Train the team or add a headcount? The 2026 arithmetic

The figure everyone quotes here is $14,170 to hire an IT professional against $5,770 to upskill one, from Pluralsight’s 2026 Tech Skills Report. It is a real number, and it loses budget reviews, because it prices two purchases while hiding what each obliges you to. One is a fee you pay once. The other is a six-figure annual commitment with a severance tail.

Put both in the same unit and the comparison sharpens. A headcount buys roughly 1,850 hours of capacity a year for about $100,000. A one-time AI enablement program for 25 people costs about $30,000 all-in and returns those same 1,850 hours if each attendee gets back 4% of their working time and keeps it. Below roughly 1.2%, the req is the better buy. Which side of that line your team lands on is not something a vendor can tell you.

The comparison that loses budget reviews

Pluralsight’s other finding is that 89% of organizations now say hiring is more expensive than upskilling for IT roles, up from 49% a year earlier. That swing in sentiment is the most interesting thing in the report, and it still does not survive a CFO’s first question. Framed as AI training vs headcount, this looks like a price comparison. The two costs have entirely different shapes.

	Open req	Enablement program
Cost shape	Recurring, indexed to salary bands	One-time fee, plus hours
Year-one extras	$5,475 cost per hire (SHRM)	None
Time to capability	39 days to fill, then a ramp	Weeks
Reversibility	Severance, morale, a rehire	You stop buying sessions
What decays	Nothing, while they stay	The skill, without reinforcement
Capacity delivered	Concentrated in one calendar	Thin slices across many

The last two rows are where the decision lives, and the two every published comparison skips. A hire is expensive and durable; a program is cheap and perishable. Those are not the same bet. Pretending they are is how training lines get cut while license spend grows, the pattern we set out in the case against treating a license rollout as a capability.

What a headcount actually costs per hour

The Bureau of Labor Statistics publishes the only loaded rate that holds up in a budget review. As of June 2026, employer compensation costs for full-time private industry workers averaged $54.00 per hour worked, split $36.97 in wages and $17.03 in benefits.

“Per hour worked” is the part that matters: it excludes holidays and paid leave from the denominator, so you multiply by hours actually worked, not by 2,080. A full-time employee logs near 1,850 hours after holidays and PTO, putting the average full-time headcount at roughly $100,000 a year in employer cost. Technical and professional roles run well above that; use your own band.

Then add the one-time charges. SHRM puts average cost per hire at $5,475 for nonexecutive roles with a 39-day median time to fill, and nobody is at full output on day 40. Call year one $105,000 for nearer 1,400 effective hours. That is the yardstick: $54 an hour, standing, until the person leaves.

What enabling the team you have costs per hour

Build the program cost the same way, by charging it for the hours it consumes. Almost nobody does, and that omission is what gets training models thrown out.

A posture that actually changes how people work is one facilitated day on the team’s own documents, plus two short follow-ups where people bring real work. That is 12 hours of attendee time, or $648 a head in payroll at $54.00 before anyone invoices you. Fees sit on top. A private cohort session falls in a $2,000 to $15,000 market range depending on length, customization, and who delivers it. For a 25-person room, $10,000 for the main day plus $4,000 for follow-ups is a fair midpoint: $560 a head.

All-in, $1,208 a head, or $30,200 for 25 people, payroll being the larger half. Discounting the fee by 20% saves $2,800; cutting two hours of session time saves $2,700 and may cost you the outcome. Any per-head figure quoted without the hours understates by roughly half, the flaw in most of the benchmarks companies quote for enablement spend.

The breakeven, by cohort size

Both sides sit in the same unit now, so you can solve for the only input in question: how much time each person gets back. Two thresholds matter. The higher one is where the program returns as many hours as a headcount; the lower one is where it simply buys an hour of capacity for less than $54.

Attendees	Cohorts	All-in per head	Program total	Gain to match a headcount	Gain to beat $54/hr
15	1	$1,581	$23,720	6.7%	1.6%
25	1	$1,208	$30,200	4.0%	1.2%
50	2	$1,128	$56,400	2.0%	1.1%
100	4	$1,088	$108,800	1.0%	1.1%

Read the right-hand column first. It barely moves. Past the first cohort the fee stops mattering, because attendee payroll dominates and payroll is linear in headcount. Negotiating the invoice is not the lever; the hours you ask people to sit still are.

At 25 attendees, the 4% threshold is 74 hours a year, about 1.5 hours a week each. If that time is real and collectible it equals one new employee, bought at $16.32 an hour against the hire’s $54. The lower threshold is under half an hour a week, a low bar, and why the arithmetic favors enablement in most mid-market cases. The arguments worth having are about its two assumptions, not its math.

The problem with recovered minutes

Here is the assumption nobody writes down: that time returned in slices adds up to a person. It does not, automatically. If 25 people each save 20 minutes a day you have not hired anyone, you have 25 people whose days are slightly less congested. Those minutes become capacity only if queued work absorbs them, and that work is rarely spread evenly across whoever freed it.

Three tests tell you whether yours will aggregate:

1. Is there a visible backlog in the same work? Proposals not going out, analyses not getting done, a support queue with age on it. Returned minutes flow into a backlog and evaporate into an empty one.
2. Is the work divisible? Twelve minutes suits a draft, a summary, or a first-pass review, and is useless for anything needing a two-hour block.
3. Does someone own the reallocation? Without a manager who reassigns freed time, it goes to whatever was already waiting, Slack included.

Fail all three and your constraint is ownership, not throughput, and ownership is what a hire buys. It is also why programs that return real hours sometimes show no output change at all, the timing problem we work through in what payback period to expect from enablement.

What the measured evidence says about the gain rate

Ask a vendor what gain to assume and you will usually get a number between 20% and 40%. The research answer is that it depends on the work, and is sometimes negative.

The strongest positive field evidence is Brynjolfsson, Li and Raymond’s study of 5,172 customer support agents, published in the Quarterly Journal of Economics in 2025. An AI assistant raised issues resolved per hour by 15% on average, concentrated among newer and weaker performers. The most experienced agents saw little or none.

The counterweight is a 2025 randomized trial by METR, which gave 16 experienced open-source developers 246 real tasks in codebases they knew well, randomizing whether AI tools were allowed. They were 19% slower with the tools, and estimated afterward that they had been 20% faster. Sit with that gap: a satisfaction survey would have recorded a win the stopwatch says never happened.

Together they give you a design rule. Gains appear where the work is repetitive, text-heavy, and done by people who are not already the best in the building at it. They go missing, sometimes backward, in deep expert work on systems the expert already holds in their head. So a survey cannot settle your case. Defending a program against a req takes a task-level measurement before the session and the same one after, in hours rather than satisfaction. The AI enablement metrics that survive that scrutiny are in our measurement playbook.

In the sessions we run, the first hour usually goes to discovering that much of the room has spent a year using their assistant as a slightly worse search engine. The gain available to a team in that state is not 4% because the tool is powerful. It is 4% because the baseline is that low.

Four cases where the hire is the right answer

A program sold into any of these usually fails, for reasons unrelated to the facilitator.

1. The gap is a capability adjacent to nothing anyone does. You can teach a finance team to work with an assistant. You cannot teach it to build and run a data pipeline in 12 hours. Depth of gap beats cost per head.
2. The capacity must be concentrated and schedulable. If the need is one person owning a build, on call and accountable on a date, 25 thin slices do not substitute.
3. The constraint is ownership. Nobody owns AI adoption and it keeps getting dropped. That is a role, not a workshop. In most companies this size it has landed on someone as a second job, and the honest fix is making it their first.
4. The team is at its hours ceiling. If everyone works 50-hour weeks, returning 4% means they go home earlier. Real value, wrong budget line.

Three cases where the program wins outright

Invert all of that and the program is the stronger buy.

1. The same task is spread across many people. Drafting, summarizing, first-pass review, routine client correspondence. Many hands doing variants of one text-heavy job is the shape the measured gains appear in.
2. Headcount is frozen and operating budget is not. A $30,000 program and a $100,000 req get approved by different people on different timetables, and only one can happen this quarter. The sequencing options are in our guide to funding enablement in a flat-headcount year.
3. You need the answer inside a quarter. 39 days to fill plus a ramp puts a hire at useful output near the end of the next one. A cohort runs in three weeks.

Most companies land on both, in order: one hire who owns the capability, plus a program that lets everyone else use what that person builds. One owner per 50 to 150 people beats two owners with no program.

Running the number for your own company

Five inputs, 20 minutes in a spreadsheet.

1. Your loaded hourly rate. Employer cost for the group divided by hours actually worked. Use $54.00 only if you have nothing better, never salary over 2,080, which understates by roughly 30%.
2. Attendees and hours. People in the room times hours of session and follow-up. Your payroll line, about half the program.
3. Delivery fees. Per cohort, not per head, and ask what changes if the room holds 30 instead of 20. Converting quotes in different units is its own exercise, covered in cost per head versus cost per hour saved.
4. A baseline on one task. Cycle time on something attendees do weekly, taken before the session. Without it you have an opinion.
5. A collectibility test. Name the backlog the freed hours flow into. If you cannot, model the program as a quality and retention investment and say so in the review.

Then run both thresholds. If the gain needed to match a headcount is under 4% and your baseline is as weak as most 2026 baselines, buy the program first and let the req wait a quarter for evidence. If you cannot name the backlog, or the gap is a skill nobody has, hire. For the full model, see how to build an ROI model your finance team cannot dismiss.

One timing note, since training only pays while people stay. BLS put median employee tenure at 4.1 years in January 2026, up from 3.9 in 2024. Hiring carries that exposure on worse terms: when a trained person leaves you are out $1,208, not $100,000 and a 39-day hole.

Frequently asked questions

Is it cheaper to train my team or hire someone?

Per hour of capacity, training is cheaper in nearly every mid-market case: a 25-person program at roughly $1,200 a head beats a $54-an-hour headcount at any sustained gain above about 1.2% of working time. The question underneath the price is whether your constraint is throughput or ownership, and training only addresses the first.

How much time does each person need to get back for the training to pay for itself?

Under half an hour a week to buy capacity more cheaply than hiring, and about 1.5 hours a week each, or 4% of working time, to replace a whole headcount’s annual hours from a 25-person cohort.

What productivity gain should I assume in the model?

Do not assume one. The published range runs from +15% among 5,172 customer support agents to −19% among experienced developers in a randomized trial. Measure one task’s cycle time before the session and again at four and twelve weeks.

My headcount is frozen but I have operating budget. Does that settle it?

It settles the timing, not the decision. A program can be approved and delivered this quarter; a req may not open at all. Spend on enablement, measure it properly, and use the measurement as the case for the req you still want next year.

Why does the per-head cost barely improve when I train more people?

Attendee payroll is over half the cost and scales linearly with headcount. Fees spread across a bigger room; the hours do not. Past the first cohort, per-head flattens near $1,100.

Does the training justify itself on retention alone?

Sometimes, and it is a weaker argument than it sounds. If the team is at its hours ceiling, returned time becomes shorter days rather than output. Worth having, but put it on the right line.

How do I prove afterward which choice was right?

By measuring something before you chose: task cycle time, a sample of output quality, and how often the tool gets used for the work it was taught for. License logins do not count, since measuring AI adoption through seat activity flatters a program that has not worked.

What about the hours a new hire spends ramping up?

Budget for them. A 39-day median time to fill plus a ramp means a hire delivers nearer 1,400 effective hours in year one against a full year of cost, not 1,850.

Key takeaways

• The $14,170 versus $5,770 comparison prices two purchases, not two commitments. A headcount is a standing cost near $100,000 a year at the BLS full-time rate of $54.00 per hour worked; a program is one-time and decays.
• Charge the program for attendee hours. Twelve hours at $54.00 is $648 a head, over half the all-in cost, and omitting it is why training models get thrown out.
• A 25-person program at $1,208 a head matches a headcount’s 1,850 hours at a 4% sustained gain per person, and beats $54 an hour above roughly 1.2%. Per-head cost flattens near $1,100 past the first cohort, so session length beats fee negotiation as a lever.
• Returned minutes are capacity only if a backlog absorbs them. If you cannot name one, your constraint is ownership and the hire is the right call.
• The gain rate is empirical: +15% in the support trial, −19% among experienced developers who believed the opposite. Take a baseline first.

Holding an open req and a training line and forced to pick one this quarter? Book a free AI readiness call. We will run both thresholds against your loaded rate, your headcount, and the task you would measure, and say plainly if the req is the better buy.