Your AI Training Is Working. Your Workflows Aren't.

An L&D director I know launched an internal AI training program last quarter. Eighty percent completion rate. Strong engagement scores. Six months later, when she pulled the productivity data, nothing had moved. Not adoption, not output, not time saved on the workflows the training was supposed to change. She is one of dozens of L&D leaders I have spoken with in the past year describing the same gap.

This is the story most L&D teams are living right now. The training is fine. The tools work. People say they like the tools. But the productivity bump that justified the investment in the first place is nowhere to be found, and nobody wants to be the one to say it out loud.

The Pattern Nobody Is Naming

Companies have spent the last two years building AI literacy programs. Prompting workshops. Tool selection guides. Use case libraries. The training infrastructure is real, and so are the completion certificates. The output metrics are not.

The MIT NANDA State of AI in Business 2025 report puts a number on it. After surveying 350 employees, interviewing 150 leaders, and analyzing 300 public deployments, the researchers found that roughly 95 percent of enterprise generative AI pilots produce no measurable impact on P&L. A separate analysis tracked AI adoption climbing from 61 percent to 71 percent of firms in a single year while 89 percent of managers reported no change in productivity over the same period. Adoption went up. Output stayed flat. The cause is not model quality. It is what the report calls the learning gap, meaning the structural distance between adopting a tool and redesigning the work that tool is supposed to change.

The Hidden Cause: A J Curve You Cannot Skip

This pattern is not new. Brynjolfsson, Rock, and Syverson (2021), writing in the American Economic Journal: Macroeconomics, documented the same dynamic across every general purpose technology of the last century. They call it the Productivity J Curve. The intangible investments required for a new technology to actually pay off, process redesign, role changes, new measurement systems, take years to build. During that period, productivity dips before it climbs. Companies that skip the intangible work never reach the climb.

The California Management Review made the practical case in late 2025. Organizations that bridge the AI transformation gap are not the ones with more training hours. They are the ones that treat AI as a reason to redesign processes rather than a layer to bolt onto existing ones.

The reason your AI training program isn’t moving productivity numbers is probably that you trained the people but never changed the work. Tool literacy is necessary. It is not sufficient.

Here is what to do about it.

Tie every training cohort to a workflow redesign. Don’t approve another AI training program that doesn’t name the specific workflow it intends to change. Who owns that workflow? What does it look like today? What does it look like after the training? If nobody on your team can answer those three questions before the program launches, you are funding literacy theater, not capability building.

Measure the work, not the learner. Completion rates, satisfaction scores, and the confidence ratings learners give themselves are leading indicators of nothing. The right metric is cycle time, error rate, or output per hour on the actual workflow you set out to change. If you cannot draw a direct line from training spend to one of those numbers, your AI program is producing engagement, not impact.

Build a workflow owner into every cohort. The MIT data shows that successful AI rollouts pair the people learning the tool with someone who has the authority to change the process around it. A learner alone cannot redesign work. A process owner alone has nothing to redesign with. Put them in the same room from day one, and give them a shared deliverable.

Expect the J Curve and budget for it. Brynjolfsson’s research is honest about the timeline. The first six to eighteen months of serious AI integration look worse on paper than doing nothing. The companies that quit during the dip never see the climb. The ones that hold the line, and keep investing in the intangibles, do.

Most AI training programs are not failing because of the training. They are failing because the work around the training was left untouched. A new tool in an old workflow gets you a slower old workflow.

Look at your most expensive AI program right now. Can you name the workflow it was supposed to change, the person who owns that workflow, and the metric that should move? If even one of those is missing, you already know what is wrong.

At Peak Potential, we help organizations design AI capability programs that pair tool training with the workflow redesign work that makes the training stick. If your AI investment is showing up in completion reports but not in productivity numbers, let’s talk.

Frequently Asked Questions

What is the ‘learning gap’ in AI adoption?

It’s the distance between adopting an AI tool and redesigning the work that tool is meant to change. The MIT NANDA State of AI in Business 2025 report uses it to explain why roughly 95 percent of enterprise generative AI pilots produce no measurable impact on P&L. People learn the tool, but the workflow around it stays as it was, so output never moves.

Why isn’t more AI training the answer?

Because training builds tool literacy, and literacy on its own doesn’t change how work gets done. Per the California Management Review, the organizations that close the gap treat AI as a reason to redesign processes rather than a layer to bolt onto the ones they already have.

What is the Productivity J Curve, and why does it matter for AI?

Brynjolfsson, Rock, and Syverson (2021) showed every general purpose technology needs years of intangible investment before it pays off, and productivity dips before it climbs. For AI that means the first six to eighteen months of serious integration can look worse on paper than doing nothing, and the companies that quit during the dip never reach the climb.

How do I know if my AI program is actually working?

Stop measuring the learner and start measuring the work. Completion rates and satisfaction scores tell you almost nothing. The signal is cycle time, error rate, or output per hour on the specific workflow you set out to change.