Before AI Enters Performance Reviews, Define What Good Work Means

“Should we put AI usage rates into this year’s performance goals?”

If this question came up in a meeting room, I would probably stay silent for a few seconds first.

It is an important question, and it will come sooner or later. The company bought Copilot, rolled out an internal knowledge base, ran AI training courses. Eventually someone asks: how do we know whether people are actually using it? And has anything gotten better?

The trouble is, once this question connects to the performance system, it becomes sensitive. Employees start guessing: if I finish a proposal with AI, does that earn points or lose them? If quality improves, will my manager see me as more capable, or decide the work no longer counts as mine? If I honestly disclose AI assistance, will I lose out to colleagues who don’t?

None of these questions has a standard answer, yet every one of them shapes how employees will use AI next. Frankly, rushing to answer this one tends to leave both sides unhappy.

Don’t Rush to Measure Usage Rates

Putting AI usage rates into performance metrics looks intuitive. The system has logs, employees have accounts, and managers want to know the adoption status. For senior leadership, this is far easier than discussing whether the way people work has actually changed.

But usage rates can only answer a narrow question: whether this person touched the tool.

They struggle to answer the questions that matter more: did this person define the problem clearly, and do they know which data must never go into the model? Whether the reasoning gaps in AI output were caught, and whether the decisions that need a manager’s judgment were preserved — none of that shows up in the logs. As for whether someone took back the responsibility they should carry before delivery, that is even less visible.

If the performance system only tracks AI usage volume, employees quickly learn one thing: performative usage. Mention AI in meetings, paste a few prompts into weekly reports, append a line saying “AI-assisted” to deliverables. These moves make adoption rates look good, while work quality does not necessarily move at all.

There is another kind of trouble. Over-tracking usage punishes the careful ones. In compliance, finance, HR, customer complaints, healthcare, or financial services work, some data simply cannot be fed into general-purpose models, some judgments must keep human review; and for some tasks, doing them with AI costs more than it returns. If these people get rated lower because they use AI less, the system is effectively rewarding risk-taking.

This depends on the organization’s risk environment and cannot be generalized.

flowchart TD;

subgraph S1["Usage-Driven Trap"];

A1["Count system logs / logins"] --> B1["Employees showcase prompts and AI labels"];

B1 --> C1["Surface adoption rises, actual quality goes unexamined"];

end;

subgraph S2["Quality-Driven Thinking"];

A2["Define output standards for good work"] --> B2["Review data sources, assumptions, and exceptions"];

B2 --> C2["Assess human judgment, verification quality, and accountability"];

end;

The Standard for Good Work Must Withstand Questioning

If a company genuinely wants to bring AI into performance management, I would suggest replacing the starting point of the discussion. File “did they use AI?” away in the IT department’s adoption report. The question the meeting room actually needs to discuss is:

“After AI participates, what makes this work good?”

The question sounds ordinary, but answering it forces the organization to lay out judgments that used to stay vague. Managers used to assess proposals by experience: complete content, polished formatting, fast turnaround meant this person was doing well. After AI arrives, polished formatting and speed lose their value together, and managers need to see something more fundamental.

Take an example. For a market analysis report, AI can help organize data and generate a first draft, and the prose can read smoothly. What the manager needs to look at may become: did the employee pick the right data sources, which assumptions did they keep, what plausible-but-wrong content did they delete, and did they pull the recommendations back within business constraints? These judgments used to live inside the employee’s head. Now they need to be written into work records or delivery standards.

Of course, no one should ask employees to log every step like a running diary. That would drive everyone mad, and managers would not want to read it anyway.

A more practical approach is to define, for different types of work, a few evidence fields that can be questioned after the fact. Like this:

Work Type	What to Look At	Traceable Work Evidence
Analysis and decision recommendations	Data sources, assumptions, risks and trade-offs	Key basis, excluded options, human judgment points
Customer or employee communication	Tone, compliance, audience context	Reasons for revisions, handling of sensitive information
Process improvement	Problem definition, scope of impact, exception handling	Before-and-after process differences, nodes requiring human handover

This table is only a directional reminder; every company needs to adapt it to its own roles and risk environment before use. What it says is simple: the focus of performance systems can gradually shift from “how much you did” to “how you judged, verified, and took responsibility.” Once AI makes execution fast, what managers need to see most is the quality of human judgment.

HR Must Protect Fairness, and Managers’ Room for Judgment

This is where HR needs to step in.

IT can set tools, permissions, data security, and system records; compliance can cover risk, policy, and regulatory requirements. But the performance system runs into something else: how employees understand “am I doing well,” how managers give feedback, and how the organization avoids turning AI use into a new form of unfairness. These questions will not appear in system specifications, nor on compliance checklists.

CIPD’s 2026 AI governance guide reminds us that the governance scope of people professionals should cover any AI deployment that affects people, reaching beyond HR systems themselves. Placed in the context of performance systems, that sentence carries real weight. Once a company starts asking about the relationship between AI use and performance, it has moved past a simple tool-adoption question into careers, evaluation, fairness, and the distribution of responsibility.¹

Microsoft’s 2026 Work Trend Index states the problem clearly as well: AI and agents expand individual capability, but for organizations to absorb that capability, they must address leadership, culture, management practices, and how work is measured. The report notes that more mature AI users discuss quality standards for AI-assisted work more often, and more often document agent workflows, human-AI handovers, and quality standards.²

Gallup’s data adds another angle. As of February 2026, half of employed Americans use AI at work at least occasionally; in organizations that have adopted AI, 65% of employees say AI has a positive effect on their productivity. Yet only 12% strongly agree that AI has changed how their organization works.³ This gap deserves HR’s attention: individuals already feel faster, while the organization’s processes, roles, and evaluation methods have not caught up.⁵

Deloitte’s 2026 Human Capital Trends also notes that as AI expands into work, workers are re-asking questions about effort, ownership, fairness, and accountability. Placed in performance management, these words become a string of very practical questions: after using AI, how does effort get seen, who owns the outcome, who is responsible for errors; and since roles differ greatly in how much AI they can use, how do performance ratings stay fair?⁴

flowchart TD;

subgraph S1["1. Quality Standard"];

A1["From output speed and volume to assumptions and logic"] --> B1["Focus on source review, error screening, and context fit"];

end;

subgraph S2["2. Accountability"];

A2["Clearly mark human review and handover points"] --> B2["Ownership and error responsibility stay clear despite AI use"];

end;

subgraph S3["3. Traceable Evidence"];

A3["Keep traces of decisions and revisions"] --> B3["The point is explaining changes to managers, kept brief"];

end;

subgraph S4["4. Fairness & Risk"];

A4["Separate high-risk compliance zones from low-risk trial zones"] --> B4["Avoid over-tracking usage that punishes careful, compliant behavior"];

end;

So what HR needs to do is bring the question back into the language of institutions.

HR sits in a delicate position, with both sides to mind: pressure from leadership to drive adoption on one side, employees’ sensitivity to fairness on the other. The more practical move is to help the organization sort out a few things: which roles can be encouraged to use AI, which tasks require disclosing AI assistance, which outputs must keep human review points, and which performance metrics would accidentally punish careful behavior.

Here is a line I would bring back to the meeting room:

“We can encourage AI use, but performance evaluation must first define quality, responsibility, and evidence. Usage rates can serve as an adoption signal, never directly as a measure of contribution.”

At minimum, this line lets the discussion move away from the anxiety of “who is actually using AI” and back to what the system truly has to answer: now that AI has entered the work, how does the organization judge the value of people? As for what the quality standards look like for each role, that still takes sitting down with one department at a time.

---

Read More

• The Judgment Economy: When AI Slashes Output Costs, Who Judges Right from Wrong?
• Where Enterprise AI Transformation Gets Stuck: When KPIs Run Ahead of Use Cases
• AI Skills Planning Beyond Tools: Process Capacity Is the Key to Transformation
• The AI Efficiency Trap: Why AI Can Make Organizations More Complacent

---

References

Frequently Asked Questions

Can AI usage rates be used directly as a performance metric?

Usage rates only tell whether an employee touched the tool; they reflect nothing about judgment quality or accountability. Writing usage into KPIs invites performative usage and punishes roles like compliance and finance that must use AI carefully. A steadier approach treats usage as an adoption signal and defines contribution criteria separately. How should managers evaluate work quality after AI enters the workflow? | Shift the focus from output volume and speed to data source selection, retained assumptions, error screening, and human judgment points. Define a few evidence fields per work type that can be questioned afterward, such as key basis, excluded options, and human judgment points for analytical work. Do employees need to disclose AI use? | It depends on the role and risk environment. Organizations should first sort out which tasks require disclosing AI assistance and which outputs must keep human review points, and ensure honest disclosure never costs anyone in ratings; otherwise the system rewards concealment. Why does AI performance evaluation need HR involvement? | IT manages tools and permissions, compliance covers policy and regulation, but how employees understand doing well, how managers give feedback, and whether the system creates new unfairness are performance and fairness questions. CIPD 2026 guidance also holds that people professionals should participate in governing all AI deployments that affect people.