Talent and performance
What Is Performance Management? What to Measure When AI Does Part of the Job.
Performance management is the year-round cycle that turns expectations into better work. The hard question in 2026 is what the cycle should measure once an agent does the execution.
Most people meet performance management as a form. Once a year a manager fills it in, scores the work on a scale and files it, and everyone on the receiving end learns what the organization really values from what the form asks about.
That is the trouble with the form in 2026. On most of them, the part of the job that gets scored is the part an AI assistant now does: the volume produced, the tasks closed, the reports delivered.
People notice. They protect the work that gets rated, even when a machine does it better, and the AI rollout stalls without anyone quite saying why.
Performance management is the whole system that form sits inside. It matters more this year than last because the system now decides, quietly, whether people hand work to the machine or hold on to it.
Performance management vs performance appraisal.
The two terms get used as if they meant the same thing, and the confusion does real damage. A performance appraisal is an event: a rating of a person's work over a period, usually once or twice a year, that produces a score and feeds pay and promotion.
Performance management is the whole cycle the appraisal sits inside, from setting expectations at the start of the year to rewarding the results at the end, with the feedback and development in between.
| Question | Performance appraisal | Performance management |
|---|---|---|
| What is it? | A rating event | The whole cycle |
| How often? | Once or twice a year | All year, with checkpoints |
| Who drives it? | HR, through a form | The manager, through conversations |
| What does it look at? | The past | The past and what comes next |
| What comes out? | A score | Better work and a development plan |
The appraisal is a part. The mistake most organizations make is running the part as if it were the whole: a year of silence, then a form. Everything that follows is about the whole.
The performance management cycle.
The clearest statement of the performance management process I know comes from the US Office of Personnel Management, which sets the rules for the federal workforce.
It lists five components: planning, monitoring, developing, rating and rewarding. Vendors rename the stages, but the five hold up, so here they are, each with what the evidence says about it.
01Plan
Set expectations and goals with the person, not for them, and write them down. This is the step most systems skip, and it shows.
In the first half of 2026, 49% of US employees strongly agreed that they know what is expected of them at work, well below the 61% peak in 2015 (Gallup, July 2026).
OPM's own guidance gives the reason for doing it with the person: people who help plan the work understand what needs to be done, why, and how well.
02Monitor
Watch the work and talk about it as it happens, not months later in the review. On Gallup's 2026 engagement tracker, fewer than a third of US employees strongly agree that someone at work has talked with them about their progress in the past six months.
A check is a conversation, not a status report: what is working, what is in the way, what should change.
03Develop
Close gaps while they are small, through coaching from the manager, training and the next assignment. Make development goals count as much as performance goals. A system with no development step is an appraisal with extra paperwork.
04Rate
Summarize the period with a rating that rests on more than one input and is calibrated across managers, not one manager's impression written up as a number.
A rating is also the data that succession runs on, so it should be tested against readiness criteria written for the roles you will need, not the roles you have.
05Reward
Pay, promotion and recognition, in a conversation held apart from the feedback conversation, because once money is on the table nobody hears the feedback.
The link to pay matters: in a December 2025 Gartner survey of 1,622 respondents, people who believed their pay was strongly linked to their performance were up to 17% more productive.
What is performance management in practice: a cycle of planning, monitoring, developing, rating and rewarding
Plan
Set expectations and goals with the person, in writing, at the start
Monitor
Watch the work and talk about it as it happens, at least quarterly
Develop
Close gaps while they are small, with development goals that count as much as performance goals
Rate
Summarize the period from more than one input, calibrated across managers
Reward
Pay, promote and recognize, in a conversation held apart from the feedback
Then plan the next period with what the last one taught you. The cycle does not end; it gets faster.
The annual review is one event in the fourth step. Most systems run only that step.
Source: US Office of Personnel Management, Performance Management Cycle
Why performance management matters, and why most systems don't work.
Performance management matters because it is the mechanism that connects what one person does on a Tuesday to what the organization is trying to do this year, and because it is the place where people learn, in a form they cannot ignore, what the organization really rewards.
When it works, people know what is expected, hear how they are doing while they can still do something about it, and see pay follow the work.
Most systems do not work, and the people who own them cannot show that they do. In an August 2026 survey of 934 HR and business leaders by Betterworks, which sells performance software, 85% of the HR leaders were very or extremely confident that performance management helps deliver the organization's priorities.
Only 44% could answer the CEO with clear business metrics. Another 23% would point mainly to engagement or activity numbers.
Employees see the gap from the other side of the desk. In a Gartner survey of nearly 3,000 employees and managers, published in April 2026, 39% of employees agreed that their manager is effective at giving clear developmental feedback.
How sure HR leaders are that performance management works, and how many can show it
One survey of 488 HR leaders and 446 business leaders in the US and UK, August 2026. The people who own the system are sure it works. Fewer than half can show it.
Source: Betterworks, 2026 Performance Management Survey, October 2026
So why do most systems fail? Not because organizations are bad at measuring. They measure constantly. The systems fail because they measure the wrong thing, activity, and measure it once a year, and then attach pay to the result.
Activity is easy to count, which is why forms count it. But the count rewards volume, volume rewards busyness, and busyness is exactly what an organization with AI in the work needs less of.
From the annual rating to continuous performance management, and halfway back.
The last decade's answer to all this was frequency: replace the annual rating with regular check-ins and let the conversation carry the year.
It never fully took. In the same Betterworks survey, 69% of the business leaders said their formal process was still annual or semiannual reviews, with or without check-ins in between.
Then the swing back. In January 2026 Meta told staff that its new review system, Checkpoint, would rate people twice a year on four levels and tie bonuses to the rating, with about a fifth of staff expected to be rated Outstanding, as Fortune reported.
In August 2026 Microsoft began retiring its peer feedback tool and moved to five rating categories with sharper differentiation, Business Insider reported. Its reason: feedback works best in regular conversations, not formal requests.
So the pendulum stops halfway: keep the conversation, sharpen the number. Pay and promotion decisions still need one, and a decade of check-ins did not remove the need.
Here is what I take from the decade. Continuous performance management is better than annual, because people need to hear how they are doing while they can still act on it.
But the frequency debate was the wrong debate. It changed how often you rate. What changed in 2026 is what there is to rate.
What performance management should measure when AI does part of the job.
Performance management and AI meet at the rating criteria. When an agent takes over the execution of a role, a form that counts tasks closed, hours logged and volume produced is now scoring the machine.
Worse, it is paying the person to compete with the machine. Reward volume and you reward teams for overriding the agent to protect their own numbers, which is how an AI investment ends up used by the enthusiasts and routed around by everyone else.
In Rewired Leadership I set out three mandates for the CHRO, and the first is this one: shift the entire performance system to measuring the quality of human judgment, not the volume of human activity.
The only thing AI cannot reproduce is judgment applied to ambiguous, high-stakes decisions. If that is what people are for now, it is what the review has to measure.
The way in is what the book calls the Judgment Audit. Map every core function's output into three categories: execute, which the agent does; synthesize, which the agent and the person do together; and judge, which stays with a person and gets elevated.
Then write the rating criteria for the third column. Doing the audit honestly means settling which decisions stay with a person and which an agent may take inside the guardrails, and that work belongs to the CHRO and the CTO together, not to HR alone.
What the new form scores looks different from the old one. The metric names below come from the dashboard the book proposes for leaders.
| What the old form scores | What the new one scores |
|---|---|
| Activity: how much the person did | Judgment: how well the person decided |
| Tasks closed | Decisions made well inside the guardrails |
| Hours logged | Decision cycle time, from signal to decision |
| Everything escalated upward, to be safe | Escalation rate, and the quality of what gets escalated |
| Individual output | System health: what the team can do without the person |
| How much the person added | How much complexity the person removed |
Two of those rows need a word. Exception quality is the share of what a person escalates that actually needed a senior decision; a leader who gets only good escalations has built a good system. And team confidence, which you measure by asking, is the earliest signal that delegated judgment is working.
The leader's own scorecard changes most of all. The book's position is that a meaningful share of a senior leader's bonus should ride on the delegation rate of their team and how many of their people move into higher-judgment roles, so that a leader is paid for what the team can do without them.
And in the chapter on why leaders sabotage their own rewire, I argue that a leader's review should be based on how much complexity and how many low-value tasks they removed from the organization.
Subtraction is the discipline. Throughput is the habit it replaces.
Gartner's Tony Guadagni put the division well in October 2025: the processes of performance management will be automated, but managing performance cannot be. In Gartner's December 2025 survey, managers using AI reported saving an average of four hours across the performance management process.
The question for the CHRO, whose own job is changing faster than anyone's, is what the four hours go to. In a performance system rebuilt around judgment, they go to the conversation.
Using AI in performance reviews without losing the review.
The tools arrived before the rules.
In August 2026 Highwire surveyed more than a thousand full-time US professionals and found that 78% of managers had used AI to prepare employee feedback or a performance review in the past year, while 16% of employees said they were told AI played a role in their last review (Highwire, 2026).
The employees were split on the result: 54% said the feedback had become more specific and actionable, 34% said it had become more generic, and 32% said it was less useful, as HR Dive reported in October 2026.
Gartner's finding from the autumn before is the uncomfortable part: most of the managers experimenting with AI in performance management have had no formal training in it, and Gartner's advice is approved tools plus training on bias.
AI in performance reviews, from Highwire's 2026 survey
Nearly four in five managers use it. One in six employees is told.
Source: Highwire, Who's Managing Whom, October 2026, employee figures as reported by HR Dive
AI performance reviews can be better reviews. A model that reads a year of project notes, customer feedback and check-in records will remember more than a manager does in late November. Three rules keep the review a review:
- Tell people. Disclose that AI helped draft the feedback, and what it was given to work from. A review that turns out to have been written by a tool nobody mentioned does the same damage to trust as any other discovery about how bad news and mistakes are handled.
- The model drafts from evidence. The manager owns the judgment. The tool can assemble what happened. Only the manager can say what it meant, and the manager has to be able to defend every line in person.
- The model never produces the rating. A score generated by a system is a score nobody is accountable for, and the rating is the one part of the process that exists to hold a person accountable.
Examples of performance management.
Three performance management examples, one illustrative and two from the book, show what rating judgment looks like in practice.
A sales team with an AI assistant. Picture a team whose assistant now drafts every proposal and follow-up. The old form rated reps on proposals sent and calls made, so the assistant made every rep look identical.
The redesigned form rates the calls the assistant could not make: which deals to walk away from, what to concede in a negotiation, when to bring in a specialist. Proposals sent still appear on the dashboard. They no longer appear on the review. This example is illustrative, not a client case.
The hero bonus. In Rewired Leadership I describe a manager who bypassed protocol and personally rescued a large delivery after a system failure. The old culture would hand that manager a $10,000 bonus.
The rewired action is for the CEO to thank the manager publicly for their speed and give the major award to the team that documented the flaw in the guardrail that caused the failure. The principle, in the book's words: we don't pay for heroes; we pay for systems.
The CLO and the thirty hours. Also from the book: a chief learning officer spent thirty hours a month correlating course completion, performance data and business outcomes by hand.
An agentic system took over the correlation and flagged the anomalies, and her review time fell to three hours on a single dashboard. Her performance was never the thirty hours.
It was what she did with the three: challenging the design of the curriculum based on what the system surfaced. The counting left. The judgment stayed, and finally had room.
How to build or fix a performance management system.
This is how the design work runs when I do it with a CHRO. The order matters, and the first two steps decide whether the rest is worth doing.
- Diagnose how the system actually works. Run the Judgment Audit on the roles that matter most, and read how reviews, pay and promotion really happen today, not how the policy says they do. The gap between the two is usually the finding.
- Decide what gets rated, by whom, and how AI-assisted work is credited. This is the performance management framework in one page. If the answer to the third question is "the same as before," stop and go back to step one.
- Design with the people who will run it. The CHRO owns the system. The managers who will hold the conversations help build it, because a system managers did not shape is a form they will fill in late.
- Give development goals equal weight, and separate the money. Make development goals count as much as performance goals, and hold the pay conversation apart from the feedback conversation.
- Pilot with one team for one real cycle, usually a quarter. A pilot that skips the rating and reward steps has tested nothing.
- Roll out with manager guides, not a memo. The system is only as good as the manager running the conversation, and in the Gartner survey published in April 2026 only 41% of employees said their manager helps them prioritize their work. Train for the conversation, not the form.
If you are not sure where the system breaks, an assessment of your performance system is the place to start, because step one is most of the work.
Common mistakes.
- Counting activity because it is easy to count. The count is the problem, not the fix.
- Running the appraisal as if it were the process. A year of silence and then a form is not performance management.
- Letting the tool draft the rating, not just the words. A score nobody is accountable for is not a score.
- Rating the person for work the system now does. You are scoring the machine and paying the person to fight it.
- Changing the frequency and calling it a redesign. Quarterly reviews of the wrong thing are still reviews of the wrong thing.
- Paying leaders for their own output instead of their team's autonomy. Nothing makes a leader hold on to execution faster than a bonus for doing it.
Frequently asked questions.
The five stages, as the US Office of Personnel Management lists them, are planning, monitoring, developing, rating and rewarding.
Plan expectations with the person, watch and discuss the work as it happens, close gaps while they are small, summarize the period with a calibrated rating, and reward the result. Then plan the next period.
To make individual work add up to the organization's goals. It does that by setting clear expectations, giving people feedback while they can still act on it, developing them, and tying pay and progression to results.
In 2026 its purpose includes deciding what people are rated on once AI does part of the job.
A performance review, or appraisal, is a single event: a rating of a person's work over a period, usually once a year.
Performance management is the whole cycle the review sits inside, from setting expectations to rewarding results. The review is one step of five, and most of the value sits in the other four.
Lists of five Cs circulate widely, usually some mix of clarity, communication, coaching, consistency and consequences, but they vary by vendor and none that I have found cites a source.
The version with a source is the five stages above, which come from the US Office of Personnel Management.
Maybe. The question that matters more is what the ratings reward. If they reward activity, dropping them changes nothing and removes the number pay decisions need. Meta and Microsoft both kept the rating in their 2026 redesigns and made it sharper.
Fix what the ratings reward first. Then decide whether you still need the number.
AI can draft one, and a draft built from a year of real evidence is often better than a manager's memory in November.
It should not produce the rating, and the manager has to own every line. Tell people it was used. In Highwire's 2026 survey, 78% of managers used AI and 16% of employees were told.
The short version.
Performance management is the year-round cycle that turns expectations into better work: plan, monitor, develop, rate, reward, and plan again. Most systems fail because they rate activity once a year, and in 2026 the activity is leaving for the machine.
The job of the cycle now is to rate judgment, and to pay leaders for what their teams can do without them. That is the first module of the talent and performance work we do, and it starts with what the form asks about.