"Digital skills" has become one of the least useful phrases in workforce policy. It collapses a 22-year-old learning Python, a 55-year-old learning how to triage an AI-drafted email, and a refugee worker learning how to scan a UPC code into a single category that funds bear no relationship to. The Organisation for Economic Co-operation and Development's 2024 Skills Outlook makes this point with characteristic restraint: "the term 'digital skills' encompasses a range of competencies whose labor-market returns differ by orders of magnitude." Treating them as one bucket is why so many digital-skills programs show such weak labor-market returns.
This piece argues that the most defensible strategy for "equipping workers with essential digital skills" is to stop trying to do it generally and start funding the three specific skill clusters where the returns are unambiguous: fluency with AI-assisted writing and analysis, comfort working in cloud-based collaboration suites, and the data-literacy minimum needed to interpret a spreadsheet, dashboard, or model output. Almost everything else in the standard "digital skills" curriculum either has weak returns or is acquired faster on the job.
Where the empirical returns actually concentrate
Daron Acemoglu and David Autor's body of work on skill-biased technical change is the cleanest framework for thinking about this. Their 2022 NBER paper "Tasks, Automation, and the Rise in U.S. Wage Inequality" shows that the wage premium for "routine cognitive complement" skills — the ability to direct, evaluate, and integrate machine output — has grown faster than any other skill category over the past two decades. That premium is rising fastest, post-2022, for workers who can use generative AI competently in their existing job, not for workers who change jobs into "tech."
That finding has held up under the 2023 Brynjolfsson/Li/Raymond study of customer-support workers (14% productivity gain on average, 35% for less-experienced workers), the BCG/MIT/Wharton consulting study (25% faster, 40% higher quality on appropriate tasks), and Goldman Sachs Research's 2023 industry analysis that estimated AI could add 1.5 percentage points per year to U.S. productivity growth — concentrated in workers who use it, not in workers whose jobs are about it.
The implication for skills policy is clear. The training that moves the labor-market needle is short, applied, and embedded in actual jobs — not multi-month "bootcamps" that produce certificates few employers value. The OECD's PIAAC longitudinal data shows that short, employer-integrated training produces meaningfully larger earnings impacts than stand-alone reskilling programs, particularly for workers over 40.
Three clusters that earn their place in the curriculum
AI-assisted work fluency
The minimum competency here is roughly: write a prompt that produces a useful first draft, recognize when the output is wrong, edit it into a final, and not paste confidential data into a public model. That is a 6–10 hour training, not a multi-week course. The Stanford/MIT field experiments cited above suggest this is the highest-return single training a typical knowledge worker can receive in 2026. Microsoft's 2024 Work Trend Index found that AI-fluent workers reported saving an average of 30 minutes per day, with the largest gains concentrated in admin-heavy roles.
Cloud collaboration suites
Mastery of Microsoft 365 or Google Workspace — shared documents, version control, basic automations, secure sharing — has become the rough equivalent of typing literacy in the 1990s. Workers who lack it pay a steep penalty in remote and hybrid roles, where the OECD's 2024 employment data shows roughly 28% of U.S. workers now spend at least part of their week. This is also where many older workers face the highest friction; AARP's 2024 workforce surveys consistently find that lack of confidence with collaboration suites is among the top three barriers to older-worker re-employment.
Data literacy
Not statistics. Not coding. The ability to read a chart honestly, recognize when a denominator is missing, interpret a percentage change, and use a pivot table is, per Brookings Institution analyses, the single largest predictor of within-occupation earnings growth in white-collar roles. Most adults underperform on these tasks; the OECD's PIAAC numeracy assessments find that roughly 30% of U.S. adults score below the level needed for routine workplace data tasks. That gap is fixable and the fix has a strong return.
The delivery mechanisms that actually work
The strongest recent evaluation of workforce training programs is MDRC's multi-site randomized study of sector-based training, finalized in 2023, which found earnings impacts of $4,000–$6,000 per worker per year — but only when programs combined three elements: (1) employer-co-designed curriculum, (2) wraparound supports (childcare, transportation, income replacement during training), and (3) direct job placement at a partner employer. Programs missing any of those three elements showed earnings impacts statistically indistinguishable from zero. The Department of Labor's own 2024 review of Workforce Innovation and Opportunity Act outcomes corroborates this — WIOA's per-participant earnings impact is essentially zero in most states absent these supports.
That is a strong finding, and it is the basis for our broader argument in Reskilling for Real →: digital-skills funding does not fail because the skills are wrong. It fails because the program structure routes most participants out before completion.
What employers can actually do this year
The most-cited recommendation in this category — "create a learning culture" — has not survived empirical testing as a standalone intervention. What has worked, in firms studied by McKinsey's 2024 "The State of AI" report and Gallup's 2024 workplace data, is concrete and unglamorous: paid training time (not "encouraged" or "after-hours"), a single accountable owner per skill area, an internal mobility marketplace so workers can apply skills they've learned without leaving the firm, and explicit recognition of skill gains in performance review. Firms that did these things saw measurable retention and productivity benefits. Firms that funded LMS subscriptions and posted exhortations on Slack did not.
The cost is real but bounded. Roughly 1.5–2% of payroll, deployed on the structure above, produces returns most CFOs would otherwise reach for an enterprise software contract to get. McKinsey's data suggests the productivity payback period is 12–18 months for typical knowledge-work organizations.
"Digital skills" is too broad to fund well. The three clusters worth investing in are AI-assisted work fluency, cloud-collaboration competence, and basic data literacy. Everything else is downstream.
What governments can do that they aren't
Singapore's SkillsFuture Credit gives every citizen over 25 a personal training wallet. The U.K. ran a similar pilot with the Lifelong Learning Entitlement. France has the Compte Personnel de Formation, funding nearly 2 million workers per year. The U.S. has nothing equivalent at federal scale. A $2,000-per-worker annual training wallet would cost roughly $300 billion over a decade — less than the Inflation Reduction Act's clean-energy provisions, and addressing a population four times larger. Whatever its sticker price, the empirical record from peer economies is that portable training accounts work. Whatever the U.S. spends on workforce development that does not include them is being asked to do the job with less effective tools.
The workers we keep saying we want to "equip with essential digital skills" already know the skills they need. The question is whether the country's training apparatus is built to deliver them — or whether it is still optimized for a labor market that no longer exists.
Updated May 21, 2026. This piece was substantively rewritten as part of NWLB's 2026 editorial refresh.
