Organizations are spending billions training employees to use AI — then sending them back to jobs, workflows, and workloads designed for a world without it, producing what one leading researcher calls "incremental and unmeasurable gains."

There is a $16 billion question sitting in the middle of enterprise AI adoption, and almost nobody is asking it.

Companies are investing heavily in AI training. Deloitte reports that 53% of organizations are "educating the broader workforce to raise AI fluency." McKinsey estimates that enterprise AI spending will double in 2026. The training market for AI skills is projected to reach $16 billion by 2028.

And the results? By the industry's own admission: disappointing.

Only 19% of AI use cases fully meet business objectives. Fewer than one in five organizations can demonstrate measurable ROI from their AI investments. And the most common explanation — "we need more training" — may be making the problem worse.

Exhibit 1 — The Adoption Paradox
MetricPercentage
Companies reporting regular AI use88%
Companies that have redesigned jobs around AI16%
Companies that have redesigned workflows around AI21%
AI use cases that fully meet business objectives19%

Sources: BCG 2026, Deloitte 2026, McKinsey 2026

Read those numbers together. 88% of companies are using AI. 19% are getting measurable results. And between those two numbers sits a gap that training alone cannot close.

The Training Trap

The logic behind AI training is sound on the surface: teach people to use the tools, and productivity improves. It's the same logic that drove IT training in the 1990s, digital transformation workshops in the 2010s, and agile training in between.

The problem is that training changes individual behavior. It does not change organizational design. And AI — unlike most previous technologies — doesn't just make existing work faster. It changes what work is possible, what work is necessary, and how work should be structured.

When you train someone to use AI and send them back to a job designed for 2019, one of three things happens:

Exhibit 2 — Three Outcomes of Training Without Redesign

Path 1: Surface adoption (most common). Employee uses AI for basic tasks — cleaning up emails, generating first drafts — but doesn't integrate it into core workflows. Usage is sporadic. Impact is unmeasurable. This is what 55% of trained employees settle into.

Path 2: Work intensification (emerging risk). Employee uses AI effectively and becomes more productive. But their job description, workload expectations, and performance metrics haven't changed. So they take on more work. A new study from Berkeley Haas found that 83% of AI-equipped workers reported increased workload — not decreased. The initial productivity gain becomes a burnout engine.

Path 3: Genuine integration (rare). Employee uses AI to fundamentally change how they approach their work. But this requires permission, redesigned processes, and organizational support that most companies haven't built. Fewer than 1 in 5 organizations have enabled this path.

Sources: HBR "AI Doesn't Reduce Work — It Intensifies It" (Feb 2026), Deloitte State of AI 2026

The Work Redesign Gap

Thomas Davenport, one of the most respected voices in enterprise AI, named this problem directly in MIT Sloan Management Review: individual uses of generative AI have resulted in "incremental — and mostly unmeasurable — productivity gains." The measurement problem is acute, he writes, because "it's too hard to measure either the productivity or the quality of output, and hardly anybody does measure it."

The reason it's unmeasurable isn't that the gains don't exist. It's that they're absorbed into unchanged job structures. If a governance analyst saves 3 hours per week on document drafting, but their job description, workload, and performance metrics remain the same, those 3 hours disappear into more of the same work. No one tracks them. No one reallocates them. No one reports them.

The organizations seeing real value from AI aren't the ones with the best training programs. They're the ones that asked a fundamentally different question.

Exhibit 3 — Two Questions, Different Outcomes

The training question: "How do we teach our people to use AI?" → Produces: workshops, licenses, usage metrics, adoption dashboards → Result: 88% usage, 19% measurable impact.

The redesign question: "Now that AI can do X, how should this job actually work?" → Produces: redesigned workflows, reallocated capacity, new performance metrics, measurable ROI → Result: McKinsey data shows workflow redesign has the single biggest impact on EBIT from AI.

Sources: McKinsey State of Organizations 2026, Davenport & Bean MIT Sloan 2026

What Work Redesign Actually Looks Like

Work redesign is not reorganization. It's not a restructuring exercise. It's the deliberate process of examining how work gets done at the task level and reconstructing it for an AI-augmented environment. It answers five questions for every critical role:

Exhibit 4 — The Five Redesign Questions

1. What tasks in this role can AI handle independently? These move to Tier 1 automation — output easily verifiable, low consequence of error, repetitive. They're removed from the human's workload entirely.

2. What tasks need AI assistance with human verification? Tier 2 — the human remains accountable, AI accelerates the work, and the redesigned workflow includes a verification step with clear criteria.

3. What tasks should remain fully human? Tier 3 — high consequence, high judgment, low verifiability. Protected from AI, and may receive more time and attention once Tier 1 is automated.

4. What does the human do with the time AI freed up? The question nobody asks — and the one that determines whether AI investment produces ROI.

5. How do performance metrics change? If the job description and KPIs stay the same, the redesign didn't happen.

The Intensification Risk

One of the most important studies of 2026 comes from Berkeley Haas, published in Harvard Business Review. Researchers studied 200 employees at a U.S. technology company over eight months (April–December 2025). Their finding challenges the core assumption behind most AI rollouts.

Exhibit 5 — The Intensification Cycle

AI makes Task A faster → employee takes on Task B with the freed-up time → expectations adjust, the new pace becomes baseline → employee uses AI for Task B, takes on Task C → workload expands, hours extend, cognitive fatigue sets in → quality declines, burnout increases, the "productivity gain" is consumed.

The data: 83% of AI-equipped workers reported increased workload. Burnout rates: 62% among associates, 61% among entry-level — compared to 38% among C-suite. The productivity gains were real — but unsustainable without workload redesign.

Source: Ranganathan & Ye, "AI Doesn't Reduce Work — It Intensifies It," HBR, February 2026

The researchers' conclusion: organizations need an "AI practice" — deliberate norms around pacing, sequencing, and human grounding that prevent productivity gains from becoming productivity traps. This is not an individual discipline problem. It's a management design problem. And it only gets solved through work redesign, not more training.

The Measurement Problem

Even when AI delivers genuine value, most organizations can't prove it. Davenport's observation — "hardly anybody does measure it" — reflects a structural problem, not a laziness problem.

Exhibit 6 — What Gets Measured vs. What Matters
What companies measureWhy it doesn't workWhat to measure instead
AI tool adoption rateLicenses don't equal usage, usage doesn't equal valueHours reallocated from automated tasks to higher-value work
Number of AI queries per monthVolume ≠ impact; high counts may indicate poor workflowsQuality and accuracy of AI-assisted output vs. manual baseline
Employee satisfaction with AI toolsSatisfaction doesn't correlate with productivityBefore/after cycle time for specific workflows
Training completion ratesCompletion ≠ competence, competence ≠ integrationDemonstrated behavior change in target workflows

The measurement framework that works starts at the workflow level, not the tool level. Measure the workflow before AI, redesign it with AI, and measure it again. The delta is your ROI.

What to Do Instead

1. Stop treating training as the strategy. Training is a component of the strategy. It is not the strategy itself. The strategy is work redesign — training enables it.

2. Redesign 3-5 critical workflows before scaling training. Pick the workflows with the highest time cost and clearest automation potential. Decompose to the task level. Classify each task. Redesign the workflow. Then train the people who do that work on the redesigned process — not on AI in general.

3. Reallocate the time AI frees up — explicitly. If AI saves a team 20 hours per week, those 20 hours need a destination. Otherwise they'll be consumed by work intensification.

4. Measure at the workflow level, not the tool level. Stop tracking adoption dashboards. Start tracking cycle time, quality, error rates, and capacity metrics for specific redesigned workflows.

5. Build judgment alongside automation. For every workflow you automate, ask: "How will the next generation of employees learn to do this well enough to verify AI's work?" If you don't have an answer, you're building a capability debt.

The Bottom Line

The AI skills gap is real. But it's not the gap most organizations think it is.

The real gap isn't between trained and untrained employees. It's between organizations that train their people and organizations that redesign their work. The first group gets 88% adoption and 19% measurable results. The second group gets measurable ROI, sustainable productivity, and a workforce that knows how to use AI wisely — not just frequently.

Training without redesign is like teaching someone to drive and putting them back on a horse trail. The capability is there. The road isn't built yet. The organizations that build the road will outperform the ones that simply buy more driving lessons.

Sources

Talk to me about your AI strategy →