Enabling managers to support AI-enabled work

Define what the manager controls

List the decisions available to the manager: task assignment, time, review expectations, approved sources, escalation, practice opportunities, feedback, and local workload. Separate these from platform configuration, legal interpretation, security approval, and enterprise policy.

Give managers a specific workflow rather than a generic AI mandate. Name the output, quality conditions, authoritative evidence, material failure modes, and release authority. Managers should be able to explain when AI may assist, when a person must decide, and what stops the work.

The NIST AI RMF Core calls for clear responsibilities, role-relevant training, critical thinking, human oversight, and feedback from relevant actors. It is voluntary guidance and does not prescribe a manager-training programme.

Sources: [1]

Build the manager enablement loop

Prepare a short workflow card with the task, source hierarchy, AI contribution, required checks, escalation route, and evidence to retain. Use a representative practice case that contains one plausible failure. Observe what the participant could actually see and do; do not score behavior that the case did not allow.

During coaching, ask the participant to reconstruct the decision: what were you trying to achieve, which source carried authority, how did AI contribute, what did you verify, what remained uncertain, and why did the final output change? The manager should cite evidence rather than infer intent from confidence or fluency.

The UK government’s 2026 employer guide to AI upskilling emphasizes practical and integrated learning connected to everyday work. Its evidence can inform the loop, but local task design and evaluation still require testing.

Sources: [2]

Worked artifact: a synthetic monthly forecast

Imagine a synthetic finance manager coaching analysts who use AI to draft monthly forecast commentary. The practice pack contains a spreadsheet, an approved planning assumption, and an email proposing an unapproved sales adjustment. The assistant produces a confident narrative that mixes the approved assumption with the email.

The analyst identifies the conflict but leaves the sales adjustment in the final paragraph without qualification. In the coaching conversation, the manager traces each material statement to its source and asks who owns the unresolved assumption. Together they revise the paragraph to separate the approved forecast from the proposed upside and record an escalation to sales leadership.

The manager then changes one operating condition: every material variance must link to an approved assumption or carry an unresolved label. In a second synthetic case, the analyst applies the rule and catches a similar conflict. The loop records the source comparison, original and revised language, coaching question, operating change, and second-case evidence.

The result is bounded evidence that the analyst applied the rule in the second exercise. It does not show durable transfer, future performance, or a business outcome. The figures and participants are synthetic, and no result drives an automatic employment decision.

Coach the work system, not only the individual

Look for hidden causes: missing access, ambiguous authority, deadline pressure, inaccessible tools, review overload, conflicting incentives, or an unsafe local workaround. If the environment prevents verification, change the environment. Repeating “be careful” is not a control.

Track manager capacity as well. Measure coaching time, unresolved escalations, review queues, repeated failure types, and whether the manager can stop release. A programme that adds extensive unpaid checking work may need redesign even if usage rises.

Verify the loop and avoid surveillance shortcuts

Sample coaching records and confirm that evidence supports the stated observation. Check that examples remain representative, sources current, accessibility needs met, and managers apply anchors consistently. Retain uncertainty and missing evidence rather than converting them into negative judgments.

Do not use prompt counts, message volume, or continuous monitoring as a substitute for task evidence. Reopen the loop after changes to workflow, model, source system, policy, or team. This method supports practical manager coaching. Two Prune can provide task-bounded evidence from controlled simulations for human interpretation; it does not manage employees or issue employment recommendations.