Where AI Helps in HR, and Where It Must Not

Two piles, and the line between them

HR work splits into two piles that look similar from the outside and are completely different underneath.

The first pile is writing and organising. Job adverts. Policy summaries. Interview question sets. Onboarding checklists. The fifth version of the same email about the new expenses process. This work is repetitive, low in judgement, and eats an astonishing amount of your week. AI is genuinely good at it, and if you use it well you will get hours back.

The second pile is decisions about named people. Who gets hired. Who gets promoted. What a performance rating says. Whether a disciplinary case is upheld. Whether someone's role is at risk. These decisions change a person's income, their standing at work and sometimes their ability to pay rent. They also have to be defensible later, sometimes years later, in a room where nobody cares how efficient your process was.

The whole course rests on one principle: AI drafts, humans decide, and a human must be able to explain any decision in their own words.

That last clause is the real test. If you cannot explain a rejection, a rating or an outcome without saying the system suggested it, you have crossed the line. Not because a rule forbids it, but because you have made a decision about a person that you cannot account for.

In plain English

Automated decision:: An outcome produced or substantially driven by software rather than a person exercising judgement.
Human in the loop:: A real person reviews and can overturn what the system suggested, with the time and information to do it properly.
Adverse impact:: A process that filters out one group at a noticeably higher rate than another, even when nobody intended it.
Audit trail:: The record of who decided what, when, and on what basis. What you rely on when a decision is challenged.

What is genuinely safe and useful

Drafting job adverts from a role brief. Summarising a long policy document into something a manager can act on. Turning a job description into a set of structured interview questions. Building onboarding material. First drafts of routine communications, the kind nobody would ever describe as sensitive.

What these have in common is that no individual is the subject. You are producing material, not judging a person. If the output is wrong, you notice and fix it before anyone is affected by it.

Prompt you can copy: summarise a policy for managers

Summarise the policy below for line managers who have five minutes.
Audience: managers with no HR background
Output: under 250 words, plain English, no legal phrasing
Cover: what the policy requires them to do, what it forbids,
and what they must escalate rather than handle themselves
End with: the three points managers most often get wrong about this
Do not add anything the policy does not say. If something is unclear
in the source, list it under a heading called Unclear rather than guessing.
Policy:
[paste it]

What must stay human

Hiring and firing decisions. Performance ratings. Disciplinary and grievance outcomes. Redundancy selection. Anything at all about a named individual's future.

This is not squeamishness. Three practical reasons hold it up.

You cannot explain it. These tools produce fluent reasoning that sounds convincing and may have very little to do with how the output was actually produced. Repeating that reasoning as your own is not an explanation, it is a costume.

It goes wrong at scale. One tired manager misjudging one candidate is a bad afternoon. One screening rule applied to four thousand applicants is a pattern, and patterns are what regulators and tribunals look for.

People can tell. Being processed by software at a moment that matters to you is its own injury, quite separate from the outcome.

Nothing in this course is legal advice. Employment and hiring law varies enormously between countries, states and sectors, and it changes. Before you automate any part of a process that affects candidates or employees, speak to your own legal or compliance function.

❌ Weak prompt

Prompt

Here are 40 candidate CVs. Rank them best to worst for this role and tell me who to reject.

Output

A ranked list of 40 names with scores from 9.1 down to 3.4, plus a recommended rejection threshold.

Confident, fast and indefensible. You cannot explain any single position on that list, and the ranking may simply reflect which CVs sound like the people who already work there.

✅ Good prompt

Prompt

Here is a job description. Produce a shortlisting checklist with 6 criteria drawn only from the essential requirements. For each criterion, describe what strong evidence looks like and what weak evidence looks like. Do not score anyone. I will apply this myself to each application.

Output

Six criteria with paired evidence descriptions, written plainly enough that two different reviewers would apply them the same way.

The AI improved the instrument. A human still applies it, and can explain any decision by pointing at the criterion and the evidence in front of them.

Checkpoint

AI drafts, humans decide. If you cannot explain a decision about a person in your own words, without referring to what a tool suggested, you should not be making it that way.

The test to apply on Monday

Before using AI for any HR task, ask three questions.

Is a named individual the subject of this output? If yes, slow down.

Would I be comfortable showing this prompt and this output to the person affected? If not, that discomfort is information.

If this is challenged in twelve months, could I explain the basis of the decision without mentioning software? If no, redesign the step.

Prompt you can copy: pressure-test a proposed AI use

I am considering using AI for this HR task:
[describe the task in two or three sentences]
Tell me:

Whether any output would be about a specific named individual
What could go wrong for a candidate or employee if the output is wrong
Which parts a human must do rather than merely review
What record I should keep so the decision can be explained later
Be blunt about the risks. Do not reassure me. If this task should not be
automated at all, say so first and explain why.

Prompt you can copy: draft a week one onboarding plan

Draft a week one onboarding plan for a new [job title] joining [team].
Include: what they need before day one, who they meet and why,
what they should be able to do by Friday, and what nobody should
expect of them yet.
Plain language. No corporate warmth, no exciting journey phrasing.
Keep it to one page. Mark anything that needs a name or a date from me
with square brackets.

A useful habit: keep the prompt alongside the output for anything that feeds a process affecting people. Six months later, the prompt is often the only thing that explains why the output looks the way it does.

📝 Quiz

Question 1 of 4

What is the core principle this course rests on?