Back to Guides
Enterprise Software 15 min read

The Real Cost of unmeasured AI training

The Real Cost of unmeasured AI training

The cost of an unmeasured AI enablement program is not the invoice. It is a short list of decisions you can no longer make: whether to renew the vendor, whether to run the same session for the next department, whether the people who went quiet in month four need a refresher or a different role. Those decisions get made anyway, on instinct, and the one input that would have settled them stops being available the moment the first session starts.

That input is a baseline. Everything else here can be rebuilt later from invoices, calendars and admin consoles. A record of how long your team took to produce a named deliverable before anybody trained them cannot. Skip it and you have bought a program whose result is permanently unknowable, which is a different purchase from the one you approved.

Three different things get called AI training ROI

Search the phrase and the results mix three unrelated questions. The first is the return on putting AI into a business at all. The second is what it costs to train a model, a capital question about compute and hardware that belongs to a lab, not to you. The third is the return on teaching your own staff to use the tools you already pay for.

This piece is about the third, and the distinction has an edge: the gloomy headline figures about AI returns come from the first question and get quoted at budget meetings as though they settle the third. A failed deployment and an untrained team are different diseases with different cures.

The gap that makes this a live problem

Adoption is no longer the open question. Stanford’s Human-Centered AI Institute put organizational AI use at 88% of surveyed organizations in 2025, with generative AI live in at least one business function at 70%. The licenses are bought. Evidence of a financial result has not followed.

McKinsey’s 2026 State of AI survey, fielded May 4 to June 8 2026 across 1,719 respondents, found that 80% report improvements in their own productivity from AI while 37% attribute any EBIT impact to AI use, flat against the prior year. About 6% qualify as high performers, attributing at least 5% of EBIT to AI. The gap between 80 and 37 is the whole commercial problem, and it is what our case against treating license count as capability sets out at the level of strategy.

One survey puts a figure on the training half of that gap. A DataCamp study fielded by YouGov among more than 500 US and UK business leaders, published in February 2026, found just over one in five leaders reporting a strong positive return on AI investment, against 42% among organizations with mature data and AI upskilling underway. Read it as correlation, not cause: a company disciplined enough to run an upskilling program is usually disciplined enough to track outcomes. Either way it points one direction. The organizations that can show a return are the ones that set up to see one.

What the missing numbers prevent

Skipping measurement does not cost you insight in the abstract. Each missing input disables one decision, and each of those decisions arrives on a date.

Input you did not take Decision it would have settled When the cost lands
Timed baseline on one named deliverable Whether capability moved at all Month 3, when someone senior asks
Rework rate on that same deliverable Whether quality held while speed rose Month 2, at the first client-facing error
Usage counted by workflow, not by seat Who needs a refresher and who needs a different role Month 4
A group trained later than the rest Whether your session or a vendor product update caused the change Month 6, at renewal
Fully loaded cost per participant Whether the next department is cheaper or dearer to train Next budget cycle
A dated record of what was taught Whether a visible gap is a training gap or a scope gap Month 9

Read the right-hand column as a schedule of bills. Month six is the expensive one: a renewal talk with no comparison group is a negotiation where you hold nothing the vendor does not also hold. Month nine is the one that quietly ends programs. A gap appears, nobody can say whether the session covered it, and the fastest reading in the room is that the training failed.

Why the baseline is the only input you cannot buy back

Four of the six rows above can be rebuilt months later. Invoices sit in accounts payable. Calendars show who attended. Admin consoles hold usage history for a known window. Managers can be asked what changed, late and imperfectly.

The timed baseline cannot, for two reasons that compound. The first is that the condition has already changed: once people have been trained, there is no untrained version of them to time. The second is memory. Ask a team in March how long a task took them in January and you get a figure shaped by everything since, including their stake in believing the training helped. Before-and-after comparisons collected after the fact drift toward what the respondent wants, and they drift most in the teams that enjoyed the session.

So the window for this work is the week before delivery. Our guide to what to capture in that week treats it as a gate rather than a nicety, and from what we see running these sessions, the measurement conversation either happens before the room is booked or it does not happen. Nobody sets up a baseline in the glow of a good workshop, and no catalog page has a commercial reason to flag the window, because the sale closes either way.

The second program you buy because you cannot tell whether the first one worked

Work it through on a 60-person team. Market rates for a hands-on fluency workshop run roughly $2,000 to $15,000 by scope and session count. Call it $9,000 for a two-session engagement.

Now price the room. The Bureau of Labor Statistics put employer costs for civilian workers at $49.46 per hour worked in its June 2026 reference period, released September 9 2026. Sixty people out of their jobs for four hours is 240 hours, or $11,870 of loaded labor cost. The time in the room costs more than the fee and appears on no invoice, which is why it rarely appears in the business case either.

All in: about $20,900, and an unmeasured program files that whole figure under a word your finance team already uses, which is unverifiable. The predictable result is not a write-off but a repeat: a year later, with adoption still uneven, somebody proposes running it again, because nobody can show where the first attempt landed. The second $20,900 buys the same unknown. Two swings at the same pitch, no record of either, and a flat-headcount year in which L&D defends its line against functions that can produce one.

Roughly twice the cost on the purchase order, then, for the same information you had before.

What measurement costs, in dollars

Most pages on this topic imply measurement is free. It is not, and the honest figure makes the better argument. Two inputs do most of the work: a timed task before training, and the same task at day 30. Budget 20 minutes per person for each. That is 40 minutes, about $33 of loaded time per participant, or $1,980 across 60 people. Add 12 hours of one analyst’s time to design the task, collect results and write them up, call it $594. Total about $2,575.

Against the $20,900 all-in figure above, measurement runs about 12%, and it is the only part of the spend that yields a number at the end. A CFO who declines a 12% surcharge for a verified result is not being prudent, they are buying the more expensive version of the same thing.

One refinement is worth the extra cost. Running the same timed task on a second team trained two months later gives you a comparison group, which is the cheapest credibility available and the only practical answer to whether your session or the tool’s monthly update produced the change. That question is harder than it looks, and separating the training effect from the tool effect is where most internal reporting quietly gives up.

Why the people selling the training measure satisfaction

Nearly every page on the first page of results for this question is published by a company that sells training or training software. None of them is lying. They report what their systems can see, and a learning platform sees enrollments, completions, time in module, and a satisfaction score collected while people are still in the room.

Those are real figures about a real thing, and the thing is not capability. A completion rate says a video finished playing. A satisfaction score rates the session, catering and facilitator warmth included, collected at the one moment when the respondent knows least about whether anything will stick. Kirkpatrick’s four levels of evaluation have had names for this since the 1950s: reaction and learning at the front, behavior and results at the back. The back two are what a buyer needs and what no platform can report.

Seat-license activation sits in the same category. It tells you a login happened, not whether the person drafts every client memo with an assistant or opened the tool once in onboarding. Counting by workflow instead of by seat changes how you query the data, not what you spend, and it changes who shows up as trained.

The metrics most favorable to the seller are the ones least useful to the buyer, and the overlap is no coincidence. Decide what counts as evidence before a vendor hands you a dashboard. Our measurement guide for AI enablement sets out the full instrument.

Four numbers worth collecting

A workable set fits on one page.

  1. One timed task, twice. Pick a real deliverable the team produces weekly: a client update, a variance commentary, a candidate screen, a spec. Time five people before training and the same five at day 30. Hours returned is what finance asks about first, and reading a time-savings claim means knowing where those hours went as well as that they were saved.
  2. A rework count on that deliverable. How many drafts came back, and for what. A program that halves drafting time and doubles corrections has helped nobody.
  3. Workflow-level usage. How many people used an assistant on that named task last week. Not logins, not seats, not any count your license console offers by default.
  4. One sentence per manager, monthly. What their team does differently now. Unweighted, and the earliest warning you will get that the change is fading. These four lines are what a board update should hold, and rethinking the quarterly AI report covers the shape that survives a hostile question.

Four inputs, one owner, a few hours a month. Most programs have none of them, and the reason is not cost. The week to set them up passed without anyone noticing it was the week.

If the training already happened

The baseline is gone and no effort recovers it. What remains stops the loss rather than reversing it.

Time the task now and treat today as the baseline for everything after. You give up the claim about the past program and keep the ability to say something about the next six months. Ask the vendor for a dated record of what was covered, so a future gap sorts into a training gap or a scope gap. Then train the next department with a real baseline a week ahead, using the already-trained group as the comparison. That one decision turns the unmeasured program from a sunk loss into a control group.

Say once, plainly, that the first program’s result is unknown. A function that reports an unknown keeps its credibility for the figure it reports next. One that reports a smile sheet as a result spends that credibility and does not get it back.

Frequently asked questions

What is a realistic return on AI enablement spend?

No defensible benchmark exists, and the per-dollar figures circulating most widely trace back to supplier-funded research with no published method. Build the number from your own hours and your own loaded rate. That model survives a finance review; no industry average does.

How long before AI enablement shows a return?

Behavior change on a named task is visible at 30 days if a baseline exists. Effects that reach a profit-and-loss line take two to four quarters, because the hours returned have to be redirected into something that bills or ships before they count. Judge a program on P&L impact at 60 days and it fails whatever it achieved.

Can we measure this without buying an analytics tool?

A spreadsheet, a stopwatch and one named deliverable cover the first round. Usage telemetry from the tools you already license fills in the rest. Software earns its place when you track several hundred people across departments, a later problem than most buyers assume.

What does an AI enablement program cost?

Hands-on fluency workshops sit in a market range of roughly $2,000 to $15,000 by scope, session count and group size. Custom automation work runs far higher, commonly $25,000 to $150,000. Whatever the fee, add participant time at a loaded rate: on a 60-person team that is usually the larger half.

Is seat-license usage a reasonable adoption metric?

It works for license renewal and fails for capability. Activation shows a login, not a changed workflow. Two companies at the same activation rate can differ fivefold in how much work the tool touches, and only the second figure predicts whether the training paid for itself.

How do we separate the training’s effect from the tool’s?

Train one team in March and a comparable team in May, then compare the same deliverable on the same dates. The lag need not be long or the groups well matched. It needs to exist, so a mid-cycle product update lands on both groups and cancels out instead of being credited to your session.

Who should own the measurement?

Whoever owns the budget, not whoever delivers the training. The owner needs four hours a month and the standing to ask a manager for a timed task. At this company size the job usually lands on a chief of staff or VP of operations as a second role, which holds up as long as the four hours are scheduled rather than hoped for.

What would make a CFO approve the next program?

One page: the baseline figure, the figure at day 30, the method, the comparison group, and fully loaded cost per participant including time away from work. The result can be modest. What gets the signature is that it is checkable, and that whoever presents it names what the measurement could not show.

Key takeaways

  • The expensive part is not the fee. It is the second program you buy because you cannot tell whether the first worked, plus the participant time in both, which on a 60-person team exceeds the invoice.
  • One input is unrecoverable: a timed baseline on a real deliverable, taken before training. The rest can be rebuilt from invoices, calendars and admin consoles.
  • That baseline’s window is the week before delivery. It does not reopen, and no vendor has a commercial reason to flag it.
  • Measurement runs about 12% of a 60-person program’s all-in cost and is the only line item that yields a number at the end.
  • Four inputs cover the first round: one timed task twice, a rework count on it, workflow-level usage, and one sentence per manager per month.

If your team already has the licenses and nobody can say what changed, that is the common starting position, not an unusual one. We run a free AI-readiness call that works through which of these four inputs you can still collect, which window has closed, and what a defensible before-and-after looks like for your own deliverables. Book a time here.

Last Updated: Nov 19, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles