If AI does the 80%, how do humans learn the 20%?

In our earlier What Am I Missing? pieces, we used the trolley problem to explore decisions that become difficult precisely because there is no perfect answer. We started with safety and quality, and then added delivery, because in the real world, particularly in healthcare and radiopharmaceuticals, decisions rarely involve one objective at a time. There may be a patient waiting, a product that cannot simply be produced again tomorrow, or consequences both to releasing and to not releasing.

There are, however, two dimensions we have not explored in the same way yet: cost and people. AI brings both of them straight into the discussion.

I use AI a lot, not because I think it can replace professional judgement, but because I see how much work it can take away that does not necessarily require me to spend my own time doing it. It can compare documents, structure information, draft, search through large amounts of material, identify inconsistencies, challenge an argument, support a risk assessment and turn unstructured thoughts into something useful. That creates enormous potential.

In our User Perspective paper, we already wrote about an 80/20 principle: use systems and technology to take care of as much of the repeatable work as possible, so that people can focus on the part where their expertise genuinely matters. AI takes that idea much further. If AI can do 80% of a process that previously required a senior professional, the senior can concentrate on the 20% that matters most: judgement, uncertainty, unusual situations, interfaces, consequences and decisions.

That can reduce cost and increase capacity. It can allow organisations to do more with the same people, and it can also allow more junior colleagues to perform work they might previously not have been able to perform at that level. All of that is positive.

But there is something in that argument that has started bothering me.

Output capability is not necessarily judgement capability

Someone with relatively little experience can now produce an impressively structured risk assessment. They can ask AI to draft a procedure, compare regulatory expectations, propose controls, analyse a deviation or even help build substantial parts of a quality management system. Some, very established, QMS platforms already include AI capabilities that can turn templates and prompts into remarkably mature-looking outputs.

Yet the ability to produce a mature-looking output is not necessarily the same as having the judgement to know whether that output makes sense.

That judgement is difficult to describe because much of it comes from experience. It comes from seeing a technically correct solution fail in reality, from discovering that the root cause everyone initially agreed on was not the root cause after all, from watching suppliers miss deadlines, equipment behave differently from its specification, people work around procedures and qualification plans fall apart when they meet actual operations. In radiopharma, it also comes from understanding that what looks like a relatively small technical decision can have a very real consequence for a patient at the end of the chain.

Those experiences change how you read information and which questions you ask. They teach you when something that looks perfectly reasonable on paper somehow does not feel right.

That is the 20%.

Today, we can often automate the 80% because experienced people are still there to assess it. The question is what happens next.

Who trains the next senior?

Many of today's senior professionals developed their judgement by doing work that AI may increasingly take over. They drafted documents, went through the data, prepared the first risk assessments, got things wrong, were corrected by someone more experienced, and sat in meetings where carefully prepared plans collided with operational reality. Over time, through repeated exposure to decisions, mistakes, consequences and corrections, they developed pattern recognition.

That does not mean juniors today should be forced to spend twenty years doing inefficient manual work simply because previous generations had to. Quite the opposite. If AI can remove work that adds little value, we should let it.

The more important question is this:

If we deliberately remove the old apprenticeship, what are we replacing it with?

Some of the work we now call inefficient was also where people learned. Not every spreadsheet needed to be built manually, not every procedure needed to be drafted from a blank page and not every regulatory requirement needed to be searched for the hard way, but somewhere inside all that work was repeated exposure to judgement.

If we automate the work without redesigning the learning, we may create something I would call competence debt.

Competence debt

Technical debt is familiar: we make choices today that allow us to move faster, knowing that somewhere in the future those choices may have to be addressed.

Competence debt may work in a similar way.

An organisation can reduce cost today because AI allows fewer people to perform more work. For years, everything may work perfectly well, especially while experienced professionals are still available to review the output and make the difficult calls. Then those people retire, and the organisation discovers that the people underneath them did not acquire the same depth of judgement.

That would not be because they are less intelligent, less motivated or somehow less willing to learn. They simply developed in a different environment: one in which the system increasingly produced the first answer for them.

At that point, saying “a human remains accountable” is no longer enough. We also need to ask which human will carry that accountability, and how that person became capable of doing so.

Human in the loop is not a development strategy

We hear the phrase “keep a human in the loop” frequently, and in many contexts I agree with it, particularly where decisions affect quality, safety, patients, regulatory compliance or significant business risk. Regulations are being built around this principle today, as I am writing this article.

But “human in the loop” describes a control. It does not describe how that human became competent.

Today we can place an experienced senior over an AI-supported process and say that professional judgement remains with the human. Perhaps we should design the same system with another person in mind as well: the senior professional of 2040.

Where will that person's judgement come from? If a career increasingly consists of reviewing sophisticated AI output, will that person develop the ability to recognise when the output is wrong, or even when the question itself was wrong?

That last point matters because it is one of the things I experience repeatedly when working with AI. The quality of the outcome depends enormously on the human using it. I challenge answers, ask different questions, ask for sources, disagree, test assumptions and sometimes realise that I asked the wrong question in the first place.

AI helps me enormously, but I also learned how to work with technology through very different stages of its development. I first taught myself programming from books and logic long before the internet was part of everyday life. By university, email was becoming normal; when I started my professional career in 2000, paper and fax were still part of routine business communication. Since then, I have watched digital tools become embedded in almost every part of working life.

That does not make the older ways better. It does mean that I learned repeatedly to understand a new tool rather than simply inherit a world in which it was already doing the work.

The challenge, therefore, is not to preserve the working world of the past. It is to make sure the next generation develops the same ability to question and judge without having to recreate it.

Perhaps AI can accelerate experience instead

There is another possibility, and it is the one I find much more interesting.

Maybe AI does not have to reduce professional development. Maybe we can use it to accelerate it.

A junior professional could potentially be exposed to far more scenarios than previous generations ever encountered personally. AI can help simulate cases, present alternative interpretations, challenge assumptions, compare strategies and explore what happens when one variable changes. It can allow someone to examine several possible solutions where previously there might only have been time to develop one.

That could be extraordinarily powerful, but only if we deliberately use it that way.

The learning cannot simply become: here is the finished answer.

The people working with the output need to understand why an approach was chosen, what alternatives were rejected, which assumption matters most, what evidence would change the conclusion, what could fail in the real world and where professional judgement entered the decision.

That also changes the role of experienced people. Their contribution should not end with reviewing whether an AI-supported deliverable is correct. They increasingly need to make their reasoning visible.

If I challenge a risk classification, I should explain why. If Wolfgang or I look at a quality or supply decision differently because of something we experienced fifteen years ago, that experience should not remain only in our heads. If we recognise a failure mode because we have seen it before, the junior colleague working with us should understand what we recognised and why.

The same applies when we work with customers. A good deliverable should not merely solve today's problem. Where possible, the people receiving it should also understand why the solution works. Otherwise we risk increasing output capability without increasing organisational capability.

There are therefore two kinds of efficiency

We usually think about execution efficiency: can we produce the required result with less time, cost and effort? AI can clearly improve that.

Perhaps we also need to think about learning efficiency: can the work we are doing develop the judgement of the people who will need to make the next decision?

The best AI-supported organisation may not be the one that removes the most humans from its processes. It may be the one that removes the least valuable human work while deliberately increasing the amount of meaningful human learning.

That takes us back to the wider trolley problem. We started with safety and quality, then added delivery. Now we can add cost and people.

Saving cost is not wrong. Increasing productivity is not wrong. Replacing unnecessary work is not wrong. But if today's cost saving creates tomorrow's competence debt, we have not removed the cost. We have postponed it, and perhaps made it much harder to see.

So what should we do with the 80%?

Perhaps the question is not whether AI will take over the work. It probably will take over a great deal of it, and in many cases it should.

The question is what we do with the human capacity that creates.

If we use it only to reduce headcount and cost, we may eventually discover that we also removed the pathway through which judgement was developed. But if we use AI to remove low-value work while deliberately exposing people to the reasoning, decisions, failures and trade-offs behind the work, perhaps the next generation does not need twenty years to become experienced.

Perhaps they can become better, faster.

But only if we teach the WHY, not just give them the answer.

So if AI does the 80%, what are we doing to make sure humans still learn the 20%?

What am I missing?

Related reading

Why are we asking someone to pull the lever? — Safety & Quality
If we do not release, that is only a business risk. — Delivery
The User Side of Radiopharmaceutical Facility Design — the 80/20 principle

Made with AI. Thought by me.