Artificial Intelligence in HR

HR teams are surrounded by tasks that look ready for AI: answering policy questions, drafting candidate emails, summarising survey comments and preparing learning plans. Some are sensible places to start. Others carry consequences that a quick demonstration hides, because a wrong answer or an unfair summary lands on a real person. The useful question is not whether a tool can produce an output. It is whether it should help with this task, using these inputs, under these controls. This chapter gives you a repeatable way to make that call. You will map tasks by volume, variability, error cost and judgement, then decide where people must stay in charge. You will check generated answers against approved sources before anyone acts on them. Finally, you will set data boundaries and a human review point that works in practice, not just on paper.

Start with the task, not the tool

An HR task is a repeatable piece of work with a clear input and output. "Improve recruitment" is too broad to assess. "Draft a first reply to candidates asking about interview times" is specific enough, because it has an input, a desired result and a person who can check it. Map each proposed task against four factors. Volume is how often the work happens. Variability is how much the right answer changes from case to case. Error cost is what happens when the output is wrong, incomplete or unfair, and who pays for it. Judgement is whether the work needs context, values or a decision about a person. Treat these as prompts for a disciplined conversation, not a scoring formula. A single total can hide one serious risk.

Compare benefit with consequence

High volume does not make a task a good candidate on its own. The benefit might be faster drafting, more consistent wording or easier access to approved information. Weigh that against the effort of review and the cost of failure. Where consequence is high, AI may still help with clerical preparation, but it must not quietly become the decision-maker.

Apply the map to a real workflow

Funmi, a talent acquisition partner in Lagos, fields dozens of applicant questions each week about interview logistics. A tool could draft replies from the approved schedule and venue guide. The work is frequent, the facts are stable and a recruiter can check each reply in seconds. A wrong time could still make a candidate miss an interview, so nothing goes out unchecked. Her manager then asks whether the same tool could rank applicants from their CVs. That output shapes who gets an opportunity. CVs vary widely, and weighing evidence against role criteria takes care. The task is just as frequent, but the error cost is far higher, so the boundary moves. Funmi records the first task as a bounded trial and keeps the second as a human-led process.

• Name the task, its user, its input and its intended output.
• Estimate volume from workflow records, not memory alone.
• Describe one plausible error and who would bear its cost.
• Choose a bounded trial, limited assistance or no AI use.

A fluent answer is not evidence

Generative AI can produce confident text that contains unsupported details, blends rules from different places or presents an old policy as current. In HR, polished wording is persuasive because it sounds like a complete explanation. Treat every generated answer as a draft to investigate, not proof of what a policy says. Checking means tracing each material claim to your source of truth: the approved, current record for that fact. Depending on the question, that might be a policy, a benefits guide or a documented process. A colleague's memory, an old slide deck or a citation the tool supplies is not automatically authoritative. Confirm that the source applies to the right employee group, location and date.

Trace claims, not links

Read an answer as a set of separate claims. For each factual statement, ask three things. Where does this appear in the source? Does the wording keep its meaning? Are there conditions or exceptions the draft left out? A citation can point to a real document and still fail to support the sentence beside it.

• Confirm the exact question and who is asking it.
• Open the current approved source, not a snippet or search result.
• Compare the draft with the relevant passage, sentence by sentence.
• Check scope, dates, definitions and exceptions.
• Correct, qualify or remove anything the source does not support.

If two sources conflict, stop and ask the policy owner rather than picking the version that reads better. For high-impact content, such as interview guidance or a summary of grievance notes, inspect the underlying material itself, because a summary can drop the one qualification that changes the meaning.

Work through a policy question

Niamh, a People operations partner in Dublin, uses an approved assistant to draft a reply about a learning allowance. The draft says every employee can claim any course fee and that approval is automatic. She opens the current guidance before sending anything. The guidance says employees may request support for eligible learning, subject to manager review. It says nothing about every course qualifying or approval being automatic. Niamh removes both claims, states that eligibility depends on review and asks the programme owner to confirm one unclear term. Her reply is less sweeping and far more useful, because it reflects the real policy rather than the tool's confident phrasing. She also logs the error, so the team can see whether the same mistake keeps appearing. Where practical, she shows the source date and version in the reply, because an accurate passage from an obsolete guide can still mislead.

Write the task boundary first

Before a trial, describe the task's boundary in plain language: purpose, permitted inputs, expected output, prohibited uses and the named reviewer. This turns an idea into something you can test, and it gives colleagues a shared picture of what the tool does and does not do. Apply data minimisation: use only the information the stated purpose needs. Ask whether the question can be answered without identifiable employee or candidate details. If personal data is needed, confirm the tool and the use are approved through your privacy, security and technology channels. Never paste confidential HR material into an unapproved service, however convenient it is.

Make human review real

Human-in-the-loop design only works if the reviewer has the knowledge, time and authority to challenge the output. Someone who clicks approve without opening the evidence is not a control. Automation bias, the pull to trust a system's suggestion more than its evidence deserves, makes this failure common. Say what the reviewer must check, what would make them reject the output and where they take their doubts. Place review before an output is sent, recorded or used in a decision about a person. Hiring, promotion, pay, performance action and access to opportunity stay human decisions. In the EU, GDPR Article 22 gives people the right not to be subject to a decision based solely on automated processing that significantly affects them.

Boundary checklist for any trial:

• Purpose: the work problem and the benefit you expect.
• Inputs: allowed material, excluded data and the approved tool.
• Output: format, audience and what it must never claim.
• Review: the named person and the checks they perform.
• Monitoring: errors, complaints and the conditions for pausing use.

Set boundaries for a feedback summary

Rohan, an L&D partner in Pune, wants AI to group comments from a course evaluation into themes. He confirms the tool is approved, strips names and unnecessary details, and limits the input to feedback collected for that course. The permitted output is a draft theme list with de-identified excerpts, never a judgement about a learner or trainer. Rohan and a colleague compare each theme with the original comments and remove anything unsupported. If a comment raises a conduct or wellbeing concern, they route it to the right process instead of letting the summary absorb it. When the data, audience or purpose changes, they revisit the boundary before carrying on. They also agree a review date, so the trial cannot drift into permanent use without a fresh look.

Checkpoint: a proposed screening assistant

What good looks like

You can explain why a task suits AI assistance by weighing volume, variability, error cost and the need for judgement. You trace material claims to a current, approved source before anyone acts on them. You name permitted inputs and an acceptable output, and you place a capable reviewer before any output affects a candidate or employee. When the evidence shows a boundary is failing, you change it rather than hoping reviewers will catch the gap.