Completion Was Never Proof: Why L&D Must Measure Real Capability Change

The Question Finance is Actually Asking.

Every enterprise learning function has a version of the same slide. Enrollment up. Completion at 87 percent. Satisfaction in the low fours. Hours consumed trending the right way.

None of it answers the question finance is actually asking, which is whether anyone can do something this quarter that they could not do last quarter.

Completion measures attendance. It was never designed to measure capability, and the profession has spent two decades asking it to carry weight it cannot hold. The consequence shows up in the numbers: The Josh Bersin Company's February 2026 study of 800 organizations found that 74 percent report they are not keeping pace with their own demand for new skills, against more than $400 billion a year in global corporate training spend. Those two figures cannot both be true unless a meaningful share of that money is landing somewhere other than capability.

It’s worth being honest about why completion won. It was cheap to count. Attendance is a fact an LMS captures for free. Proving capability change requires measuring the same people twice, on instruments rigorous enough to survive scrutiny, and for most of the last twenty years that was operationally out of reach for most organizations outside of national assessment bodies. So we reported what we could count and hoped the training transfer held.

One baseline is not a measurement system

The organizations that did move past completion metrics then hit a second wall. A single skills assessment served as a training course capstone to "prove" training worked.

But a capstone is administered only after the fact, and that is a bigger problem than it sounds. With no measurement taken beforehand, there is no point of comparison, which means there is no way to know whether the program had any effect whatsoever. A score of 82 at the end of a course is not evidence the course did something. Absent a baseline, it is indistinguishable from what that person already knew walking in.

And even a capstone with a baseline behind it is still a photograph. The business needs a feed.

PwC's 2026 Global AI Jobs Barometer, built on more than a billion job advertisements across 27 territories, found that the skills employers ask for in the most AI-exposed roles are changing more than twice as fast as in the least exposed ones. That gap matters more than any headline rate, because it means decay is uneven. The same twelve-month-old baseline can still be roughly accurate for one function and badly stale for another, and nothing in a completion report will tell you which is which. The data has not become wrong. It has become historical, aging at different speeds, across different parts of the company.

Executive expectations have moved with it. "We ran the upskilling program" was acceptable when boards treated development as a cost of doing business. It is not acceptable now that the same boards are funding AI initiatives against explicit readiness commitments.

What proof actually requires

Three design requirements. Most organizations get the first, some get the second, almost none get the third.

Measure rather than infer. Resumes, self-ratings, manager assessments, and completion records all measure something. None of them measure capability.

Self-ratings are the best-documented failure. Across more than 88,000 enterprise assessments in Workera's 2026 benchmark, only 11 percent of employees accurately placed their own skill level, with the rest split between significant overconfidence and significant underestimation.

Manager ratings are noisier than they appear, for structural reasons rather than careless ones. A manager observes a fraction of any individual's actual work. Organizational politics colors what they do observe. And there is a standing incentive to rate generously, since a team that looks strong reflects well on whoever is said to have built it.

The remaining inputs share a different flaw. Resumes, project histories, and course records are exposure data. They establish that someone was present for an opportunity to develop a skill, which is a yes-or-no fact about their calendar. They do not establish whether the person took that opportunity, and they say nothing at all about how far they got. A development portfolio built on exposure data is optimizing against noise.

Measure the same people before and after. The unit of analysis has to be the individual, tracked across the intervention, rather than the population in aggregate.

That distinction is easy to wave past, so it is worth making concrete. Aggregate averages move for reasons that have nothing to do with learning. If the group measured in March is not the same group measured in September, because people joined, left, or opted in at different rates, the average will shift on composition alone. Aggregate scores that rise while the population underneath them changes tell you about your sampling, not your program.

Use comparable forms, not identical ones. This is where in-house efforts break, and it is the difference between evidence and theater. Reuse the same questions before and after an intervention and the improvement you observe includes memory, with no way to separate genuine gain from recall. Swap in a different set of questions that was never held to a comparability standard and you have two numbers on two scales and can’t interpret delta at all.

Taylor Sullivan, PhD, who leads Product and Assessments at Workera and trained as an industrial-organizational psychologist, puts the problem this way:

"Percent correct is form-dependent. It isn't always comparable across two different sets of questions, which means the most common way organizations report growth is measuring the forms as much as it's measuring the people. We build every form to be equivalent in structure, content, and difficulty, so scores are truly comparable across all of them. That said, what I wouldn't claim is that you establish comparability once and then stop thinking about it. You establish it as well as your data allows, then you keep monitoring live responses, because that's where the evidence actually accumulates."

That last point is the one worth sitting with, because it reframes what a learning organization is buying. Comparability is not a checkbox cleared before launch. It is an ongoing obligation, and the honest version of the pitch says so. A growth number a CFO can dismantle in one question is worse than no number, because it costs the function credibility it will need later.

Four ledgers, one input

Once the delta is defensible, it becomes the shared input to four separate financial arguments. Illustrative arithmetic below for a 5,000-employee organization with 1,000 in a measured technical population and a $2 million learning budget. Every assumption should be replaced with your own.

Build talent instead of buying it. 

Roles filled internally × fully loaded cost to fill

Result: Fifteen avoided external hires at $30,000 returns $450,000

The measurement is what makes that number real rather than aspirational. Mobility programs stall not for lack of ambition but because nobody can identify who is genuinely ready, so the safe default is always to post the role externally. Verified capability data is what turns a hiring manager's hunch about someone into a decision they will actually sign.

‍

Rationalized spend. 

Learning budget × share redirected from low-impact to high-impact programs

Result: Ten percent of $2 million returns $200,000.

That reallocation is hard to justify today, because the usual signal for whether a program earns its renewal is usage, and usage always looks fine. Seats fill. Courses complete. Dashboards stay green. None of it distinguishes content people needed from content they merely consumed. You find the waste by comparing where the capability gaps actually sit against where the money actually went, which of course requires knowing where the gaps actually sit.

‍

Recovered time.

Participants × hours not spent on already-mastered material × loaded hourly rate 

Result: One thousand people, 20 hours each, at $70 an hour returns $1.4 million.

Personalization is only real if the system can exempt someone from sitting through content they can already demonstrate. The reverse case is at least as valuable and gets discussed far less: the same measurement surfaces gaps nobody had flagged, in people whose managers had no particular reason to suspect one. Unrecognized strength wastes time. Unrecognized weakness gets shipped into production.

‍

Productivity, including token spend.

Measured population × current per-person token spend × efficiency gain

Gartner forecast in June 2026 that AI coding costs will overtake the average developer's salary by 2028. Roughly a quarter of technology leaders already report $200 to $500 per developer per month, and about 6 percent exceed $2,000. Applied to a measured population, that market data becomes a specific number: a thousand engineers at $350 a month is $4.2 million annually, and a 15 percent efficiency gain returns $630,000.

The takeaway

Stop defending the budget with completion data. It has never worked, it works less every year, and leading with it trains the C-suite to treat development as an expense to manage rather than an investment to size.

Three questions for the next leadership review. For our largest program last year, what was the measured capability delta, and against what instrument? If we ran that measurement again today, would the result be genuinely comparable to the baseline, or only similar-looking? Which of the four ledgers could we populate with evidence rather than assumption?

Completion proves people showed up. Verified capability change proves the investment worked. Only one of those belongs in a board deck.

‍

‍

‍

‍

‍