Scovai Scovai
People Analytics 2026-10-29 1 min read

Learning Access Is Falling While AI Use Climbs โ€” and It Is Your Strongest AI Users Who Say They Are Leaving

DSL

Dr. Sarah Liu

Learning Access Is Falling While AI Use Climbs โ€” and It Is Your Strongest AI Users Who Say They Are Leaving

Two lines in the same survey moved in opposite directions this year. Daily generative-AI use among workers rose from 14% to 22%, and 64% used AI at work in the past twelve months โ€” a ten-point increase. Over the same period, learning access went the other way: the share of workers saying they have the learning and development resources they need fell to 51%, down from 59% (PwC, 2026). That is 49,364 workers across 48 countries and regions and 29 sectors, fielded in May and June 2026.

Read it once more: adoption up ten points, learning access down eight. Whatever is funding your AI rollout, it is not a growing development budget. And the people best placed to notice are the ones most able to leave.

The Adoption Curve and the Development Curve Have Separated

The temptation is to treat 51% as a soft number โ€” a survey artefact, a complaint line. The surrounding data in the same release makes that difficult.

Workers reported more change, not less: 58% of those who experienced workplace change said there was more of it in the last year than in previous years, and 40% said they had to apply new skills in their job to a large or very large extent. Meanwhile 27% said fatigue or burnout limits their productivity, and only a third had money left at the end of the month, down 8% year on year (PwC, 2026).

So the demand for new skills rose, the pressure on workers rose, and the supply of structured development fell. That is not a morale story. It is a capacity story, and it has a specific shape.

Change is also concentrating at the top of the house. PwC found 51% of senior executives and 46% of managers had applied new skills in their role, against 29% of non-managers and 38% of entry-level workers. The learning is happening where the discretion already is.

One Budget, Two Retention Problems

PwC segments the workforce into four groups by skill scarcity and AI capability:

SegmentShareDescription
Front-runners14%Scarce skills, strong AI capability
AI insurgents18%Less scarce skills, using AI to deliver more
Indispensables11%Scarce, highly valued skills, early on the AI curve
Engine room56%Skills not scarce, early on the AI curve

The retention consequence lands at both ends of that distribution, and the two ends need different money.

Among front-runners, 29% say they are very or extremely likely to change employer in the next year. They over-index on being rewarded for challenging existing ways of working (1.5x the global average) and for learning new skills (1.2x), and they are +33 points more likely to ask for a promotion (PwC, 2026). This is a flight risk created by success. You rewarded them for building a portable capability, and the capability is portable.

At the other end, of the 56% in the engine room, only two in five say they have the learning access they need. They are the least likely to be rewarded for using AI, learning new skills, or challenging how things are done โ€” and they do most of the day-to-day work. That is not a flight risk. It is a capability risk, and it shows up later, as a bottleneck rather than a resignation.

An L&D budget spread evenly across headcount funds neither. It is too thin to hold the 14% and too generic to move the 56%.

The Confidence Gap Is Doing Quiet Damage

The survey's comparison between daily and infrequent AI users is the part worth taking to your next operations review. Daily users reported 68% confidence in their job security against 57% for infrequent users, and were +12 points more likely to ask for a promotion, +15 on trusting top management, and +21 on confidence they can learn new skills (PwC, 2026).

Those are correlations, not causal estimates โ€” confident people may simply adopt faster. But the operational implication survives either reading. If AI fluency tracks with trust in management and belief in your own learnability, then the half of your workforce without development access is also the half least likely to believe your change programme, and least likely to tell you so.

Learning Access Is an Organisational-Design Outcome, Not a Budget Line

Here is the counter-argument an ops lead should sit with before approving more training spend: firms that underinvest in portable skills are often behaving rationally, and more budget alone does not fix it.

New evidence from Deming, Sadun and Stanton reframes firm-sponsored training as an organisational-design outcome rather than a scale or generosity outcome (NBER Working Paper 35920, 2026). Drawing on a survey of 954 US firms with 100 or more employees, they find that having a skills taxonomy is associated with a 26-percentage-point increase in the probability that a firm runs a structured training programme, and a skills inventory with 17 points. Firms with 10,000 or more employees are 13 points more likely to run such programmes than firms with 100 to 999 โ€” and that scale effect is already controlled for in the specification that yields the 26 points.

That ordering matters for a 50โ€“500 FTE company. The measurement infrastructure predicts structured development twice as strongly as being twenty times larger. You cannot buy your way out with headcount, and you do not need to be enterprise-scale to fix it.

Their mechanism also explains the front-runner flight risk. Organisational design, the authors argue, is how firms make general skill firm-specific โ€” and firm-specific skill is how firms capture the return on training. Train someone on a generic capability with no role architecture around it and you have funded their next employer's onboarding. Train them inside a defined competency model, with a visible internal next step, and the return stays with you.

Their cluster analysis adds a warning: training tends to cluster around either managerial or frontline development, rarely both. Which is precisely the failure mode PwC measures โ€” development concentrated where the seniority already is.

The cost side compounds it. Gartner puts AI-related roles at three to four times average worker pay, with AI skill life cycles shortening to two to five years, and expects up to 30% of AI-displaced employees to be rehired by 2029, often at higher cost (Gartner, 2026). Expensive, fast-depreciating, mobile skill is the worst possible asset to hold without a development and succession structure around it.

Before You Act on the Segments, Check How They Were Measured

One caution, because it changes what you can do with this data internally. PwC's segmentation rests on workers' own accounts of their AI capability. Self-reported AI skill is a known weak signal: in GCheck's 2026 Automation Anxiety Report (N=1,500 US full-time workers, fielded April 2026), 63% said they had exaggerated or lied about their AI skills to appear more capable, rising to 80% among workers under 30, while 64% said their employer had never attempted to verify the claim (GCheck, 2026). It is a vendor survey โ€” GCheck sells verification โ€” so treat the magnitude as directional and the direction as solid.

The implication is not that PwC's global picture is wrong. Across 49,364 respondents, self-report inflation is largely noise in the aggregate. The implication is local: if you reproduce this segmentation inside your own company using a self-assessment form, you will mis-sort people, and you will mis-sort them in a biased direction โ€” your loudest adopters will land in the front-runner box whether or not they belong there.

Capability segmentation is only as good as the measurement under it. Assessed competency, demonstrated output, and observed task performance are different instruments from a confidence question, and they disagree systematically.

What a Head of Operations Should Do This Quarter

Four moves, in order of leverage.

1. Recompute your learning access rate by segment, not in aggregate. You almost certainly have a single company-wide training participation or satisfaction number. Split it: by level, by function, and by whoever your actual heavy AI users are (licence telemetry will tell you, not a survey). A 51% average can conceal 80% at the top and 40% in the middle โ€” which is what PwC finds globally. The gap, not the average, is the number to manage.

2. Build the taxonomy before you buy the content. If a 26-point effect sits with having a skills taxonomy, and that effect survives controls for size and sector (NBER Working Paper 35920, 2026), then the cheapest high-yield move is defining the competencies your roles actually require and measuring against them per cycle. This is the work that turns a training line item into a development system โ€” competency measurement feeding succession readiness and internal mobility rather than a course catalogue, the distinction Scovai's performance and succession layer is organised around.

3. Give your 14% scope, not another licence. The front-runners who are 29% likely to leave are being rewarded for challenging the status quo and learning fast. More tooling does not retain them; a visible internal next step does. Name the seats they could move into and the readiness window โ€” ready now, or twelve to twenty-four months โ€” and tell them. An internal market that works is the retention instrument here.

4. Verify capability before you staff AI-critical roles on it. Install a short applied-competency gate before assigning anyone to run, supervise, or redesign an AI-dependent workflow. Given a self-report inflation rate that tops 60% (GCheck, 2026), capacity planning built on declared fluency is planning on a number that is wrong by design.

The Decision in Front of You

Learning access falling eight points while AI use rises ten is not a line-item problem awaiting a bigger budget. It is a distribution problem, and distribution is a design choice you are already making by default.

So the question for this quarter is narrow: do you know your learning access rate for the 14% you cannot afford to lose, and for the 56% who do the work? If those two numbers are not on a page by the end of the quarter, you are not managing an AI skills divide. You are funding one.

Ready to go beyond the CV?

Scovai's AI-powered Talent Passport reveals what resumes can't: personality, potential, and true job fit.