Measuring readiness for AI-enabled knowledge work

Define readiness for a bounded use case

Name the work, users, inputs, output, decision, and consequence. State the proposed AI contribution and human judgment point. A general writing assistant used for private ideation has different readiness requirements from a workflow that recommends customer action or changes a controlled record.

Avoid averaging unrelated dimensions into a reassuring score. A team may have strong technical access and weak source governance. Another may have capable reviewers but no approved data route. Report each condition and its evidence separately, including not tested and unavailable.

The NIST AI RMF Core organizes work through Govern, Map, Measure, and Manage and calls for documented context, roles, risk tolerance, task scope, oversight, testing, and risk treatment. It is voluntary guidance rather than a readiness certification.

Sources: [1]

Build the readiness evidence map

Use rows for purpose and task fit, authoritative data, permitted access, tool configuration, user proficiency, reviewer competence and capacity, quality testing, exception handling, monitoring, change control, and release authority. For each row, record the evidence, owner, date, status, material gap, and next test.

Use decision states such as discovery, controlled practice, restricted operation, broader operation, and hold. Define entry and exit evidence for each. Training completion may support a capability row, but it does not close data, quality, control, or operating-ownership gaps.

The OECD’s 2026 Building an AI-ready public workforce discusses governance, skills, strategic alignment, and continuous learning in public administration. Its examples inform workforce planning but do not validate a universal corporate readiness model.

Sources: [2]

Worked artifact: a synthetic contract-summary workflow

Imagine a synthetic procurement team proposing an assistant to summarize supplier contracts before renewal meetings. The evidence map shows approved access to a controlled contract repository, but metadata quality varies and several amendments are stored separately. Reviewers understand the agreements yet have limited capacity during quarter end.

The controlled test includes an agreement whose amendment changes the termination period. The assistant retrieves the base contract but misses the amendment. A reviewer catches the conflict only by searching manually. The data row therefore remains not ready, and the review-capacity row is amber because the workflow depends on extra investigation.

The team adds an amendment-completeness check, displays the source hierarchy beside each summary, and blocks release when linked records are missing. A second test confirms that the stop condition activates, but no production-volume evidence exists. The decision is to continue controlled practice rather than declare broad readiness.

The artifact also records that the procurement director owns the renewal decision, technology owns repository integration, and a contract specialist owns interpretation. The example is synthetic and does not establish legal accuracy, control effectiveness, or Two Prune customer results.

Distinguish capability from operating conditions

Assess people with representative tasks that create observable opportunities for framing, evidence navigation, orchestration, verification, judgment, and communication. Mark missing opportunities explicitly. Do not infer proficiency from confidence, tool activity, prompt style, or a final artifact alone.

Then inspect the environment: time, source access, interface, workload, escalation, incentives, accessibility, and policy. A capable participant can fail in a poorly designed workflow, while a good output may result from hidden expert correction. Readiness requires both human and system evidence.

Verify readiness over time

Reconstruct sampled cases from source to release. Reperform material calculations and check versioned documents, logs, reviewer actions, exceptions, and stop mechanisms. Compare described controls with actual permissions and behavior. Test degraded conditions and realistic volume.

Reopen the evidence map after changes to model, prompt, retrieval, data, autonomy, population, reviewer staffing, policy, or downstream consequence. This method supports explicit stage decisions and honest gaps. It does not establish a mature benchmark or predict performance outside the tested workflow. Two Prune supplies task-bounded assessment evidence for human interpretation, not an enterprise readiness verdict.