Foundations for AI-enabled knowledge work

Literacy, tool use and performance answer different questions

AI literacy describes knowledge, skills and attitudes for understanding and using AI critically. Tool use shows that a person interacted with an AI system. Applied performance concerns what happened in a particular task. These ideas overlap, but one should not be treated as automatic proof of another.

The OECD and European Commission describe AI literacy as knowledge, skills and attitudes that help people understand how AI works, critically evaluate outputs and use AI ethically and creatively. Skills England's workplace benchmark likewise combines technical, non-technical, responsible and ethical skills. Both sources make literacy broader than knowing a collection of prompt patterns.

In a workplace setting, activity data can answer whether a tool was opened, how often it was used or which feature appeared in a process. That information may be useful for adoption. It does not by itself show whether the worker selected the right evidence, noticed an incorrect assumption or delivered a sound recommendation.

Applied performance is therefore best described within a bounded task. A person may show strong verification in one familiar workflow and need more support in another. The evidence should stay attached to the conditions under which the work occurred.

Sources: [1] [2]

The unit of analysis is the human-AI work system

Useful evaluation looks beyond an AI model in isolation. The work system includes the person, objective, source material, AI tool, organizational constraints, review process and intended audience. Changing any of these can change both the quality of the result and the risks that need attention.

NIST's AI Risk Management Framework treats context as central to mapping and measuring AI risk. Its ARIA program extends that perspective by examining what happens when people interact with AI in realistic settings. That does not turn every workplace task into a formal AI-system evaluation, but it provides a sound reason to examine work in context.

Consider a manager preparing a supplier recommendation. A fluent-looking answer may still be weak if it uses the wrong reporting period, ignores a contract condition or repeats an unsupported vendor claim. The quality of the AI output alone cannot reveal whether the surrounding work was properly framed and checked.

A practical analysis therefore asks what decision was being made, which evidence mattered, what the AI was asked to do, which checks were performed and whether the final communication matched the audience's needs.

Sources: [3] [4]

Six capabilities organize observable AI-enabled work

Two Prune organizes task-bounded evidence into six connected capabilities: context framing, evidence navigation, AI orchestration, verification and risk control, judgment and synthesis, and communication and delivery. The structure describes aspects of observed work rather than permanent traits.

Context framing concerns the objective, audience, constraints and definition of success. Evidence navigation concerns finding, weighing and using relevant information. AI orchestration covers decomposition, prompting, iteration and the allocation of work between a person and an AI tool.

Verification and risk control concerns checks, uncertainty, sensitive inputs and operational consequences. Judgment and synthesis concerns trade-offs and the reasoning behind a recommendation. Communication and delivery concerns whether the final artifact is accurate, clear and usable by its intended audience.

The capabilities interact. Weak context can make technically competent prompting solve the wrong problem. Strong verification can reveal missing evidence and force a revised judgment. Clear communication can expose uncertainty rather than conceal it. A useful assessment keeps those relationships visible.

• Use knowledge measures when the question is what someone understands.
• Use activity measures when the question is adoption or frequency.
• Use realistic tasks when the question is how knowledge becomes work.
• Keep conclusions bounded by the task and available evidence.

Sources: [3] [2]

Good evidence connects claims to observable work

A defensible statement about AI-enabled work should identify the task, the behavior that was observable, the evidence supporting the interpretation and the limitations on that evidence. Missing evidence should reduce confidence rather than be silently converted into poor performance.

A final deliverable is important but incomplete. A polished document can hide weak source use, while a visible correction can show that a worker caught and repaired an AI error. Process evidence is also incomplete because an event log can show that an action occurred without fully revealing the person's reasoning.

The strongest interpretation combines the deliverable, relevant workflow evidence and the conditions of the task. It distinguishes direct observation from inference and states what could not be determined. NIST's framework similarly emphasizes context, documentation, measurement and ongoing review rather than a single decontextualized indicator.

This foundation matters for training, measurement and governance. Clear vocabulary prevents an organization from treating course completion as performance, usage as quality, or one simulation as proof of future results across every role.

Sources: [3] [4]