How to Use AI for Interviews and Onboarding

Preparing a fair process is not the same as judging people

This distinction carries the whole lesson, so it is worth stating plainly before anything else.

Using AI to prepare a fair process means building better questions, writing clearer scoring guides, and producing onboarding material. The output is an instrument. Humans then use it, and humans remain answerable for what they conclude.

Using AI to judge people means letting software rate a candidate's answers, read their face, infer their personality or produce a score that decides their future. The output is a verdict about a named individual, and you cannot explain how it was reached.

The first is one of the best uses of AI in HR. The second is where the serious trouble lives.

Structured questions, built from the actual job

Unstructured interviews are where bias does its best work. Everyone conducts a slightly different conversation, everyone forms an impression in the first two minutes, and the notes afterwards are a reconstruction of why that impression was correct.

Structure fixes a surprising amount of this, and structure is admin, which means AI can help you build it quickly.

Here is the job description:
[paste it]
Write 8 interview questions tied directly to the essential requirements.
Rules:

Every question must map to a named requirement. Say which one.
Ask about what the person has actually done, not what they would
hypothetically do
No brainteasers, no culture fit questions, no questions about
personality, hobbies, family or personal circumstances
Each question must be answerable by someone from a different sector
who has done comparable work
Add one follow-up probe per question for thin answers
Then list any question you were tempted to write and rejected, and why.

The last instruction is worth keeping. It surfaces the tell me about yourself and the where do you see yourself in five years, and gives you the reason they are weak.

Scoring guides that humans apply

A scoring guide written before the interviews is one of the most effective fairness tools available, and almost nobody writes one, because it takes an hour and the interviews are on Thursday. That hour is now fifteen minutes.

For each of these interview questions:
[paste your questions]
Write a scoring guide with three levels: does not meet, meets, exceeds.
For each level, describe what the content of the answer contains.
Describe evidence, not impressions.
Rules:

Never reference confidence, enthusiasm, polish, communication style,
eye contact or how the person came across
Two different interviewers should reach the same level from the same answer
Keep each description under 30 words
Add one line at the top reminding interviewers to score each question
before moving to the next.

Note what the rules exclude. Confidence and polish track background, schooling and first language far more reliably than they track competence, and they are the easiest thing in the world to mistake for ability.

Onboarding, where AI is simply useful

After the caution, some relief. Onboarding is mostly material, and material is what these tools are good at.

New starter guides. Week one plans. The explainer that turns your benefits documentation into something a human can read. Role-specific reading lists. The thirty questions every new joiner asks in month one, answered once properly.

Two rules keep this safe. Never paste identifiable employee information into a consumer AI tool, so no names, no salaries, no records. And have a human write the welcome, because a personal message from a real manager is the one part of onboarding that cannot be produced in bulk and is the part people remember.

Here is our internal documentation on [benefits / systems / ways of working]:
[paste it, with names and any personal details removed]
Write the 20 questions a new joiner most commonly has in their first month,
with an answer to each drawn only from this documentation.
Rules:

Plain English, two or three sentences per answer
If the documentation does not answer a question, write
Not covered, ask [role] and do not guess
No welcome language, no enthusiasm, just answers
Group the questions under simple headings.