Every HR platform is bolting AI onto appraisals, and the demos are seductive. Paste in a year of feedback, get back a tidy paragraph on where someone is strong and where they need support. It looks like the end of the worst week in the HR calendar.
Then you connect it to your own company and it has almost nothing to read.
That is not a failure of the model. Language models are genuinely good at this kind of work: compressing a lot of text, noticing what three people said in different words, drafting a first pass a human then fixes. What they cannot do is invent the evidence. If your review process produces one manager’s notes, once a year, written the night before the meeting, then a summary of it is a tidier version of one person’s opinion — and the tidiness makes it more convincing, not more true.
The input problem
Ask what a model would actually be reading. In most companies the honest answer is a single rater, no comparison, and a scale nobody agreed on. Was that four out of five generous? Did "a great team player" mean the same thing from the two managers who wrote it? Nothing in the data says, so nothing in the summary can either.
Worse, the gaps are invisible. A model given thin evidence does not refuse — it writes you a confident paragraph. The output looks identical whether it was built from forty considered answers or four rushed ones.
What a model would need to be worth having
- More than one rater on the same person, answering the same questions — manager, peers and direct reports, not a single point of view.
- A competency library mapped to roles, so "leadership" means a defined level for that grade rather than a word everyone reads differently.
- The self-assessment alongside the reviewer scores, so blind spots and undervalued strengths are in the data rather than in a conversation nobody recorded.
- Calibration, so a four from one panel means roughly what a four from another panel means.
- A history — the same person, the same competencies, cycle after cycle. Most of what is useful about a review is the shape of the line, not this year’s dot.
None of those are AI features. They are process features, and they are the thing to fix first — whether or not you ever point a model at the result. They are also exactly what a 360° cycle in Hubb People produces: multi-rater campaigns against one question set, a self-versus-reviewers radar per competency, calibrated final ratings, and the same structure repeated every cycle so the comparison holds.