Measuring applied AI capability in knowledge work

Define the capability in context

Choose a bounded workflow, intended output, audience, source environment, time limit, and decision consequence. Identify the AI actions available and the human judgments that remain. Break the task into observable capabilities instead of relying on a broad label such as “advanced user.”

Design a fair opportunity for each behavior to appear. If source verification matters, provide a material claim and an accessible authoritative source. If escalation matters, include a question whose authority is genuinely outside the participant's role. Missing evidence should be recorded as missing, not converted into poor performance.

Skills England's 2026 AI foundation skills for work describes basic technical, non-technical, and responsible or ethical skills for using simple AI tools at work. It is a foundation benchmark, not a validated measure of performance in every role or task.

Sources: [1]

Build the capability evidence grid

Create rows for task framing, source selection, AI instruction, iteration, output inspection, calculation or claim checking, uncertainty handling, judgment, escalation, and delivery. Record the opportunity, observable anchor, evidence captured, observed action, output consequence, missingness, reviewer note, confidence, and next evidence needed.

Separate process and result. A strong final document may conceal unsupported AI work; a visible correction may show valuable verification even when the first draft was wrong. Do not award credit for activity alone: many prompts, long sessions, or numerous citations do not establish whether the actions were appropriate.

Specify evidence capture before the task. Identify which interactions, source openings, revisions, calculations, and submitted artifacts are available to reviewers, which are not captured, and how participants are informed. A richer log can improve interpretation while also increasing privacy, access, and review obligations.

The UK government's 2026 employer guide to AI upskilling recommends practical and contextualized learning tied to real roles, tasks, decisions, responsible use, and reflection. It supports using work-relevant practice, while local transfer and outcome claims still require direct evidence.

Sources: [2]

Worked artifact: a synthetic operations brief

Imagine a synthetic operations analyst asked to recommend which service queue needs attention. The source pack includes a dashboard, a spreadsheet with one duplicated week, two manager notes, and an assistant. The capability grid gives opportunities to frame the service decision, inspect source dates, direct analysis, detect duplication, reconcile conflicting notes, and communicate uncertainty.

The participant asks the assistant for trends, then notices an implausible spike and rebuilds the weekly totals. They remove the duplicate rows, compare the manager notes with ticket timestamps, and narrow the recommendation to a two-week staffing experiment. They also label one category uncertain because its routing rule changed mid-period.

The grid records evidence for verification, synthesis, and communication but marks escalation as unobserved because the task did not require it. A second varied task would be needed before interpreting that behavior. This example is synthetic and is not a customer result, mature benchmark, or proof of general workplace performance.

Verify the measure before broad use

Use multiple tasks with different sources and failure patterns. Have reviewers apply prewritten anchors to the same sample, examine disagreements, and revise ambiguous criteria. Check whether access, instructions, time, assistive needs, or task design created unequal opportunities.

Recheck the construct when tools or work change. A task that once required explicit prompting may later be automated, shifting useful evidence toward problem framing, source selection, exception handling, and oversight. Preserve versioned interpretations rather than comparing scores across materially different opportunities without qualification.

Report pillar-level evidence, confidence, missingness, and follow-up questions rather than hiding them in one score. Two Prune provides task-bounded assessment evidence for human interpretation. It does not establish future performance or make automated employment decisions.