AI training fails when the workflow stays the same

The workshop goes well. People are engaged, a few are genuinely excited, and everyone leaves able to do things with AI they could not do that morning. Three weeks later, the team is working exactly as it did before. This is the normal outcome of AI training, and the frustrating part is that the training was usually fine. The problem is what happened, or rather did not happen, around it.

Training changes what a person can do. It does not, by itself, change the approvals, the incentives, the data access, the responsibilities or the quality checks that surround their work. Leave all of that untouched and the old way remains the path of least resistance, so people drift back to it the moment the enthusiasm fades. The skill was never the binding constraint. The workflow was.

Why teaching the tool is not enough

This is one of the most robust findings in the economics of technology. Brynjolfsson, Rock and Syverson's productivity J-curve describes how the payoff from a new general-purpose technology lags precisely because the value depends on complementary changes, redesigned processes and new ways of working, not on the tool alone. A workshop delivers the tool and skips the complement, which is why it so often produces a brief lift and no lasting change.

The pattern shows up in the field too. The MIT Media Lab's GenAI Divide report concluded that what separated organisations getting value from those not getting it was learning and integration, not raw capability or training volume. And in the large customer-support study by Brynjolfsson, Li and Raymond, the productivity gains came from the AI being built into the workflow, with good examples and context in the flow of the work, not from agents being sent on a course. The lesson is consistent: adoption is a property of the workflow, not the headcount you have trained.

Learn, apply, redesign, reinforce

Training works when it is one step of four, not the whole plan. We use this sequence.

Learn. Build the skill. This is the workshop or the lesson, and it is necessary. It is simply not the finish line, which is where most efforts stop.

Apply. Use it immediately, on real work, while it is fresh. A skill practised once on a sandbox exercise and then parked for a fortnight is a skill lost. The gap between learning and first real use should be days, not weeks.

Redesign. Change the workflow around the task so the new way is the easy way. Move the approval step, update the template, rewrite the standing instructions, reassign who checks what. This is the step that makes adoption stick, and it is the one almost everyone skips because it is harder than booking a trainer.

Reinforce. Keep it alive. Share the examples that worked, give a champion the time to help, recognise the people doing it well, and revisit it as the tools change. Without reinforcement, even a redesigned workflow erodes under the pull of old habits.

Train the person, leave the workflow untouched, and nothing changes.

It is worth naming why the last two steps get dropped. Learning and applying feel like progress and fit neatly into a calendar; redesigning and reinforcing are slow, sometimes political, and never quite finished. The government's own AI Opportunities Action Plan frames adoption as a "scan, pilot, scale" sequence for the same reason: the value comes from scaling and sustaining a change, not from the moment someone first meets a tool. A workshop is the easy quarter of the work. Treating it as the whole job is the most common way organisations spend real money on AI training and see nothing change, then conclude, wrongly, that the training itself did not work.

What this looks like in practice

Consider an accountancy firm that trains everyone to draft client letters with AI. The session is good and people leave keen. But the letter still goes through the same partner approval, the old template is still the one saved in the system, and nobody has said whose job the AI draft now is. Within a month the team is back to writing from scratch, because nothing in the actual workflow rewarded the new way. The training did not fail. It was never given a workflow to land in.

Contrast a firm that treats the same training as step one. The standing prompt and an updated template are built into the letter process, the approval step is adjusted to check the AI draft rather than ignore it, one person owns the workflow, and the team reviews what is working each month. Same workshop, completely different result, because the other three steps happened. (Illustrative examples, not specific firms.)

The honest limits

Two cautions. First, redesign is genuinely harder than training, and that is exactly why it gets skipped. A workshop is a line item you can book; changing a workflow means touching approvals, habits and sometimes egos. There is no way around this: if you are not willing to follow training with redesign, the honest move is to spend the training budget on something else. The work is the point.

Second, do not try to redesign everything at once. The discipline of choosing one or two workflows applies here too; a redesign spread across ten processes lands on none. Pair each piece of training with one specific workflow change, finish it, then move on.

What to do next

Before the next workshop, decide which single workflow it is meant to change, and what specifically will be different about that workflow afterwards: which step moves, which template changes, who owns it. If you cannot answer that, you are buying a morale boost, not a capability. Pair every piece of learning with a redesign and a way to reinforce it, and the same training that used to fade will start to stick.

The tool

To turn a single lesson or workshop into a changed way of working, we have built the 30-Day AI Learning-to-Adoption Plan: a week-by-week follow-through across learn, apply, redesign and reinforce, with the workflow change named, an owner against each step, and the review questions that tell you whether it actually took.

Download the 30-Day AI Learning-to-Adoption Plan (PDF)

Redesigning real workflows around AI, not just training on them, is the heart of a 90-Day Adoption Build with a team. It follows directly from the 90-day plan for getting started and from the difference between task productivity and genuine business value.

Sources and further reading

• Brynjolfsson, Rock and Syverson, The Productivity J-Curve, NBER Working Paper 25148, 2018. Independent. Source for value depending on complementary process change, not the tool alone.
• MIT Media Lab (Project NANDA), The GenAI Divide: State of AI in Business 2025. Industry report, not peer-reviewed; the headline 95% figure has been contested. Source for value sitting in learning and integration.
• Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161, 2023. Independent. Source for gains coming from AI built into the workflow with good context.
• UK Government, AI Opportunities Action Plan, January 2025. Source for its "scan, pilot, scale" framing of adoption as a sequenced effort, not a one-off event.