How Managers Should Evaluate Work Done by People and AI

Managers are starting to evaluate work that was produced by a person and AI together. The employee framed the problem, gathered context, directed a model, checked parts of the answer, changed the output, and made the final call. Looking only at the employee's effort misses what the tool contributed. Looking only at the model's output misses the judgment that made it useful.

A 2026 CESifo working paper makes the manager's role hard to ignore. Using longitudinal Gallup Workforce Panel data covering more than 30,000 U.S. employees, the study found that frequent AI use rose from under 10 percent in 2023 to more than 27 percent by early 2026. Employees who perceived a clear organizational AI strategy were 27 percentage points more likely to report frequent use. The paper also found that the relationship between AI use and worker outcomes varied with strategic clarity and trust.

Those findings are associations, not proof that a management memo caused better outcomes. They still point to a practical truth: employees take their cues from what managers explain, reward, review, and make safe to learn. AI adoption becomes an operating practice through day-to-day management, not through the purchase order alone.

Start with whether the work should use AI

The first management decision comes before the prompt. Is this a good place to use AI? High-volume work with clear inputs, inspectable outputs, and a repeatable standard is often a reasonable candidate. Work that depends on confidential context, unsettled policy, personal discretion, or a hard-to-reverse decision may require tighter boundaries or a different approach.

Ask what the employee is trying to improve. Speed is one answer, but it is not the only one. AI might help a team explore more options, apply a standard consistently, summarize a large record, draft a first version, or make expert practice easier for a newer employee to use. The purpose determines how the work should be evaluated.

The NBER study Generative AI at Work illustrates why context matters. Researchers studied an AI assistant used by 5,179 customer support agents and found an average productivity increase of 14 percent, with larger gains for novice and lower-skilled workers and little impact on experienced, highly skilled workers. The same tool can create different value for different people because the work and the starting point differ.

Define good work before measuring the workflow

Managers need a standard for the output that does not depend on whether AI was involved. A customer response should be accurate, useful, on policy, and appropriate for the situation. An analysis should use the right evidence, show its assumptions, and answer the business question. Code should be correct, maintainable, tested, and safe to operate.

Then add the workflow requirements created by AI. Which sources may the system use? What information may enter the tool? Which claims need verification? What uncertainty must be surfaced? Which parts of the work need a trace or retained evidence? A clear standard lets an employee experiment without guessing what the manager will consider acceptable afterward.

• Quality: does the result meet the same accuracy, completeness, policy, and customer standards as other work?
• Judgment: did the employee recognize where context, ambiguity, or consequence required more than a fluent answer?
• Evidence: can a reviewer inspect the important sources, assumptions, tests, and decisions behind the result?
• Efficiency: did the workflow reduce total time or cost after review, correction, and coordination are included?
• Learning: did the employee improve the process, document an exception, or make the next run more reliable?

Put human review where judgment changes the result

A requirement that a human reviews everything sounds safe and often says very little. Review belongs at the points where an error matters and a person can evaluate it. A manager should identify the decision, claim, action, or threshold that needs human judgment, then make sure the reviewer has enough context and authority to reject the work.

Some workflows need review before an external action. Others can use sampling after the fact because the work is low consequence and reversible. A mature process may automate routine cases and route exceptions to a specialist. The review design should follow the cost of error, the ability to detect it, and the difficulty of recovery.

Count the review burden. G2's 2026 AI Agents Insight Report found that oversight and rework can erase a meaningful share of expected productivity gains for many buyers. An agent that finishes a task quickly but creates hours of checking has moved work around. Managers should measure the whole workflow from request through accepted result.

Keep one person accountable for the result

AI participation does not create a third place for accountability to go. The employee or business owner who submits the work still owns the result. That person should understand what the system contributed, what was checked, what remains uncertain, and what happens if the output is wrong.

Ownership should also exist at the workflow level. Someone needs to decide whether the process is approved for its purpose, which tools and data it may use, how quality is monitored, and when the process should change or stop. The operator can own an individual result while a manager or process owner owns the recurring system around it.

Review the process without turning it into surveillance

Managers do not need every prompt or keystroke to evaluate hybrid work. That level of monitoring encourages performance theater and can expose sensitive employee thinking without answering whether the work was good. Focus on the workflow: the task, approved tool, relevant inputs, quality checks, owner, accepted output, cost, and result.

The employee should be able to explain the role AI played and the judgment they supplied. A concise review can ask what changed, where the model struggled, what the employee verified, and what they would do differently next time. That conversation produces better management information than a leaderboard of messages sent to a chatbot.

Graduate useful workflows from habit to team practice

The best individual workflows should not remain private tricks. When a process repeatedly produces good work, the manager should help turn it into a team asset. That means naming the purpose, inputs, tools, steps, quality standard, review points, owner, exceptions, and fallback. Another employee should be able to run it and reach a comparable result.

Graduation should be earned. Start with an individual experiment, test it on representative work, compare the output with the existing process, and document failures. Move it to team practice when the quality and economics hold up beyond the original creator. Expand it again only when ownership, access, training, and measurement can support the larger scope.

• Personal experiment: one employee tests a bounded use and remains responsible for every result.
• Repeatable workflow: the employee documents the process, review standard, known failures, and required tools.
• Team practice: another person reproduces the workflow, the manager names an owner, and results are sampled over time.
• Scaled operation: access, cost, quality, exceptions, training, and outcomes are monitored across the wider deployment.

Reward contribution, not visible AI activity

Raw usage is a poor performance target. It rewards employees for generating activity whether or not the activity improves the work. It can also penalize an expert who uses AI selectively because they know exactly where it helps and where it introduces noise. Usage can show where adoption exists, but it cannot grade an employee on its own.

Performance conversations should stay anchored in contribution. Did the employee improve the result, serve the customer, reduce cycle time, make a sound decision, teach the team, or build a reusable process? Did they handle uncertainty honestly and take responsibility for mistakes? AI changes how the work was produced. It does not remove the need to judge the work itself.

Give managers an operating view

Managers cannot coach what they cannot see. They need a grounded view of which tools and agents are being used, which workflows matter, who owns them, what they cost, where adoption is spreading, and which activity is connected to outcomes. That view should support a conversation about work, not a scorecard that treats volume as value.

Proxon connects AI usage, tools, agents, workflows, ownership, spend, adoption, and outcomes in one management view. It does not evaluate an employee's performance or replace managerial judgment. It gives managers a shared operating record for deciding where to coach, where to standardize, and where an AI workflow deserves more investment.

Every employee using AI is learning to manage a new kind of collaborator. Their managers have a new job too: set the standard, place the review, keep ownership clear, and help good individual practice become a durable team capability.