serve option beats the live cohorts.
Price moved nothing in either case. The keep rate moved everything, in both directions.
That is the reason the useful question to a vendor is not whether they can sharpen the number. It is what in their design produces weekly use at week 40, and what they will measure to show it. A paid seat is not a capability, and a per-head discount buys you a cheaper version of the same bet.
What a vendor can honestly commit to
Three inputs, three different owners.
Reach is theirs. Attendance, completion, and whether the exercises used your documents or a demo dataset. A vendor should commit to all three in writing, with a remedy if they miss.
The hours baseline is yours. It has to exist before the vendor starts, or there is nothing to compare against afterward. The measurement playbook has a one-week protocol for capturing it, and that week is the cheapest line in this whole exercise.
The keep rate is shared, and nobody should guarantee it alone. It depends on their design and on whether your managers make room for the new way of working. We quote these programs, and reach is the only one of the three we would put in writing before a baseline exists. A vendor who guarantees a keep rate without one is guessing with your money.
Ask each bidder to fill in columns like the ones above with their own numbers. The gap between their keep-rate guess and yours is where the deal gets made.
A vendor who has run enough rooms will put a number down and tell you what it depends on. One who has not will talk about engagement.
Where cost per hour saved stops being the right test
Hours saved is a floor on the argument, not the ceiling, and the evidence on that is awkward for anyone selling this math.
The Wharton and GBK Collective study The Adaptive Advantage came out on October 13, 2026, from a survey of 1,416 US enterprise leaders and employees. It found 80% reporting positive returns from AI, and only 28% calling those returns strongly positive.
The split that matters sits under that. Among teams tracking at least one growth measure, 35% reported strongly positive returns, against 13% of those watching efficiency alone.
So the AI enablement metrics worth keeping past month three are the ones tied to output, not to time returned.
Counting hours alone tracks with the weaker outcome. Use cost per hour saved to rank quotes, because it is the only common denominator three unlike proposals share.
Do not use it to declare the program a success twelve months later. By then you should be looking at what the team shipped that it could not ship before, which is the harder version of the same question and the one worth building into a model your finance team cannot dismiss.
The same study found 78% of leaders believe employees get adequate AI training while 56% of employees agree. Both numbers come from the same organizations. Whatever the per-head figure on your quote, somebody is already wrong about what it bought.
Frequently asked questions
What should a per-head AI training number include?
Ask what the number covers before you compare it to anything. A usable per-head figure covers design, delivery, handouts, and any follow-up session. It usually excludes travel, platform licenses, and every hour your own staff spend, which at BLS loaded rates runs $74 to $89 an hour for the people most likely to attend. In the three-quote comparison above, payroll was the larger half of total cost in both live formats.
How do I compare a software license to a live workshop?
Put both through the same five inputs. A $399 annual license and a $7,500 session can look alike per head and behave nothing alike, because the license reaches everyone shallowly while the session reaches a cohort deeply. The gap shows up in the keep rate and in hours saved per person, not in the price.
What hours-saved number should I use if we have no baseline?
Use 2.2 hours a week per active user, which is the measured average for people who use these tools at all, then discount it for formats where usage stays shallow. Flag the discount as an assumption in your model. Then go get a real baseline, because a comparison built on two guesses only ranks the guesses.
Is a lower cost per head ever the right choice?
Often, when the audience is already fluent and needs reference material rather than teaching. Self-serve licenses also win on cost per hour saved whenever the keep rate holds up, as the 35% run above shows. The mistake is assuming a low per-head price produces that keep rate on its own.
How do I get a keep rate out of a vendor who will not commit to one?
Ask for the number from their last three jobs of similar size, measured at 90 days, and ask who measured it. A vendor with delivery history has the figure or can say honestly why they do not. Then write a checkpoint into the contract rather than a guarantee: the baseline, the 90-day read, and what happens if the read lands under the number they quoted.
Should our own people’s time count as a cost?
Count it for anything you plan to repeat. A one-off session can be argued either way, but across a rollout the hours are a real claim on the same budget, and leaving them out is what makes a cheap license look free. Count the hours of people who engage, not everyone on the invite list.
What counts as a good cost per hour saved?
No benchmark is published, which is part of why this comparison is useful rather than final. The honest test is relative: compare each quote against the others and against your loaded hourly rate. Anything above your loaded rate is returning less capacity than it consumes, and should lose.
Does this math work for a 30-person team?
The arithmetic works at any size, but the sampling does not. At 30 people a keep rate is six or seven people, so one enthusiast or one skeptic swings the answer. Below about 50 staff, rank quotes on reach and on the baseline you can measure, and treat the hourly figure as a sanity check.
How long a horizon should I use?
Fifty weeks. A shorter window flatters the cheap option by hiding the decay, and a longer one asks you to forecast tools that change every quarter. If you want a second view on when the money comes back, the payback-period question deserves its own answer.
Where to start
Email every bidder the same five-row sheet: total cost, people reached, keep rate at 50 weeks, hours saved per week per user, and the resulting cost per hour saved. Tell them to use their own assumptions and to mark which ones they will stand behind. You will learn more from who fills it in than from the numbers they write.
Then measure your own baseline week before anyone starts. It costs a few hours and it is the only thing that converts a quote comparison into evidence.
If you want a second set of eyes on three quotes you are holding, a free AI-readiness call is a direct way to get one. We run these sessions, so we will also tell you when the cheapest quote on your desk is the right one.
Arthur Wandzel
