Sim Chen Xing

Artificial intelligence is increasingly presented as a means of accelerating work. Generative AI can assist with drafting, analysis, coding, research, information synthesis and idea development at speeds that would have been difficult to imagine only a few years ago. Organisations are therefore understandably investing in AI platforms, experimentation and workforce capability, often under the broader objective of developing “AI literacy”.

AI literacy is important. Employees need to understand how to interact with AI systems, frame effective instructions, provide appropriate context, interrogate outputs and recognise some of the technology’s limitations. Yet this emphasis may also encourage an overly narrow understanding of what effective AI use requires. It can suggest that the principal challenge lies in learning how to operate the tool well.

The challenge is considerably broader.

AI does not enter a vacuum. It enters existing domains of knowledge, organisational processes, datasets, systems of governance, institutional practices and relationships of accountability. The quality of AI-enabled work therefore depends not only on what the technology can produce, but on whether the people using it understand the context within which that output is generated and applied.

A user may be highly proficient at prompting and still define the wrong problem. An AI system may execute an analysis competently even though the available data are unsuitable for the inference being made. A workflow may be automated efficiently while overlooking a control that exists for regulatory or organisational reasons. A polished answer may appear reliable despite being based on incomplete context. In each case, the difficulty lies not simply in the performance of the AI system, but in the surrounding judgement required to determine whether its use is appropriate and whether its output can be trusted.

This suggests that AI-enabled work requires a broader set of capabilities than AI literacy alone. These include problem and domain literacy, process literacy, data and methodological literacy, technological literacy, governance and ethical literacy, institutional literacy, and an understanding of the people ultimately affected by AI-mediated work.

These literacies perform different functions. Some help us determine what problem should be solved. Others help us assess the evidence available, select appropriate methods, understand the limits of the technology, establish the boundaries of acceptable use, or evaluate the consequences of an AI-enabled decision. Together, they provide the intellectual and organisational infrastructure within which AI can be used responsibly.

This matters because generative AI substantially lowers the effort required to produce an output. Tasks that once exposed uncertainty through time, technical difficulty or the need for specialist assistance can increasingly be completed through a relatively simple interaction. Yet reducing the effort required for execution does not remove the reasoning required to establish whether the execution is valid.

Indeed, the more easily an answer can be produced, the more important it becomes to understand the conditions under which that answer should be accepted.

The central question of AI adoption, therefore, should not be limited to whether employees know how to use AI. It should also ask whether they possess the wider literacies necessary to define the work, evaluate the evidence, select appropriate methods, recognise technological limitations, operate within organisational boundaries and remain accountable for the result.

The sections that follow examine these literacies in turn and consider why each becomes more, rather than less, important as AI becomes embedded in everyday work.

Problem and Domain Literacy: Defining the Work Before Delegating It

Before AI can help solve a problem, someone must first understand what the problem actually is.

This may sound self-evident, but much organisational work depends on tacit knowledge. Requests such as “improve this process”, “analyse these results”, “draft a response” or “find a solution” may appear sufficiently clear when communicated between colleagues who share an understanding of the surrounding context. The same request given to an AI system may conceal substantial ambiguity.

The quality of AI output therefore depends in part on the quality of the problem definition that precedes it. Yet problem definition is not merely a prompting skill. It is a domain skill.

A person working in taxation, for example, needs to understand the legal and policy context surrounding an issue before deciding what information is relevant. Someone working in communications needs to understand the audience, institutional position, and objective of the message. A researcher needs to understand the conceptual or theoretical question being investigated before deciding what evidence would meaningfully address it. An operations manager needs to know which constraints are fundamental to a service and which are simply historical features of the way work has traditionally been performed.

Without such knowledge, AI can create the appearance of progress without necessarily improving the quality of the underlying work. A well-written output may seem complete because it is coherent, organised, and responsive to the wording of the prompt. Yet coherence is not the same as correctness, and responsiveness to a prompt is not the same as responsiveness to the underlying problem.

This distinction becomes more consequential when AI is used by people entering unfamiliar domains. Generative AI lowers barriers to specialised language and technical concepts. A user can request an explanation of a legal principle, generate code in a programming language they have never formally studied, or ask for a statistical analysis they would not previously have known how to perform.

This can be highly valuable. AI can facilitate learning, exploration, and interdisciplinary work. But it can also create an important asymmetry: the ability to produce sophisticated-looking work may develop more quickly than the ability to evaluate whether that work is substantively appropriate.

AI can expand the range of possibilities a person is able to consider. It does not remove the need for sufficient domain knowledge to distinguish a plausible suggestion from a defensible one.

Problem and domain literacy therefore concern the intellectual work that precedes the AI interaction: What are we actually trying to achieve? Why does the problem matter? Who defines success? What constraints are non-negotiable? What assumptions are being made? What would count as evidence that the problem has been adequately addressed?

These questions are not merely preparation for the work. They are part of the work itself.

Process Literacy: Understanding How Work Produces Outcomes

Once the objective is understood, the next requirement is an understanding of how work is translated into organisational outcomes.

Processes are often described in terms of visible steps: information is submitted, checked, approved, acted upon, and recorded. AI naturally invites questions about which of those steps can be automated, shortened, or removed.

Such questions are legitimate, but process improvement becomes risky when efficiency is pursued without understanding the function that each step serves.

A step that appears redundant may operate as a control. A delay may be caused by a necessary approval rather than administrative inefficiency. Two apparently duplicated checks may protect against different risks. An escalation pathway may seem unnecessary in routine cases but become essential when unusual circumstances arise.

If an AI system is asked to optimise a workflow without sufficient context, it may reasonably infer that fewer steps, reduced manual intervention and faster completion constitute improvement. In many situations they do. However, organisations may also value auditability, segregation of duties, fairness, accessibility, legal compliance, quality assurance and the ability to recognise exceptional cases.

Process literacy therefore requires an understanding not only of what happens, but of why it happens.

This becomes more important as AI moves from assisting with individual tasks to coordinating multiple stages of work. An AI agent that retrieves information, drafts a response, recommends an action and triggers a subsequent process is not simply supporting the workflow. It is participating in the execution of that workflow.

At that point, tacit organisational knowledge becomes a significant constraint. Experienced employees often know when a seemingly standard case requires escalation, when a formal procedure is no longer current, or when two written instructions conflict and professional judgement is required. Much of this knowledge may never have been documented explicitly.

An AI system cannot reliably reproduce organisational judgement that the organisation itself has not made legible.

AI adoption may therefore require organisations to examine their processes more critically than conventional automation sometimes has. Before asking how AI can execute a process, an organisation may first need to establish whether the process is coherent, current, documented, and understood.

In this respect, AI often reveals weaknesses that existed long before the technology was introduced. Conflicting procedures, ambiguous ownership, obsolete controls, and undocumented exceptions are not created by AI. However, AI can magnify their consequences by reproducing a flawed interpretation rapidly and at scale.

The relevant objective is therefore not the indiscriminate automation of existing work, but the careful understanding of how work produces the outcome an organisation intends.

Data Literacy: Understanding the Evidence Before Analysing It

Few areas illustrate the limits of AI-enabled acceleration as clearly as data analysis.

Modern AI systems can perform tasks that once required considerable technical knowledge. They can write SQL queries, generate Python or R scripts, produce visualisations, identify potential patterns and explain statistical results in accessible language. Employees who previously depended on specialist analytical support may now be able to explore datasets independently.

This represents a significant expansion of analytical capability. It also makes data literacy more important.

A common temptation is to treat AI as though the analytical process begins with uploading a dataset and asking the system to “find insights”. Yet analysis does not begin with computation. It begins with an understanding of the evidence.

Before examining relationships or producing visualisations, the analyst needs to know what the data represent. How were they collected? For what original purpose? Which variables are complete and which are not? Have definitions changed over time? Are there duplicates, missing values or measurement errors? Are particular categories meaningful, or are they artefacts of an administrative system?

These questions matter because the mere existence of data does not mean that the data are suitable for the question being asked.

A dataset created for an operational purpose may not support a research inference. A field that appears numerical may encode categories rather than quantities. Missing values may reflect systematic patterns rather than random omissions. Apparent changes over time may result from revised definitions rather than genuine changes in behaviour.

An AI system may be able to process all of these data successfully while remaining unaware of the institutional history that determines how they should be interpreted.

Data literacy therefore includes an understanding of provenance: where the data came from, how they were produced, what they measure and what limitations accompany them. It also includes the ability to distinguish between available data and appropriate evidence.

This distinction becomes especially important because AI can produce polished outputs from poor inputs. Tables, charts and numerical summaries may appear authoritative even when the underlying dataset is unsuitable for the intended purpose.

The quality of presentation should therefore never be treated as evidence of the quality of the underlying data.

Analytical and Methodological Literacy: Choosing an Appropriate Way to Answer the Question

Understanding the data is necessary, but it is not sufficient. The analyst must also understand how the available evidence can legitimately be used to answer the question.

AI has made analytical execution substantially easier. A user can request a regression, clustering analysis, forecast or classification model, and an AI system may generate the relevant code almost immediately. It may also explain the method and interpret the output.

Yet the ability to execute a technique does not establish that the technique should have been used.

Suppose an organisation wants to understand whether participation in a programme improved a particular outcome. It may be technically straightforward to compare participants and non-participants and ask AI to conduct a statistical test. But if participation was voluntary, the two groups may already differ systematically. If the outcome was measured differently across groups, the comparison may be misleading. If the dataset includes only those who completed the programme, the analysis may exclude precisely those cases most relevant to understanding effectiveness.

The computational task may therefore be simple while the inferential problem remains difficult.

This applies widely. Correlation does not establish causation. A forecast based on historical trends may become unreliable when underlying conditions change. Statistical significance may have little practical significance. A model that performs well on average may perform poorly for specific subgroups. A compelling visualisation may conceal denominator problems, selection effects or inappropriate comparisons.

AI can often explain these limitations when prompted. The deeper problem is that the user must know that such questions need to be asked.

Methodological literacy therefore cannot be delegated entirely to the model. If users do not understand the assumptions on which a method depends, they may not recognise when those assumptions are violated. If they do not understand the distinction between prediction and explanation, they may draw causal conclusions from a predictive model. If they do not understand the data-generating process, they may attribute significance to what is merely an artefact of collection or measurement.

A robust analytical workflow should therefore preserve the distinction between different stages of reasoning: defining the question, assessing the available data, selecting an appropriate analytical approach, preparing and cleaning the dataset, generating or reviewing the necessary code, executing the analysis, examining diagnostic outputs, interpreting the results, and finally verifying that the conclusions are supported by both the chosen method and the underlying evidence.

AI can assist at each stage. It should not collapse them into a single instruction to “analyse the data”.

Where AI generates code, the user should still establish what the code actually does. Important calculations should, where appropriate, be checked using alternative methods or validated against samples of the source data. Unexpected findings should trigger investigation rather than immediate interpretation.

AI substantially reduces the effort required to execute analysis. It does not reduce the methodological standards required to draw a defensible conclusion.

In this sense, AI democratises execution more readily than it democratises judgement.

Technological Literacy: Understanding the Nature and Limits of the System

Effective use of AI also depends on an adequate understanding of the technology itself.

Generative AI systems are especially persuasive because their outputs resemble human communication. They respond conversationally, synthesise information fluently, and often present explanations in language that appears deliberate and authoritative. This can encourage users to treat them as though they were conventional information systems.

They are not.

A traditional database query retrieves information according to defined rules. A calculator executes a specified mathematical operation. A generative AI system produces responses probabilistically, influenced by its training, instructions, available context, retrieved information and access to external tools.

This difference has practical consequences.

An AI system may generate information that is plausible but incorrect. It may misunderstand an ambiguous instruction, omit a relevant qualification or produce an answer based on incomplete context. It may retrieve the wrong document, fail to retrieve the right one, or present uncertainty with greater confidence than the evidence warrants.

Context limitations are particularly significant in organisational settings. AI systems can only reason over information that is available to them in the relevant interaction or retrieval environment. An organisation may possess the correct policy, but an AI assistant cannot apply it if the policy is not retrieved, is outdated, or exists in a form the system cannot interpret adequately.

The conversational experience can obscure this distinction. A response based on complete evidence and a response based on incomplete evidence may both appear equally polished.

Tool selection therefore matters as well. Different systems may have different capabilities, context limits, retrieval architectures, data-handling arrangements and access to organisational information. The fact that one AI system is appropriate for a particular task does not imply that another system is equally suitable.

Technological literacy does not require every employee to understand machine-learning architecture in depth. It does require a sufficiently accurate mental model of what the system is doing and where reliability may break down.

Perhaps the most important distinction is between fluency and verification.

An AI response can sound certain without having established that its claims are true. It can generate a persuasive interpretation from inaccurate premises. It can write code that appears reasonable while containing a subtle logical error. It can summarise a policy while overlooking the provision that materially changes its application.

The appropriate response is not universal distrust, but calibrated trust.

The level of confidence placed in an AI output should depend on the nature of the task, the quality of the available evidence and the consequences of error. Low-stakes brainstorming can tolerate uncertainty that regulatory interpretation cannot. A preliminary draft can be treated differently from a financial calculation. A suggestion may require less scrutiny than a factual assertion intended to guide an external party.

Technological literacy therefore concerns not merely how to use an AI system, but how to judge the degree of confidence its output deserves.

Governance Literacy: Establishing the Boundaries of Appropriate Use

The fact that AI can perform a task does not mean that it should.

Governance literacy concerns the formal and organisational boundaries within which AI is used. It asks whether a particular task, dataset, decision or system configuration is appropriate in the first place.

Before organisational information is introduced into an AI system, users need to understand the nature of that information and the rules that govern its handling. Does it contain personal, confidential, commercially sensitive or privileged material? Is the selected tool approved for that type of information? Where is the information processed? Is it retained? Who may gain access to it? What contractual, legal or internal requirements apply?

These questions cannot be resolved through better prompting because they concern the conditions of use rather than the content of the output.

Governance also determines the appropriate boundary between assistance and decision-making. An organisation may permit AI to summarise applications but not make final eligibility decisions. It may allow AI to draft responses while requiring human approval before those responses are issued. Routine cases may be automated while higher-risk cases require mandatory escalation.

The appropriate boundary depends on the consequences of error, the rights of affected individuals, regulatory requirements and the organisation’s tolerance for risk.

This is especially important because AI can produce a subtle displacement of responsibility. When a recommendation originates from a system perceived as authoritative, employees may become more inclined to defer to it. The more embedded the system becomes in organisational processes, the easier it may be to treat its recommendations as though they carry independent institutional legitimacy.

They do not.

Responsibility remains with the organisation that chooses the system, determines where it is used, specifies what information it can access, establishes the controls around it and decides how its outputs are incorporated into work.

Good governance therefore does more than restrict AI use. It creates explicit conditions for appropriate use, allowing employees to experiment productively without forcing each individual to independently reconstruct the organisation’s risk boundaries.

Ethical Literacy: Evaluating Consequences Beyond Compliance

Governance establishes formal boundaries. Ethics asks whether those boundaries are sufficient.

An AI application may comply with policy and still produce undesirable consequences. A model may reproduce inequalities embedded in historical data. An automated service may technically provide correct information while making it more difficult for people with complex circumstances to obtain human assistance. A productivity system may improve measurable efficiency while reducing professional autonomy or encouraging staff to optimise for indicators that do not reflect the quality of the underlying work.

Ethical literacy therefore requires attention to the broader consequences of AI-mediated decisions, including consequences that may not be captured by technical performance measures.

Bias is one dimension, but the issue extends further. AI systems can influence which information receives attention, which options are considered and how problems are framed. If employees habitually consult AI before forming their own view, the system may gradually shape organisational reasoning even where it has no formal decision-making authority.

The ethical question is therefore not limited to whether a particular answer is accurate. It also concerns how repeated reliance on AI changes institutional behaviour and the distribution of human judgement.

Transparency is another consideration. When should users know that AI has materially contributed to an answer? When should an AI-generated recommendation be disclosed to a decision-maker? When is a human explanation necessary? How should people contest an outcome influenced by an automated system?

These questions do not admit a single universal answer. Their resolution depends on context, consequence and the relationship between the organisation and those affected by its decisions.

Ethical literacy is therefore the capacity to recognise when efficiency, accuracy and compliance are not the only values at stake.

Institutional Literacy: Making Organisational Knowledge Legible

As AI systems become more deeply embedded in operations, they also become dependent on the quality of the organisation’s own knowledge environment.

Organisations contain far more knowledge than their formal policy repositories suggest. Work is shaped not only by documented procedures but by conventions, professional judgement, historical decisions, escalation practices and relationships between teams.

An AI assistant cannot be expected to operate reliably within such an environment merely because it has access to a large number of documents.

Some documents may be outdated. Different versions may conflict. A circular may modify an earlier procedure without replacing it. A management decision may create an exception that becomes common practice but is never formally incorporated into the original policy. An experienced employee may know which source is authoritative, while the AI system sees several documents of apparently equal status.

The reliability of AI therefore depends partly on the coherence of the institutional knowledge on which it relies.

This creates an important reversal in the usual discussion of AI readiness. Organisations frequently ask whether employees are sufficiently prepared to use AI. They may also need to ask whether the organisation itself is sufficiently coherent for AI to interpret.

AI implementation can therefore expose weaknesses in knowledge management: unclear ownership of documents, inconsistent version control, conflicting procedures, weak metadata and tacit practices that have never been formalised.

Addressing these weaknesses has value beyond AI. Clarifying authoritative sources, resolving contradictions and documenting organisational knowledge can improve consistency and reduce dependency on informal transmission even in human-only workflows.

The effort required to make an organisation legible to AI may therefore also make the organisation more legible to itself.

Human and Stakeholder Literacy: Understanding the People Affected by the Output

AI-enabled work ultimately exists within a human context.

An organisation can define the problem correctly, design an efficient process, validate the data, select an appropriate method and govern the technology carefully, yet still produce a poor outcome if it does not understand the people affected by the resulting service or decision.

This becomes particularly visible in customer-facing applications.

A person seeking assistance from an organisation is unlikely to evaluate the sophistication of its underlying AI architecture. They are more likely to care whether the answer is relevant, comprehensible and trustworthy. They need to know what the information means for their situation and what they should do next. In difficult cases, they may need access to someone capable of exercising discretion.

AI can improve such interactions by increasing responsiveness, translating complex information and extending service availability. But it can also create frustration when users are trapped in automated pathways, repeatedly receive generic responses or cannot establish whether the information they have been given is authoritative.

Trust therefore becomes central.

Trust depends partly on accuracy, but it also depends on consistency, transparency, contestability and accountability. If an AI assistant gives one answer while an employee gives another, there must be a credible way to resolve the difference. If the AI cannot handle a particular case reliably, the system should not simply continue generating increasingly elaborate answers; it should recognise the need for escalation.

Organisations should therefore assess AI-enabled services not merely in terms of task completion, response time or technical accuracy, but in terms of whether the person affected by the process receives a satisfactory and defensible outcome.

Efficiency is an internal measure. Legitimacy is relational.

Verification as a Core Discipline of AI-Enabled Work

Across all these literacies runs a common requirement: verification.

Verification should not be understood simply as proofreading an AI-generated output once the system has completed its work. By that stage, significant errors may already have entered much earlier in the reasoning process.

Verification should instead be embedded throughout the chain of work.

• The question must be verified: are we solving the correct problem?
• The evidence must be verified: do the data, documents and assumptions accurately represent the situation?
• The method must be verified: is this an appropriate way to answer the question?
• The execution must be verified: did the code, workflow, or AI system actually perform what was intended?
• The interpretation must be verified: do the conclusions legitimately follow from the evidence?
• The final output must be verified: is it accurate, appropriate, compliant, and useful to the person who will rely on it?

The intensity of verification should vary according to consequence. It would be disproportionate to apply the same standard to generating ideas for a workshop as to calculating a financial figure, interpreting a regulatory requirement, or recommending an action that materially affects another person.

The principle, however, remains consistent: the ease with which an AI system can produce an output should not determine the confidence placed in that output.

Indeed, the speed and fluency of AI may make verification more important precisely because they reduce the friction that once exposed uncertainty. A task that previously required several intermediate steps may now appear to move directly from question to answer. Those intermediate steps have not necessarily become intellectually unnecessary; they have simply become less visible.

Responsible AI-enabled work therefore requires organisations to preserve the reasoning structure of a task even when AI reduces the effort required to execute it.

Conclusion: AI Changes Where Human Capability Matters

The rapid development of generative AI has understandably generated enthusiasm about productivity. Tasks that once demanded substantial time, specialist expertise, or technical skill can increasingly be supported through systems that are accessible to a much wider range of people.

This represents a significant expansion of human capability.

Yet its most important implication may not be that expertise matters less. It may be that expertise increasingly matters at different points in the work.

When AI can generate a first draft, human judgement becomes more important in deciding what should be written, for whom, and to what standard. When AI can generate analytical code, human expertise becomes more important in defining the question, evaluating the data, selecting the method, and determining whether the result is valid. When AI can retrieve policies and recommend operational actions, organisational capability becomes more dependent on whether those policies are coherent, authoritative, and appropriately governed.

AI therefore does not eliminate judgement. It redistributes where judgement is required.

This is why AI literacy, understood narrowly as the ability to operate an AI tool effectively, is insufficient as an organisational objective. AI-enabled work also depends on problem and domain literacy, process literacy, data and methodological literacy, technological understanding, governance and ethical judgement, institutional knowledge, and an understanding of the people ultimately affected by the work.

These literacies should not be treated as independent competencies. In practice, they interact. A data problem may also be a governance problem. A technically correct analysis may be methodologically inappropriate. An operationally efficient process may create an ethical concern. A customer-facing system may be accurate in most cases yet institutionally incapable of recognising when discretion is required.

The deeper AI becomes embedded in work, the more important it becomes to understand these interdependencies.

Perhaps the more consequential risk of AI is therefore not simply that it will occasionally produce an incorrect answer. Organisations have always made mistakes. The subtler risk is that the extraordinary ease with which AI generates plausible outputs encourages us to compress stages of reasoning that should remain distinct: defining the problem, establishing the evidence, choosing a method, executing the work, interpreting the result and deciding what should be done.

AI can assist at every one of these stages. It should not cause them to disappear.

The easier AI makes it to produce an answer, the more important it becomes to understand what makes an answer valid.

The future of AI-enabled work may therefore depend less on how quickly organisations place AI tools in the hands of employees than on whether they develop the wider capabilities that allow those employees to use AI with discernment.

The objective should not be to create workers who unquestioningly trust AI, nor workers who reject it out of caution. It should be to develop people who understand the work they are undertaking, understand the systems they are using, recognise the limitations of both, and know when verification, escalation, and human judgement remain necessary.

That is a considerably more demanding ambition than teaching people how to prompt. It is also a more meaningful conception of what it means to be literate in an AI-enabled world.