Vendors call everything "AI". Sorting each pitch into one of three kinds tells you what it can do, what can go wrong and how much control it needs.
1. Predict: scores and forecasts
Models trained on historical data estimate what is likely: which candidate matches a job, who is at risk of leaving, how many people a team will need. This is the oldest kind of HR AI. Its big risk is bias learned from the past — if past decisions were unfair, the model repeats them.
2. Draft: generate and summarize
Generative AI writes and condenses: job descriptions, interview guides, onboarding plans, answers to policy questions, summaries of engagement-survey comments. The risk is wrong or generic content, and leaking personal data into tools that are not approved.
3. Act: agents
An agent is AI connected to tools so it can take steps on its own: answer an employee's question and open a ticket, schedule interviews, update a record, send an offer letter prepared by a recruiter. Agents multiply value and risk, because mistakes reach people at machine speed.
Why the difference matters
| Kind | Typical HR use | Main risk | Control you need |
|---|---|---|---|
| Predict | Candidate ranking, attrition risk | Bias, opaque logic | Bias testing, notice, human decision |
| Draft | Job ads, policy answers | Errors, data leaks | Review before use, approved tools |
| Act | HR service agent, scheduling | Errors at scale | Limits, escalation to humans, logs |
Where the market is going
Leaders expect to use more "digital labor". In Microsoft's 2025 Work Trend Index, 82% of leaders said they expect to use digital labor to expand their workforce in the next 12 to 18 months. That puts the third kind — agents — onto HR's agenda: who manages them, who is accountable, and what they mean for jobs.