We're throwing billions at AI training. Very little of it sticks. Roughly 1 in 50 enterprise AI investments produce meaningful ROI, largely because organizations invest in technology without upskilling the people who use it. That's not a training failure. That's a workflow failure.

I've watched this unfold in three different organizations over the past 18 months. The pattern is always the same: Leaders see the skills gap, green-light the training spend, launch LMS courses, track completion rates, celebrate their "upskilling initiative," and then watch people leave because they learned skills they can't actually use in their real jobs.

Over 90% of global enterprises will face critical skills shortages by 2026, representing an estimated $5.5 trillion in unrealized productivity globally. The crisis is real. The response is theater.

The Workflow-Before-Training Rule

McKinsey argues that upskilling should be treated as a change management initiative, not a training program. The companies that succeed are the ones that redesign workflows around AI first, then train people for the new workflows. Training without workflow redesign just teaches people skills they cannot use.

That's the decision fork every leader needs to understand. You cannot train your way out of a process problem.

Let me make this concrete. You have a financial analyst team. You deploy an AI forecasting tool. The temptation is immediate: run everyone through a 4-week course on the tool, send them back to their desks, and assume the problem is solved. Six months later, your best analysts have left because they're spending 80% of their time manually validating AI outputs instead of doing strategic analysis. You didn't fail at training. You failed at redesigning the role.

The Three-Step Framework

Step 1: Redesign the workflow, not the tool.

Before any training begins, map how the job actually changes when AI enters it. Not the optimistic version—the real version, with all its friction. Where does human judgment still matter? Where does AI automate mundane work? Where does it create bottlenecks? Redesigning the workflow around AI capabilities, then training people for the new process, produces real productivity gains, whereas adding AI tools to old workflows creates confusion.

If your legal team uses an AI document analyzer, don't train them to use the tool. First, redesign the document review process so the tool sits upstream of their work, not alongside it. Then train them to use the redesigned process.

Step 2: Tier your upskilling, not your training.

89% of organizations say upskilling existing employees is more cost-effective than hiring new talent. But cost-effectiveness only works if people actually stay and use what they learned. Avoid the "spray and pray" approach where everyone takes the same AI course.

Instead, identify which roles need upskilling vs. reskilling. An upskilling approach works best when roles and expectations stay the same, but AI skills will make employees more productive. Reskilling becomes essential when AI changes workflows or shifts responsibilities. A business analyst learning to use AI tools within an unchanged workflow? Upskilling. A data entry person whose job is being absorbed by automation? Reskilling—or honest conversation about their future.

Not every displaced worker can transition to AI roles. The skills gap between an administrative assistant and an AI engineer is not bridged by a 6-week course. Honesty here prevents the false promise that everyone can reskill into senior roles.

Step 3: Measure retention and workflow productivity, not course completion.

Stop counting finished courses. Track whether reskilled employees stay 18 months post-training, and whether the workflows they support actually became more productive. One-third of workers experienced 15 or more major changes in the past year alone, and only 27% think their organizations manage change well. Upskilled employees leave when change feels chaotic—even if training was excellent.

Meaningful metrics: Are people using the new skills in their actual jobs? Are they completing projects faster? Are they staying? If the answer to any of these is "no," the problem isn't training quality. It's that you haven't built the workflow to support what you taught them.

The Real Retention Lever

Companies that reskill before they replace have lower costs and institutional knowledge stays. Amazon, JPMorgan, and AT&T are proof that retraining works at scale. The cost is lower and the institutional knowledge stays. That's the case for reskilling as a retention tool—but only if you do it backwards from how most organizations approach it.

Workflow first. Training second. Retention follows.

Stop building better courses. Start redesigning the work.