How to enable managers for AI adoption

Define the manager's decision

Choose a real workflow and ask what the manager must decide: approve a bounded trial, allocate review capacity, set a non-use boundary, respond to an error pattern, or determine whether the task is ready to expand. Separate that decision from technical administration and specialist approvals.

The OECD paper on using AI in the workplace describes potential benefits and workplace risks and discusses consultation with managers, workers, and representatives. Consultation can surface workflow knowledge and concerns, but it does not replace a tested operating design.

Write the task, intended outcome, evidence inputs, permitted data, AI contribution, required human action, and consequence of error. A manager should be able to explain why the workflow exists and what would make it unsuitable.

Sources: [1]

Build a manager decision card

Use sections for task boundary, participating roles, source of truth, tool and configuration, review standard, exception path, evidence sample, workload effect, affected groups, and decision date. Add an owner for each unresolved issue.

The NIST AI RMF Core calls for clear responsibilities, practitioner proficiency, human oversight, feedback, and documented risk responses. A manager card makes those responsibilities concrete at the work level without asking a manager to become the privacy, security, legal, or model expert.

Provide four dispositions: approve a small test, approve with conditions, redesign, or do not use for this task. Avoid a binary “AI yes or no” framing because the same tool can be appropriate for drafting public material and inappropriate for processing sensitive records.

Sources: [2]

Work a synthetic finance-team example

Imagine a finance operations manager considering an AI assistant for drafting monthly variance commentary. The source pack includes an approved spreadsheet, accounting definitions, a prior narrative, and a schedule. The assistant may summarize calculated variances but may not invent causes, alter the workbook, or send commentary externally.

The manager card records:

In the synthetic trial, the assistant correctly identifies a 12 percent variance but attributes it to supplier prices without evidence. The participant changes the statement to “cause unresolved” and attaches the transaction review needed. The manager approves continued testing with a rule that causal explanations require a cited internal record.

This result does not establish productivity improvement or organization-wide capability. It shows that one bounded review process caught one material failure and produced a clearer condition for the next test.

• purpose: reduce first-pass drafting time while preserving accountable analysis;
• source of truth: approved workbook version and controller definitions;
• participant action: investigate material variances and label unverified explanations;
• review: controller checks numbers, causal statements, and disclosure language;
• workload: reviewer has capacity for five trial reports, not the full portfolio;
• stop condition: totals disagree, source version is unclear, or the assistant presents an unsupported cause as fact;
• evidence: rendered commentary, correction log, source links, and reviewer disposition.

Prepare managers for difficult cases

Practice cases where a high performer wants an unapproved tool, an approved tool lacks access to the required source, a deadline exceeds review capacity, an employee reports surveillance concerns, and a generated output looks polished but cannot be traced. Managers should know who owns each decision and how to pause safely.

Give managers a short escalation map: workflow owner, data owner, security, privacy, legal, accessibility, workforce representative, and tool administrator. Not every case needs every specialist. The map prevents unresolved issues from being silently converted into manager approval.

Track support needs without using them as punitive measures. Questions, corrections, and appropriate non-use can indicate responsible operation. Review whether managers have time, access, examples, and authority to enforce the stated boundary.

Review the operating decision

At the review date, inspect sampled artifacts, material corrections, unresolved issues, review time, access failures, and participant feedback. Decide whether to continue, narrow, redesign, or stop. Record the evidence and what would justify a different decision later.

Repeat the review when the task, tool, data source, policy, participant group, or consequence changes. A one-time approval is not evidence that future configurations remain suitable.

Two Prune can provide a consistent simulation and structured, task-bounded evidence about how participants frame, verify, and deliver AI-assisted work. It does not supervise employees, approve workplace deployment, set performance consequences, or make employment decisions. Managers and qualified reviewers retain responsibility for actual operating choices.