HR: AI Agents
Recruitment screening, workforce analytics and personalised learning, with the legal and ethical limits that make this the most constrained function in the framework.
Markdown. No sign-up, no email.
This is the most constrained function here, and deliberately so. Every other department's automation mistakes cost money or time. This one's cost people their livelihoods, and they do it in ways that are hard to see and harder to reverse.
Everything below assists a human decision. Nothing below makes one.
AI recruitment#
| Capability | What it does | What it must not do |
|---|---|---|
| CV screening | Surfaces relevant experience against written criteria | Reject anyone. It orders a queue; a person reads it |
| Candidate matching | Suggests fit against a role's stated requirements | Infer anything protected, directly or by proxy |
| Interview assistance | Prepares structured questions and notes what was covered | Score a person, or judge them from video |
| Skill assessment | Marks work samples against a published rubric | Be the only input to a decision |
No automated rejection. A person reviews every rejection, including the ones the system ranked lowest. This is slower and it is the price of the tool.
Historical hiring data encodes historical hiring bias. A model trained on who we hired before learns who we overlooked before, and it will reproduce that pattern with more consistency than any individual manager ever could. The mitigation is to score against written criteria for the role, never against similarity to existing employees.
Proxies are the trap. Removing names and photographs does not remove bias when postcode, university, career gaps or a language pattern carry the same signal. Audit outcomes by group, not the inputs.
AI workforce analytics#
| Measure | Legitimate use | Illegitimate use |
|---|---|---|
| Productivity | Spotting a team drowning or a process broken | Ranking individuals |
| Skills gaps | Planning training and hiring | Identifying people to remove |
| Training needs | Building learning paths | Performance evidence |
| Workforce planning | Capacity against roadmap | Justifying a decision already made |
Aggregate, not individual. These measures identify where the organisation needs help. Turned on a person, they become surveillance, and they are poor at it: they measure what is easy to count, which is rarely what matters about someone's work.
Nobody is managed out on a metric. If the data suggests a problem with an individual, that is a prompt for a conversation, not evidence.
AI learning#
Personalised paths built from role, current skills and where the person wants to go. The last of those three is the one that makes it work, and it comes from asking rather than inferring.
Genuinely well suited to automation: everyone needs something different, and building sixty paths by hand is why most companies build one and call it onboarding.
What stays with a person, always#
- Hiring. Every offer, every rejection.
- Termination. Not a capability question, an accountability one.
- Pay and promotion.
- Performance judgement. Assisted by evidence, never produced by a model.
- Anything involving a protected characteristic, and anything that might correlate with one.
- Investigating a complaint. A person, every time, with due process.
The compliance position#
Checked 9 August 2026. AI in recruitment and workforce management is the area regulators are watching most closely, and rules differ sharply by jurisdiction. Some require candidates to be told AI was used, some require bias audits, some restrict automated decisions about people entirely.
Two rules that hold regardless of where the law lands:
- A person can always ask how a decision about them was made, and get a real answer. If the answer is "the model ranked you low", the process is wrong regardless of its legality.
- Outcomes are audited by group, quarterly. Not inputs, outcomes. Bias does not announce itself in the design; it shows up in who gets through.