The easiest way to use AI at work is to give it a task that previously belonged to a person. Draft this email. Summarise this report. Analyse these comments. Write this code. Prepare these slides. Compare these proposals.
The productivity case is immediately apparent because the work that once took an hour might take minutes. Yet this way of thinking about AI also creates an important problem. If we define successful adoption by how much work can be handed over, it becomes easy to hand over parts of the work that were never simply production tasks in the first place.
A report is not only a collection of paragraphs. Someone decided which question the report should answer, which evidence was relevant, what should be excluded, how uncertainty should be treated and which conclusions the evidence justified. A communications plan contains decisions about audiences, objectives and priorities before anyone writes the copy. Software embodies assumptions about users, processes and acceptable failure. An economic analysis depends on choices about definitions, counterfactuals, methods and the interpretation of evidence.
AI can contribute to all of these activities, and increasingly it should. The important distinction is not between work humans should do and work machines should do, but between delegating cognitive labour and surrendering judgement.
Those are not the same thing.
Judgement starts before AI enters the task
Consider a manager who asks an AI system to develop a strategy for improving customer retention. The model can produce an impressive response containing customer segmentation, personalised communications, loyalty initiatives, predictive churn models and performance measures.
The obvious question is whether the recommendations are good.
The more important question comes earlier: was customer retention actually the right problem to solve?
Perhaps retention has deteriorated because the company deliberately acquired a different customer segment. Perhaps a product defect is driving cancellations. Perhaps profitability would improve by accepting higher churn among low-value customers. Perhaps the reported decline is simply a measurement artefact caused by a change in how customers are classified.
If the framing is wrong, a sophisticated answer can make the organisation more confidently wrong.
Human judgement therefore begins with deciding what deserves attention. AI can help interrogate that decision by identifying assumptions, testing alternative formulations and asking what evidence would distinguish between competing explanations, but someone remains responsible for deciding which problem the organisation is actually trying to solve.
That is not a limitation to be worked around. It is part of management.
Do not confuse speed with progress
AI makes producing things extraordinarily cheap.
We can generate more reports, more analysis, more marketing copy, more software, more presentations and more ideas than most organisations could reasonably consume. The constraint consequently begins to move away from production and towards selection.
Which analysis should be commissioned? Which idea deserves development? Which report needs to exist? Which customer interaction should be automated? Which piece of software creates enough value to justify maintaining it?
This changes the productivity problem.
When producing a first draft was expensive, reducing the cost of production generated an obvious benefit. When producing ten alternatives becomes almost free, generating another ten alternatives may add very little value. Someone still has to determine which one is appropriate, whether any of them solve the problem, and whether producing the artefact was worthwhile in the first place.
AI can increase output without increasing organisational performance if it simply allows everyone to produce more material for everyone else to process.
The scarce resource becomes judgement about what is worth doing.
Use AI before the answer, not only after the decision
Retaining judgement does not require restricting AI to administrative work.
In fact, using AI only after important decisions have already been made leaves much of its value unused. AI can be useful much earlier, when a problem is still being formulated.
Suppose an organisation is considering building a new customer application. AI can help identify assumptions about user behaviour, compare alternative service models, develop personas from research material, examine requirements, identify conflicting stakeholder objectives, simulate edge cases and question whether an application is the appropriate intervention at all.
In research, it can help interrogate a hypothesis, identify alternative mechanisms, critique a proposed method and suggest robustness checks. In communications, it can examine whether the proposed audience segmentation follows from the available evidence before drafting a single advertisement. In software development, it can challenge requirements and identify architectural risks before generating code.
Using AI strategically therefore does not mean asking it to make the strategy. It means allowing it to participate in the reasoning that precedes the strategy.
The human role shifts from supplying an instruction and checking the finished product to directing an analytical process.
Ask AI to disagree with you
One of the least useful relationships with AI is one in which the model continually confirms the user’s initial view.
Generative models are often highly responsive to framing. If you present an idea enthusiastically and ask how to implement it, the model can readily become an implementation consultant for the idea. If you describe the same proposal sceptically and ask why it might fail, it can produce an equally persuasive critique.
That flexibility is useful, but it means the first framing should not automatically determine the analysis.
If you believe a particular AI project will save money, ask the model to construct the strongest case for why it will not. If you favour one software architecture, ask what conditions would make an alternative preferable. If your analysis appears to show a relationship between two variables, ask for plausible confounders and alternative causal mechanisms.
Better still, ask what evidence would change the conclusion.
This turns AI from a confirmation mechanism into an adversarial analytical tool. The objective is not to manufacture disagreement for its own sake, but to expose assumptions before they become embedded in decisions.
Keep evidence separate from inference
Judgement becomes particularly important when AI moves from organising information to interpreting it.
Imagine providing an AI system with customer feedback and asking why satisfaction has declined. The model may identify recurring complaints about response times, product reliability and pricing. Those observations may be directly supported by the data.
It may then conclude that customers are leaving because competitors offer better service.
That could be true, but unless competitor information or customer switching data were supplied, the conclusion has moved beyond the evidence.
Generative AI can make these transitions almost invisibly because the prose connecting observation and interpretation is so fluent. A finding becomes an inference, the inference becomes an explanation, and the explanation becomes a recommendation without any obvious break in the narrative.
One of the most important human responsibilities is therefore to keep asking what kind of statement is being made.
Is this something we observed? Is it a calculation? Is it an inference from the evidence? Is it an assumption? Is it a hypothesis? Is it a recommendation based on a particular objective?
AI can help make these categories explicit, but the person responsible for the work should understand the difference.
Do not outsource objectives
AI can optimise only against some conception of what constitutes a good outcome.
This becomes important when apparently technical decisions contain competing values.
Suppose a recruitment system can reduce the average time required to shortlist candidates. That is an efficiency objective. The organisation may also care about identifying unconventional candidates, providing equitable access, maintaining transparency and giving hiring managers discretion.
A system optimised solely for speed might perform extremely well against its stated objective while making the recruitment process worse according to other criteria.
The same problem appears throughout organisations. A customer service system can minimise handling time by ending conversations quickly. A scheduling system can maximise labour utilisation while reducing employee flexibility. A marketing system can maximise engagement by producing material that damages trust. A predictive model can improve aggregate accuracy while performing poorly for a small but important population.
These are not primarily technical failures. They arise because objectives are incomplete.
Human judgement is required not simply at the end of the process to approve an AI recommendation, but at the beginning when deciding what the system should be trying to achieve and which constraints should apply.
Automation should not make responsibility disappear
As AI systems become capable of taking actions rather than merely producing recommendations, the relationship between automation and judgement becomes more important.
Consider the difference between an AI system that identifies potentially fraudulent transactions and one authorised to freeze customer accounts automatically. The underlying prediction might be identical, but the organisational consequences are not.
The question is no longer simply whether the model is accurate. It becomes necessary to consider the cost of false positives, the ability to reverse an action, the speed at which intervention is required, the vulnerability of particular customers and who is accountable when the system is wrong.
Not every action requires a person to approve it individually. That would defeat the purpose of many forms of automation. Human oversight can instead be designed into the system through thresholds, exceptions, sampling, monitoring, escalation rules and limits on what the AI is authorised to do.
The objective is not necessarily to keep a human in every loop. It is to retain meaningful human control over the loops that matter.
Expertise still matters, although its role changes
Generative AI creates an appealing possibility that anyone can perform specialist work simply by asking the right question.
There is some truth in this. AI substantially reduces the cost of accessing explanations, analytical techniques and technical capabilities that previously required specialist knowledge. A manager who cannot program can prototype software. A researcher with limited design experience can produce a credible visual concept. A small business can perform forms of analysis previously available mainly to larger organisations.
But access to production capability is not the same as access to judgement.
A person with little statistical knowledge can ask AI to run a regression, but may not recognise endogeneity, inappropriate functional form or selection bias. Someone unfamiliar with software engineering can generate working code without recognising security vulnerabilities or maintenance problems. A novice can produce a polished legal argument without knowing whether the authorities cited actually support it.
AI can lower the expertise required to begin a task much faster than it lowers the expertise required to evaluate the result.
This makes domain knowledge more important in some circumstances rather than less. Expertise increasingly determines how effectively a person can direct AI, identify weak assumptions, distinguish plausible output from defensible output and decide when additional specialist review is required.
Judgement does not mean checking every word
There is a risk of responding to these concerns by insisting that a human must carefully review everything AI produces. That sounds safe, but it is neither practical nor particularly meaningful.
If a person must reproduce every piece of work to establish that the AI completed it correctly, very little labour has actually been saved.
The better approach is to design verification around the structure and consequence of the task.
For document extraction, sample records and reconcile totals. For code, use tests, review security-sensitive components and monitor behaviour. For research, verify citations, inspect important claims and reproduce key calculations. For classification, measure performance against a labelled sample rather than reading every case. For automated decisions, monitor error rates, distributional effects and exceptions.
This moves oversight from vague human supervision towards quality assurance.
The distinction matters because effective use of AI will increasingly involve systems producing volumes of work that no person could realistically inspect line by line. Maintaining judgement therefore requires better methods of evaluation, not simply more manual checking.
Know when the human decision matters
Some decisions can be delegated almost completely because errors are inexpensive and reversible. Others warrant considerably more caution.
The relevant questions concern consequence, uncertainty and reversibility.
If AI selects the order in which internal documents appear in a search result, a poor choice is easily corrected. If it recommends which employee should be investigated for misconduct, the consequences of error are considerably greater. If an automated marketing system chooses an ineffective subject line, the cost may be modest. If a system determines eligibility for an essential service, the decision affects people’s lives.
The appropriate degree of human involvement should rise with the consequence of error, particularly where decisions affect rights, opportunities, safety, employment, access to services or significant financial outcomes.
This does not mean that humans are inherently better decision-makers. Human processes can be inconsistent, biased and poorly documented. AI may improve them. The important point is that introducing AI does not remove the need to decide how errors should be managed and who bears responsibility for them.
The strongest model is human plus machine, properly arranged
Debates about whether AI will replace human judgement often assume that the relevant comparison is human versus machine.
For most organisations, that is the wrong unit of analysis.
The useful comparison is between different ways of arranging people, models, data and processes.
A human might frame the problem, while AI identifies alternative explanations. AI might analyse thousands of records, while a specialist determines which patterns warrant investigation. A model might generate options, while a team evaluates them against organisational objectives. AI might execute routine decisions automatically while unusual or consequential cases are escalated.
The allocation can also change as evidence accumulates. A new system may begin with intensive human review, then move towards greater automation once its performance is understood. Areas where errors remain difficult to detect or particularly consequential can retain stronger controls.
The aim is not to preserve human activity for symbolic reasons. It is to allocate different forms of work according to their strengths, limitations and consequences.
Judgement is becoming more valuable, not less
As AI lowers the cost of producing analysis, code, text, images and recommendations, organisations will face an abundance of possible output. More ideas can be generated than can be implemented, more analysis produced than can be acted upon, and more content created than anyone has time to consume.
The economic value of production therefore begins to shift.
When generating an answer is expensive, the ability to produce one is scarce. When answers become abundant, the scarce capability is deciding which questions deserve attention, which evidence deserves confidence, which trade-offs matter and which actions should follow.
That is judgement.
Using AI well at work consequently requires neither resisting delegation nor handing over every task the technology can perform. It requires becoming more deliberate about where judgement sits within the process.
Let AI draft, calculate, classify, compare, criticise, simulate, search, code and propose. Let it participate before the decision as well as after it. Let it identify things that a person might miss and perform work that would otherwise be too expensive to undertake.
But retain clarity about who defines the objective, who decides what counts as evidence, who determines acceptable risk and who is responsible for the consequences.
The most capable AI system in an organisation should not make judgement disappear. It should give people better material on which to exercise it.