AI in Performance Management
What it is already doing, what it still cannot judge, and what goes wrong when nobody is watching it.
Artificial intelligence arrived in performance management through the back door. Most organisations did not decide to adopt it. Their review platform shipped an update, and suddenly the system was drafting summaries, scoring sentiment in written feedback and sending nudges when a check-in was overdue. The decision to be made now is not whether AI belongs in the process. It is which parts of the process it should be allowed near.
That distinction matters because performance management contains two very different kinds of work. One kind is administrative: gathering, collating, reminding, formatting, reporting. The other is judgement: deciding what good work looks like in this role, weighing a result against the circumstances it was achieved in, and telling someone the truth in a way they can act on. Machines are already better than people at the first kind. They are not close on the second, and the organisations getting into trouble are the ones that have not drawn the line between them.
Where it is already doing real work
Administration
- Chasing reviewers who have not submitted feedback
- Drafting a review summary from goal records, notes and the previous cycle
- Producing calibration views across teams without spreadsheet work
- Translating feedback for distributed teams
Pattern finding
- Showing where ratings cluster unusually high or low by manager
- Flagging goals that have had no activity for months
- Identifying teams where check-ins have quietly stopped happening
- Comparing the language used in feedback across groups
Manager support
- Suggesting wording when a manager cannot phrase a difficult point
- Matching development content to a stated goal
- Answering routine process questions so HR is not the first stop
- Prompting a conversation before a problem has set hard
None of this is speculative. If your organisation runs performance reviews through a modern platform, several of those functions are probably switched on already. The useful question for an HR team is which of them are being used deliberately and which are running unsupervised.
What it cannot judge
An algorithm sees the record. It does not see the context that produced the record. A sales figure that missed target by fifteen per cent reads the same whether the account was handed over mid-year, whether the territory was restructured, or whether the person carried two roles for a quarter while a colleague was on leave. A manager who was paying attention knows which of those happened. The system does not, unless someone wrote it down, and most of what makes a fair appraisal fair never gets written down.
This is why the strongest use of AI in appraisal is as preparation rather than conclusion. The system assembles the evidence, surfaces what the manager may have forgotten and points at the inconsistencies. The manager then does the part that requires having been there. Reverse that order, and you get a review that is internally consistent, beautifully written and wrong in ways nobody can trace.
Where it goes wrong
The model inherits the bias in the history
A system trained on past ratings learns past patterns, including the ones the organisation would rather not have. If a particular group was historically rated lower, or if one manager's generosity has been baked into years of data, the model treats that as signal. The result looks objective, which is exactly what makes it hard to challenge. Checking for this is an ongoing audit, not a one-off procurement question: compare outcomes by group, by manager and by location, and do it every cycle rather than at launch.
Nobody can explain the output
Employees will accept a rating they disagree with far more readily than one nobody can account for. When a manager has to say the system produced it, the process loses the thing it depends on. Before any AI-assisted output reaches an employee, somebody should be able to state in plain language what evidence it drew on and what weight it gave each part. If that explanation cannot be given, the output is not ready to be used in a decision about someone's pay or promotion.
It gets introduced as an efficiency
Announce AI in performance management as a way to save manager time and you confirm the suspicion employees already hold: that their appraisal has become a processing task. The organisations that have introduced it without damage tended to lead with what it removes rather than what it decides, and were specific about which judgements stay with a person. Vagueness on that point is read as evasion, usually correctly.
What this asks of the people running the process
The skill that matters is no longer writing the review. It is knowing when the system's output is wrong and having the standing to overrule it. That requires understanding what good performance management looks like independently of whatever the software proposes, which is harder than it sounds once a tool has been producing plausible text for a year and nobody has had to argue with it.
It also requires a working knowledge of how these systems reach their conclusions. Not the mathematics, but enough to ask a vendor what the model was trained on, what it does when data is missing, and how an employee contests an output. Those questions tend not to get asked by people who are relieved the paperwork has gone away.
Where the Institute stands on this
Our certification programme teaches the judgement side of performance management: how to set objectives that survive a changing year, how to run an appraisal conversation that changes behaviour, and how to spot when a rating does not reflect what actually happened. Those are the capabilities that decide whether AI in your process is an asset or a liability, because they are what you use to check its work.
Further reading: Artificial Intelligence and its Applications: The game changer in People management, ResearchGate.