Welcome Back to XcessAI

Artificial intelligence has produced a remarkably consistent result across professional work.

Give people AI and their output gets better. They write faster. Produce better documents. Solve more problems. Complete work that previously required more experienced colleagues.

For companies, this sounds almost entirely positive. But a fascinating new experiment with patent lawyers raises a more uncomfortable question.

What if better work today creates weaker professionals tomorrow?

Researchers gave 133 practising patent lawyers across eleven US intellectual-property firms access to an AI drafting assistant for three months.

The immediate result looked exactly as you might expect.

After 90 days, lawyers using AI produced patent drafts that scored 0.38 standard deviations higher than those produced by the control group. Junior lawyers benefited particularly strongly.

Then the researchers did something that most workplace AI studies don't. They took the AI away.

Everyone was asked to review and correct a flawed patent application without assistance, a task designed to test the professional judgment lawyers had developed during the previous three months.

The result changed completely.

Senior lawyers who had used AI performed substantially better than senior lawyers who had not.

Junior lawyers showed no average improvement.

The people who appeared to benefit most from AI while using it were the people who retained the least once it disappeared.

That creates a paradox with implications far beyond law.

AI can improve the quality of the work. But the work we are automating may also be the process through which people learn how to become experts.

Consider what happens when a junior lawyer drafts a patent.

The obvious output is the patent. Management can observe it. A client can pay for it. The firm can measure how long it took.

But another output is being produced at the same time.

The lawyer.

The junior encounters awkward cases. Makes mistakes. Receives corrections. Learns which details matter. Begins recognising patterns. Gradually discovers that rules which looked simple in a textbook become ambiguous in the real world.

After thousands of repetitions, something difficult to describe begins to emerge.

Judgment.

Traditional professional organisations have always produced these two outputs simultaneously:

Today's work + tomorrow's expert

AI changes that equation because many of the tasks easiest to automate are precisely the repetitive tasks historically given to juniors.

And those tasks may look low-value only if we measure their immediate output.

Their second function is training.

That is what makes the patent study particularly interesting.

Most AI productivity experiments measure people while AI is available.

That answers an important question:

Does AI help someone perform the task?

This experiment asked another:

Does working with AI help someone become better at the task?

Those are not the same thing.

After ten days, AI-assisted lawyers' drafting scores were 0.34 standard deviations higher. After 90 days, the gap had risen slightly to 0.38.

The researchers translate that latter improvement into roughly an 11-percentile-point increase relative to the control distribution.

The improvement was especially noticeable among junior lawyers, who also completed an early drafting exercise faster.

So if the experiment had stopped there, the conclusion would have been straightforward.

AI works.

Juniors benefit most.

Deploy it.

But after three months, participants were given the unaided redlining exercise.

Across the whole group, people previously exposed to AI still performed somewhat better.

Look beneath the average, however, and something surprising appears.

The improvement came entirely from senior lawyers.

Seniors with AI exposure outperformed their senior control group by 0.45 standard deviations.

Among juniors, there was no statistically discernible average improvement. Instead, their results became more polarised: fewer middling performances, but more poor ones as well as more good ones.

AI had clearly helped juniors produce better work.

It had not clearly helped the average junior build better independent judgment.

Imagine the traditional professional career as a ladder.

At the bottom are relatively repetitive tasks.

Research. Drafting. Checking. Reconciling. Building models. Reviewing documents. Writing first versions.

Above them sit progressively harder tasks involving interpretation, prioritisation, client management and judgment.

For decades, we have looked at the bottom rungs and thought:

Surely we can automate these.

And technically, we increasingly can.

But perhaps we misunderstood what the rung was for.

It wasn't merely there to get someone higher.

Standing on it was part of learning how to climb.

Remove enough of the routine work and something strange could happen.

We may create organisations with extraordinarily productive senior professionals and AI systems capable of performing much of the junior work.

But where do the next generation of senior professionals come from?

The study offers a clue about why experience matters.

Junior lawyers tended to approach the unaided redlining exercise mechanically. They often worked sequentially through the document, spent substantial effort correcting lower-value prose and sometimes identified important problems without actually restructuring the patent to solve them.

Experienced lawyers behaved differently.

They focused on the commercially important claims, rebuilt sections more aggressively and connected their edits to underlying legal principles.

In follow-up interviews, senior lawyers described using AI less like an answer machine and more like something they could interrogate and critique. Google Research describes one interpretation as treating AI as a kind of "logic auditor."

That distinction is important.

An expert already possesses a mental model.

AI gives the expert something against which to test that model.

The expert can ask:

Why did the AI do this? What has it missed? Is this assumption correct? Would I structure the problem differently?

For the junior, the same AI output can serve a different function.

It can simply provide the answer.

The technology is identical. The learning environment isn't.

This suggests a broader principle:

AI may amplify expertise more easily than it creates expertise.

This problem should feel familiar far beyond law.

Consider investment banking. A junior analyst historically spends countless hours building financial models, reconciling accounts and assembling presentations.

Much of that work is repetitive. AI will increasingly be capable of doing it faster.

Now imagine an analyst entering the industry in 2030.

The AI imports the financial statements. Builds the model. Normalises the numbers. Calculates valuation. Creates the charts. Drafts the investment case.

The junior becomes dramatically more productive.

But five years later, when something unusual happens inside the accounts, what exactly have they learned to recognise?

The same question applies to consulting associates who no longer build analyses from scratch.

Doctors whose diagnostic systems identify patterns before they do.

Software developers who increasingly review generated code rather than struggle through writing it.

Accountants who no longer perform the reconciliations underneath the numbers they review.

In each case, automation may eliminate work that appears inefficient while simultaneously eliminating repetitions through which intuition was built.

That doesn't mean those professions should reject AI.

It means productivity and capability development need to be measured separately.

This creates a subtle management problem.

Most organisations will evaluate AI using today's metrics.

All sensible. But imagine that an AI system allows ten juniors to produce the work previously requiring fifteen.

The business case looks excellent.

Five years later, however, the organisation discovers that fewer of those juniors developed the judgment required to become strong managers, partners or senior specialists.

That cost does not appear in the original ROI calculation.

It appears years later as a talent problem.

We might call this human-capital depreciation.

The organisation improved the efficiency with which it produced work while reducing the amount of learning embedded in producing that work.

There is an uncomfortable irony here.

The junior tasks businesses are most eager to automate are often the tasks executives themselves performed earlier in their careers.

And many of the executives now saying that those tasks are unnecessary may have developed their own judgment by doing them.

There is a more optimistic interpretation of the study.

Senior lawyers didn't merely avoid losing expertise.

They appeared to learn more when given AI.

That suggests the problem is not AI itself.

It is how AI is used.

The senior lawyers already had enough foundational knowledge to argue with the system, identify weaknesses and understand why an alternative was better.

The challenge, therefore, may be designing junior workflows that preserve that cognitive engagement rather than bypassing it.

Perhaps AI should sometimes critique a junior's draft instead of writing it.

Perhaps the junior should produce an answer before seeing the machine's.

Perhaps training systems should deliberately remove AI for certain exercises.

Perhaps managers need to evaluate whether employees can explain why an AI-generated answer is correct rather than merely whether the final answer looks good.

We don't yet know the optimal design.

And this study certainly doesn't establish it.

It involved one specialised profession, 133 participants and only three months. The researchers themselves emphasise those limitations, including the fact that professional expertise normally develops over years rather than months.

The result should therefore be treated as a warning signal, not a universal law.

But it is a warning signal worth taking seriously.

For most of the AI era, we have asked which tasks machines can perform.

Perhaps companies now need a second question.

What function did that task perform inside the organisation?

Some work exists mainly because the work needs to be done.

Automate it.

But other work has always had two functions.

It generates an output and develops the person producing it.

Those tasks deserve more thought.

Because a company is not merely a system for producing today's output.

It is also a system for creating the people capable of producing tomorrow's judgment.

That is particularly important in professions where senior expertise cannot simply be taught from a manual.

Judgment comes from seeing hundreds of cases.

Getting things wrong.

Receiving feedback.

Recognising patterns.

Learning which details matter and which don't.

There may be no shortcut around all of those repetitions.

AI will almost certainly make junior professionals more productive.

The patent lawyers in this experiment already demonstrate that.

The more important question is what happens next.

If AI performs the mundane analysis, prepares the first draft, finds the errors and proposes the answer, the junior may reach the output much faster.

But perhaps they arrive without travelling the road that taught previous generations why the answer was right.

For companies, that creates a new optimisation problem.

We are no longer simply deciding:

Human or AI?

We are deciding which work AI should perform, which work humans should still experience, and how the two should interact so that better tools create better professionals rather than merely better outputs.

Because the most dangerous consequence of automating junior work may not be that junior jobs disappear.

It may be that the jobs remain, the output improves, everyone looks more productive... and one day we discover that fewer juniors learned how to become seniors.

Until next time,
Stay adaptive. Stay strategic.
And keep exploring the frontier of AI.

Fabio Lopes
XcessAI

💡Next week: I’m breaking down one of the most misunderstood AI shifts happening right now. Stay tuned. Subscribe above.

Read our previous episodes online!