The Skill AI Can’t Generate: Why Discernment Is the New Data Literacy

For more than a decade, I have argued that the most valuable skill in a data-driven organisation is not access to information. It is the judgment to know when that information is wrong.

AI has made that skill more important, not less.

When I wrote about AI literacy in 2023, the pushback I heard most often was that the technology was not yet good enough for the question to matter. Now it is. AI can generate answers, summaries, recommendations, code, analysis, and increasingly, actions. The question for leaders is no longer what AI can do. It is what happens inside the organisation when AI produces something that looks authoritative, sounds confident, and is wrong.

Because it will.

The gap between organisations that catch those errors and organisations that do not is not a technology gap. It is a judgment gap.

We called it data literacy a decade ago. Before that, it was critical thinking. The label changes. The underlying capability does not. Every technology wave elevates a different human skill. The internet elevated knowledge access. Search elevated synthesis. AI has now automated synthesis. That means the next competitive edge belongs to the people and organisations that can evaluate what AI produces and make confident calls about what to trust, what to question, and what to discard.

That is discernment.

And it is not a capability that comes from using AI more. It comes from knowing a domain deeply enough to recognise when the answer is wrong, even when it looks right.

A quick word on terminology, because the market is increasingly calling this context. Context is what you give the AI: the data it can access, the business logic it understands, and the parameters that shape its outputs. Discernment is what you bring when the AI gets it wrong. One is a data problem. The other is a leadership one. Both matter.

The evidence is uncomfortable

BCG’s fourth annual Global AI at Work survey, published June 2026, found that 74% of frontline employees are now regular AI users. Of those, 41% report increased cognitive load alongside the productivity gains. That is the detail that tends to get buried under the headline numbers.

AI is not reducing the cognitive demand of work. For a significant proportion of the people using it regularly, it is increasing it, while simultaneously reducing the time available for the judgment that cognitive load requires. The same survey found that 72% of all respondents say expectations for the skills they need have shifted, yet only 36% feel they have received adequate upskilling. Adoption has outpaced direction, and discernment is the casualty.

The pattern holds across independent research. Findings by Dell’Acqua and colleagues, conducted with BCG and now formally published in Organisation Science, introduced the concept of the jagged technological frontier: the uneven boundary between tasks AI handles well and tasks where it fails or hallucinates. The danger is that professionals have no obvious signal telling them which side of the line a given task falls on. In the study, experienced knowledge workers with AI access performed worse on tasks outside the frontier, not because they were careless but because the output looked plausible. That is the discernment problem in its purest form. None of this is an argument against using AI. It is an argument for being deliberate about how AI is used, what standards apply, and what is expected from the people working alongside it.

The risk for leaders is not simply that teams use AI too much or too little. It is that they build AI-dependent workflows staffed by people who are no longer equipped to catch the errors. The output may be confident, well-presented, and wrong. At scale, that is not just an operational problem. It is a strategic liability.

Individual discernment does not scale. Decision architecture does.

Here is where most of the conversation about AI judgment stops, and where I think it needs to go further.

Even a leader who has personally developed strong discernment cannot be in every workflow. The organisations that get ahead of this are not the ones where every individual has excellent AI judgment. They are the ones that have built the institutional structures to act on good judgment at scale.

That requires decision architecture: clear ownership of AI-generated outputs, defined review thresholds based on the consequence of being wrong, and accountability structures that encode judgment at the process level rather than relying on individual vigilance. Finance functions spent decades building controls around spreadsheet outputs precisely because spreadsheets are authoritative-looking, easy to trust, and capable of propagating errors invisibly at scale. AI outputs have the same properties, but at greater speed and volume.

Most organisations have not built the equivalent controls. They are still at the stage of assuming that because someone reviewed the output, the output was reviewed well.

That is not a governance model. It is an optimistic assumption.

The board conversation has not (yet) caught up

Governance has not kept up with deployment. AI agents are now integrated into more workflows than most boards realise: booking vendors, triggering procurement, executing pricing decisions, drafting regulatory responses. With agents, the loop can close without a human in it. The governance question is no longer only about catching errors after they are made. It is about deciding, before deployment, where human judgment must remain in the loop when the consequences of being wrong are hard to reverse.

That is a fundamentally different conversation from the one most boards are having. Boards are asking about AI capability and regulatory risk in the abstract. The most important questions are operational and specific.

Where in our workflows is AI output being acted on without adequate human review? Which of those touchpoints are high-consequence and hard to reverse? Who owns the answer when it is wrong?

That is the process audit many organisations have not yet done. Not because leaders are complacent, but because the conversation has not been framed in operational terms. Once it is, the gap between where most governance architectures are today and where they need to be becomes clear very quickly.

Discernment fails when the data does

There is a dimension of this that sits below the model and below the process, and it is the one that will catch many organisations by surprise.

Discernment also operates at the data layer as much as the output layer. An AI system built on poor, ungoverned, or incomplete data will produce plausible-looking conclusions with complete confidence. The model cannot flag the gaps in what it was trained on or the data it is querying. It does not know what it does not know.

That is a human responsibility, which means it is a data architecture responsibility. Leaders who invest heavily in model capability without investing equivalently in making their data work for AI are building on unstable ground. The model may perform exactly as designed. The question is what it was designed on, what data it is being asked to act on, and whether the organisation has enough trust in that foundation to let AI outputs influence decisions.

Making data work for AI is not the same as making data available to AI. Availability is a pipeline problem. Working is a quality, governance, and trust problem. It means data that is current and not stale at the moment an agent queries it. It means data with clear provenance so that when a model produces an answer, someone can trace what it was trained on and verify whether that source is still valid. It means data with defined ownership so that when an agent acts on a wrong answer, the accountability chain is clear before the failure happens, not assembled after it. And it means data that has been structured with the AI use case in mind, not simply migrated from legacy systems that were never designed for machine consumption. Most organisations are doing the first of these. Very few are doing all four.

Getting that right is not just a technology decision. It is a leadership one.

Three things leaders should do now

The argument above is not abstract. It has direct operational implications. Here is where to start.

• First, audit the workflows where AI output becomes action. Map where AI-generated outputs are being used without structured human review. Prioritise by consequence, not volume. The most important workflows are not necessarily the ones using AI most often. They are the ones where a wrong answer triggers an action that is hard to reverse.

• Second, build decision architecture, not just review culture. Individual review is not a governance model. Leaders need clear ownership of AI-generated outputs, defined escalation thresholds, and accountability structures that do not depend on one person being vigilant at the right moment.

• Third, ask the agentic question before deployment. 
  For every agent, ask: what decisions can this agent make that a human cannot easily reverse? Who has authority to override it? What triggers human review? Those questions should be answered before the first failure, not after it.

The organisations that will be ahead in three years are not simply the ones with the most sophisticated AI or the fastest adoption curves. They will be the ones that know where judgment belongs, how to scale it, and what data foundation it depends on.

AI can generate the answer. It cannot decide whether the answer deserves to be trusted.

That is the human call. And it starts with discernment.