A polished draft can hide the learning problem

On one B2B marketing engagement, I reviewed AI-assisted drafts alongside the comments and revisions senior people were making to them.

The prose was rarely the main problem.

The same repairs kept appearing. The audience was too broad. Pieces began with the product rather than the reader's need. The value emerged too late. Evidence that should have shaped the argument had not been resolved before the writing started.

AI made those drafts cleaner.

It also made some of those weaknesses harder to see.

A first draft used to tell me quite a lot about the person who produced it. Research gaps appeared in the argument. Poor structure exposed weak prioritisation. Sometimes the writer was clearly trying to write around something they did not yet understand.

A polished draft never proved good judgement. But weak reasoning tended to leave visible marks.

That signal is getting harder to read.

An early-career marketer can now take a partially understood brief, spend a few minutes with a generative-AI tool and return with something coherent, structured and professionally written.

The difficulty for the manager is working out what has actually improved: the person's judgement, their ability to use the tool, or the finished artefact.

Those are different kinds of progress.

Performance can improve before judgement does

There is good evidence that less-experienced workers can benefit significantly from generative AI.

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of a generative-AI assistant to more than 5,000 customer-support agents. Productivity increased by about 15 per cent overall, with substantially larger gains among less-experienced and lower-skilled workers. The authors also found evidence that the system helped spread practices associated with stronger performers.[1]

That upside is real.

I see the attraction clearly in marketing and communications work. Someone who struggles to organise their thinking can get help creating a structure. A weaker writer can produce cleaner prose. A junior team member can explore alternatives much faster than they could from a blank page.

Strong early-career people can move especially quickly.

I would not want to take that away from them.

The management problem is that assisted performance and independent capability can develop at different speeds.

Research in education makes the distinction unusually visible. In one study of students using GPT-4 for mathematics practice, unrestricted access improved performance while the tool was available, but those students subsequently performed worse than the control group when they had to work without it. A version of the AI tutor designed with learning safeguards largely removed that effect.[2]

Marketing work is obviously different from a mathematics test. But the distinction matters.

The deliverable in front of the manager shows what the person and the tool produced together.

It does not reveal, by itself, which decisions the person could make again without the tool supplying the missing reasoning.

That matters because understanding a brief, finding useful evidence, deciding what matters, structuring an argument and reviewing the result are part of the capability we are trying to build.

AI can compress several of those stages into one interaction.

A marketer who has not properly understood the problem can receive a plausible interpretation of it. Someone who has not weighed the evidence can receive a confident recommendation.

The artefact improves while more of the reasoning disappears from view.

A better prompt still needs someone who understands the task

A lot of AI training focuses on prompts.

There is value in that. A clear instruction usually produces a better response than a vague one.

But the prompt sits downstream of more important decisions.

Before asking a model for an answer, someone needs to understand which decision the work supports, who needs to act, what is already known, what constraints matter and where the evidence is weak.

An experienced marketer may work through much of that almost automatically. An early-career marketer often cannot yet do so consistently.

AI creates an awkward shortcut here because it can help formulate the brief at the same time as it answers it.

That is efficient when the person can evaluate the assumptions being introduced. It is less useful for development when those assumptions pass unnoticed.

This is why I am less interested in whether a junior marketer knows an elaborate prompting formula than whether they can explain what they are asking the tool to solve.

A detailed prompt may contain good thinking.

It does not create that thinking by itself.

Look at what the senior reviewer is still repairing

The B2B engagement made this especially clear.

The senior reviewers were not spending most of their time correcting grammar or improving sentence flow. They were narrowing the audience, moving the message away from product description towards customer value, challenging weak evidence and correcting the argument itself.

Those are different interventions from polishing copy.

A manager under deadline pressure will often just make them.

I do it too.

The audience gets changed. The generic paragraphs disappear. The stronger value argument is inserted. The claim gets qualified. The document goes out.

The business has the deliverable it needed.

The developmental question is what happens next time.

If the person who produced the original draft cannot explain why the audience narrowed, why the argument changed or why one claim needed stronger evidence, the senior reviewer has repaired the output without transferring much of the judgement behind the repair.

Then the same lift returns a week later.

This is where AI can make the pattern stranger. It removes more mechanical effort from production while leaving a larger share of senior review concentrated on relevance, evidence and choice.

That senior attention is expensive.

Using some of it to make the decision visible can therefore matter more than another lesson in producing cleaner drafts.

The point does not require turning every review into a coaching session. Working businesses have deadlines. Clients need answers. Sometimes making the edit is the sensible thing to do.

I would focus the deeper discussion on decisions that recur.

When a manager has corrected the same audience problem several times, that is teachable. So is a repeated tendency to start with the product, accept generic evidence or avoid making a clear choice.

The aim is to spend senior attention where the lesson has a reasonable chance of transferring.

Some judgement needs exposure

There is also a limit to what review technique can solve.

A junior marketer may understand the brief and still make a weak recommendation because they have never spoken to the customer.

They may produce sensible messaging without understanding what Sales hears every week.

They may prioritise the wrong benefit because they have not yet seen which commercial trade-offs actually affect a buying decision.

That capability develops partly through exposure.

Customer conversations. Sales meetings. Failed campaigns. Difficult stakeholders. Post-mortems. Seeing what senior leaders approve, reject and question.

Over time, those experiences create patterns.

Managers can accelerate the process by explaining decisions and giving junior people access to more of the context around the task. They cannot completely substitute for the experience itself.

Neither can AI.

A model can describe a buyer objection convincingly. Hearing a real buyer make it, asking the next question and discovering that your original interpretation was wrong teaches something different.

The goal is not less AI

I would be concerned if an early-career marketer on my team avoided AI simply to prove they could work without it.

That is not the capability I need.

I want them to use the available tools well. They should move faster, explore more options and spend less time on work that does not require their judgement.

I also want them to know what the task is before they automate it.

They need to notice when an answer is merely plausible, recognise which evidence matters and know when they are missing context that the model cannot supply.

And when they bring work for review, I need enough of the reasoning to see what they are learning, rather than judging their development entirely from how polished the document looks.

Research on self-explanation has long suggested that learning improves when people have to explain the principles and reasoning behind a solution rather than merely inspect a worked answer.[3] The practical version at work can be modest. Ask the person to explain the important choice they made, the alternative they rejected and the part they are least certain about.

A polished draft can no longer be treated as sufficient evidence that the person behind it understands the work.

Managers need to inspect a little of the reasoning as well.

Done properly, that should make senior review less necessary over time, rather than turning permanent senior repair into part of the workflow.

Notes and sources

[1] Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work, The Quarterly Journal of Economics, Vol. 140, Issue 2, May 2025. Study of 5,172 customer-support agents; productivity gains were substantially larger among less-experienced and lower-skilled workers. https://www.gsb.stanford.edu/faculty-research/publications/generative-ai-work

[2] Hamsa Bastani et al., Generative AI without guardrails can harm learning: Evidence from high school mathematics, Proceedings of the National Academy of Sciences, 2025. The study distinguishes improved assisted performance from subsequent unaided performance and tests a safeguarded AI-tutor design. https://doi.org/10.1073/pnas.2422633122

[3] Alexander Renkl and Alexander Eitel, “Self-Explaining: Learning About Principles and Their Application”, in The Cambridge Handbook of Cognition and Education, 2019; see also Renkl et al., Learning from Worked-Out Examples: The Effects of Example Variability and Elicited Self-Explanations, Contemporary Educational Psychology, 1998. https://www.cambridge.org/core/books/abs/cambridge-handbook-of-cognition-and-education/selfexplaining/847D1A99961216B215E93C9762220FDE