The AI skill gap is not prompting. It is judgement.

Prompting has become the poster child of AI adoption.

Prompt libraries. Prompt frameworks. Prompt hacks. Prompt engineering. Prompt cheat sheets.

The idea is everywhere. Learn better prompts and you will get better results.

There is truth in that. A clearer prompt can generate a better answer. A better instruction can produce a more useful draft. A stronger frame can help the model understand the task, the tone, the audience and the format.

Prompting matters. But it is not the whole game. In fact, the more AI improves, the less the real advantage will sit in prompting alone.

The bigger skill is judgement:

• Knowing what to ask.
• Knowing what to challenge.
• Knowing what to ignore.
• Knowing what to verify.
• Knowing what to rewrite.
• Knowing what to leave out.
• Knowing what not to delegate to AI at all.

That is the real AI skill gap.

Prompting gets you an answer. Judgement decides whether it is useful.

A prompt can get you output. It cannot tell you whether the output is right.

It cannot tell you whether the answer is relevant. It cannot tell you whether the tone fits the relationship. It cannot tell you whether the recommendation fits the customer. It cannot tell you whether the logic will survive a sceptical room. It cannot tell you whether the answer is too generic to be useful. It cannot tell you whether the facts are current, complete or defensible.

That is where judgement starts.

AI can produce a confident answer in seconds. That confidence is part of the power. It is also part of the risk.

A fluent answer can create the feeling of correctness. A clean structure can create the feeling of substance. A polished paragraph can create the feeling of insight.

But fluency is not truth. Structure is not strategy. Polish is not judgement. The human still has to decide what is worth using.

Prompting gets you output. Judgement gets you value.

The evidence does not support blind optimism

The evidence on generative AI at work is encouraging, but it is not simplistic. AI can improve productivity and quality in some tasks. But the gains depend on the task, the user and the quality of human oversight.

A Harvard Business School and Boston Consulting Group study tested GPT-4 with consultants working on realistic, knowledge-intensive tasks. The study found that consultants using AI completed 12.2% more tasks, completed them 25.1% faster and produced outputs rated more than 40% higher in quality when the tasks were within AI's capability range.[1]

That is a powerful finding. But it comes with an important warning. The same study introduced the idea of the "jagged technological frontier". AI's performance is uneven. It can be highly capable on some tasks, while failing on other tasks that may appear similar to humans.[1]

That means users need to know where AI helps and where it can mislead. This is precisely why judgement matters. If AI were reliable across every task, prompting would be enough. But it is not. The user has to understand the task, identify the risk, check the reasoning and decide whether the answer belongs inside or outside the AI frontier.

The workplace evidence is encouraging, but uneven

A separate study published in the Quarterly Journal of Economics looked at more than 5,000 customer-support agents using a generative AI assistant. The study found that access to AI increased productivity by 15% on average, measured by issues resolved per hour.[2]

The biggest gains were among less-experienced and lower-skilled workers. The study also found evidence that AI helped worker learning and improved English fluency, particularly for international agents.[2]

That is the positive version of the story. AI can raise the floor. It can help people who have less experience reach stronger output faster. It can support language quality, improve consistency and guide people towards better responses.

But the same study also showed variation across workers. More experienced and higher-skilled workers saw smaller gains, and the research noted small quality declines in some areas for the most experienced group.[2]

That matters. The tool alone is not the full story. How it is used matters. Where it is used matters. How much it is trusted matters. What the human does with the suggestion matters. That is judgement.

AI makes output abundant

AI has changed the supply of output. Anyone can now generate a strategy outline. Anyone can draft an email. Anyone can ask for a summary. Anyone can create a meeting agenda. Anyone can produce a presentation structure. Anyone can ask for three options, five risks and ten questions.

That is useful. It is also disruptive. When output becomes abundant, output becomes less impressive. The differentiator shifts. The value no longer sits only in producing something. It sits in producing something useful.

That requires filtering:

• Is this specific enough?
• Is this grounded in reality?
• Is the recommendation commercially sensible?
• Is the tone right for the stakeholder?
• Is the evidence credible?
• Is the assumption valid?
• Is the answer solving the real problem or only the prompt?

AI can help generate options. Humans still need to decide which option is worth defending.

The weak user and the strong user

The difference between weak and strong AI use is often visible in the second move. The first move is the prompt. The second move is the judgement.

A weak AI user asks: "Write me a strategy."

A stronger AI user asks: "Give me three strategic options for this market. Include the trade-offs, execution risks, key assumptions and evidence I would need to defend each option to an executive audience."

A weak AI user asks: "Write this email better."

A stronger AI user says: "Rewrite this email for a senior stakeholder. Keep it direct, constructive and commercially grounded. I need alignment without sounding defensive."

A weak AI user asks: "Summarise this document."

A stronger AI user asks: "Summarise this document, identify the claims that require verification, highlight any missing risks, and list the questions a sceptical board member would ask."

A weak AI user accepts the first answer. A stronger AI user interrogates it. They ask:

• What is this assuming?
• What is missing?
• Where could this be wrong?
• What would a sceptic say?
• What evidence would prove or disprove this?
• What would a customer, regulator, investor or employee hear differently?
• What should be removed because it sounds impressive but adds little?

That is the difference. The prompt starts the work. Judgement makes it valuable.

Prompting is a literacy skill. Judgement is a leadership skill.

Prompting is important because people need basic AI literacy. They need to know how to give context. They need to know how to define the task. They need to know how to specify audience, tone and format. They need to know how to iterate. They need to know how to ask for assumptions, risks and alternatives.

That should be taught. But organisations should not stop there. Prompting is the entry point. Judgement is the capability.

A team trained only on prompting may produce more content. A team trained in judgement can produce better decisions. That is the distinction leaders should care about.

AI training should not only cover how to get an answer. It should cover how to evaluate an answer:

• How do we verify claims?
• How do we identify hallucinations?
• How do we test assumptions?
• How do we protect sensitive information?
• How do we adapt outputs to customer reality?
• How do we know when the work needs a human conversation instead?
• How do we avoid turning every task into a generic AI-shaped output?

Those are not prompt questions. They are operating questions.

The danger of "AI-shaped" work

As more people use similar tools, a new risk emerges. Work can start to look the same. The same structure. The same cautious phrases. The same generic recommendations. The same polished but empty transitions. The same bullet-heavy logic. The same confident framing without enough substance.

This is the danger of shallow AI use. It does not produce bad work in an obvious way. It produces work that is acceptable, polished and forgettable.

That can be worse. Bad work gets challenged quickly. Polished average work can travel further before anyone notices that it does not contain much original thinking.

This is why judgement becomes more important as AI becomes more common. The better user does not only ask AI to improve the text. They ask whether the work is worth saying at all.

The best AI users are better filters

The strongest AI users are not necessarily the people with the longest prompts. They are the people with the strongest filter.

They know when the answer is generic. They know when context is missing. They know when a recommendation does not fit the customer. They know when a paragraph sounds clever but avoids the real issue. They know when a summary is too clean for a messy problem. They know when a "balanced" answer is avoiding a decision. They know when the model is producing a safe middle rather than a useful view.

That is a human skill. It comes from experience, curiosity, commercial context and the courage to challenge fluent output.

AI can generate. The human must discriminate. That word matters. Discrimination, in this context, is not bias. It is discernment: the ability to separate useful from useless, specific from generic, supported from unsupported, relevant from decorative, sharp from soft, true from merely plausible.

That is the skill gap.

What leaders should measure

If leaders want better AI adoption, they should be careful about what they measure.

Do not measure success only by usage. Usage is not value. A team can use AI heavily and still produce shallow work.

Do not measure success only by output volume. More documents, more decks and more emails do not automatically mean better execution.

Do not measure success only by speed. Speed is useful, but faster poor judgement is still poor judgement.

Better measures are different:

• Are decisions improving?
• Are customer responses clearer?
• Are assumptions being surfaced earlier?
• Are teams verifying more effectively?
• Are junior employees learning faster?
• Are managers spending less time correcting basics and more time coaching judgement?
• Are outputs more specific, more grounded and more useful?

That is where the AI return shows up. Not in more work. In better movement.

What organisations should teach

Organisations should teach prompting. But they should also teach the judgement layer around prompting.

Teach people to ask AI for alternatives, not only answers. Teach them to ask for assumptions. Teach them to ask for counterarguments. Teach them to request evidence requirements. Teach them to identify uncertainty. Teach them to separate fact from interpretation. Teach them to mark what must be verified. Teach them to use AI as a thinking partner, not an authority.

A useful internal standard might be:

Never submit AI-assisted work unless you can explain what you changed, what you verified and what you chose not to use.

That standard is simple. It turns AI from a production shortcut into a thinking tool.

The human premium

The more AI improves, the more certain human skills matter: context, taste, timing, ethics, accountability, customer understanding, commercial judgement, relationship awareness, strategic courage.

These are not abstract soft skills. They are the difference between output and value.

AI can help a person draft a customer message. But it cannot fully understand the trust that has been built over five years. AI can summarise a market. But it cannot know which signal matters most to the team's current strategic moment unless the human provides that context. AI can generate options. But it cannot own the trade-off.

The human still has to decide.

The question

So here is the question. Where do you think the real AI skill sits: in prompting, or in judgement? And if you had to choose only one capability to strengthen in your team, which would it be?

Prompting gets you started. Judgement makes it worth using. That is where the real AI advantage is moving.

---

Sources

[1] Fabrizio Dell'Acqua et al., "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality", Harvard Business School working paper with Boston Consulting Group, 2023. [2] Erik Brynjolfsson, Danielle Li and Lindsey Raymond, "Generative AI at Work", Quarterly Journal of Economics, 2025. [3] Shakked Noy and Whitney Zhang, "Experimental evidence on the productivity effects of generative artificial intelligence", Science, 2023. [4] Microsoft and LinkedIn, "2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part", 2024.