You bought the tools. The work didn't get better.

Someone bought the licences. Someone said transformation a few times. Four years into the AI adoption journey, new questions emerge. How can we re-design work processes to get the best out of humans and technology? How do we build judgement and experience when AI does the grunt work?

Your managers were handed the mandate with a training budget and a usage dashboard. The dashboard says eighty percent active, the leaderboard ranks people by how many messages they sent to your LLM.

Underneath it, things look less promising:

• People experimenting with no method, often in tools nobody approved
• Work coming back polished and hollow, and the person who received it quietly redoing it
• Some of your best people refusing, for reasons they're careful not to say out loud
• A question from finance about ROI that you cannot currently answer

Most organisations are exactly here. Almost none of them are saying so.

Because the people getting the most out of AI have the most to lose by saying so.

If AI has changed your work a great deal, admitting it invites a question about what you're still valuable for. If it hasn't changed at all, admitting that invites a question about whether you're keeping up. Both answers cost something. The safest available answer, for almost everyone, is that it's going fine.

My research calls this the silence cascade. The deeper someone's engagement with AI, the higher their identity stakes, and the less likely they are to disclose it. It's why usage numbers can look healthy while the organisation learns nothing. The learning is happening. It just happens privately, one person at a time, and never becomes shared.

No course fixes that. The thing being avoided isn't a skill.

Three things get sold as the answer. None of them are.

Drive adoption. Measures whether people opened the tool, not whether the work got better. You already have that number. It's why you're reading this.

Maximise seat usage. Makes usage the goal, so you get usage. You're left with a higher bill and more output to check, produced by people who were told volume was the point.

Prompt engineering. Teaches people to hoard prompts, search for the perfect formulation for each problem, rather than when to hand work to AI, and how to tell whether what came back is any good.

That gap has a cost, and it lands on whoever receives the work. Stanford and BetterUp Labs named it workslop: work that looks finished and isn't, passed to a colleague who then spends close to two hours repairing it. Around two in five workers reported receiving some within a single month. The productivity doesn't disappear. It moves downstream, onto someone with no budget line for it, and it takes trust with it.

There's a longer reason underneath, and it predates AI by about twenty years.

Judgement was designed out of most organisations on purpose. Process was standardised. Oversight was added. Innovative behaviour was replaced with innovation frameworks. It was a rational trade at the time: distributed judgement is expensive to coordinate, and standardisation is cheap.

AI is what makes the bill arrive. You can't ask people to exercise discernment over machine output if the last two decades were spent designing the discernment out.

The magic word for AI-powered workflows is "human in the loop". Agent are built to autonomously complete complex tasks and a human specialist checks the agent's work before it's forwarded on. That works because the specialist likely has decades of experience in their field and knows what to look for. They built judgement through years of experience, learning from senior peers and their own mistakes. That path is getting lost, and it has to be rebuilt at speed. Junior work is being absorbed by AI, grad programmes are scaled down or redesigned, and nobody knows exactly how.

Take insurance for example. The product sold is an educated guess how likely it is for something to go wrong, and a promise to make it right if needed. Both at underwriting and at claims stage judgement and compromise are possibly the most important skills. Even with the best models, you can never have complete visibility, and claims often need to navigate what's not been spelled out.

An experienced professional can make those judgements even without perfect data. They will ask the right questions and eventually find a compromise that works for everyone, balancing the bottom line, relationships and trust. The question is how we build that judgement when AI does the work your grads and juniors used to do.

Capability — in the people. Not prompt technique. The choice before the prompt. When to outsource thinking to AI and when to use AI to make you think harder. Being able to reason what data the AI needs to complete the task. Learning to pick deliberately, step by step, is the capability. It survives the next model release, which prompt formulas don't.

Conditions — in the organisation. Making it safe to say what's actually happening. Deciding where a person stays in the loop, and being able to say why. Building judgement while automating processes. Redesigning the handful of workflows that matter, with the people who do that work rather than for them.

Some of this looks like slowing down, and it is. Getting hands-on with the teams as they figure out what might work for them. Crucially, it needs permission to produce things that don't. Skipping the critical thinking is not faster. It moves the cost to whoever receives the work.

You can show what changed in the work, not just how many messages were sent. You know which licences and which use cases are worth paying for.

My MSc research produced a measurement instrument for AI adoption, designed to find what usage data structurally cannot. Where adoption is real. Where it's being performed. Where it's stalled, and why. And what people won't say in a meeting. Bookend surveys at either end of the sprint show whether the conditions and outcomes shifted, and lightweight pulse surveys show how the team uses AI and whether learning spreads between people.

When this is working, what you hear back doesn't sound like a productivity metric. It sounds like something returned.

Your people build judgement on real work. After one sprint, judgement on how and when to use AI; after a few, an answer to how you build long-term judgement and capability across the workforce.

You might not be the only person involved in this decision. Or maybe you just work better reading things offline. Download the fact sheet for a quick reference, or to distribute to your peers.

Phase one. 90 days. £8,000 per month.

• A measured baseline of where AI adoption actually stands
• A readout for your board, with the gap costed
• A weekly working session with your team, and support in between
• One or two workflows redesigned with the people who run them

At the end you have evidence, a plan aligned to your business goals, and enough internal capability to keep going.

After that it continues in 90-day sprints, or it stops. No minimum beyond the first ninety days. I work with only two organisations at a time. This way I can focus on adding real value and embedding the learnings in your team.

For twenty years I was the supplier. Brought in to run the training, deliver the system, hand over whatever had been signed off. It is the best seat in the house for watching an organisation pay for something and then fail to absorb it. Six onboarding calls. An invite list that keeps growing. People showing each other that they care by asking stupid questions. And then nobody feels safe to engage with the actual subject matter.

I have also been in the rooms where the penny drops and a team unlocks a way of working it keeps. So I know what makes adoption fail, and I know what makes it stick.

Behind that: an MSc in organisational psychology, an AI-adoption measurement instrument built from original research, and a continuing set of interviews with people doing this work inside real organisations. Some of the quotes on this page come from those.