AI in talent work
·
6 min read
Key findings
A defensible prompt is recorded with its output, checked by a named person, and written without personal data.
Name roles instead of individuals, because a name attached to performance detail sends identifiable data to a third party.
Statistics, citations, legal statements and claims about research need a check every time, and counts from the model need a recount.
A register with a column for what changed after checking is the evidence that judgement was applied.
A prompt that cannot be shown is a method that cannot be defended.
The best ChatGPT prompts for HR are the ones a practitioner can show, check and defend.
Each prompt names roles and never individuals, is recorded with its output, and is checked by a named person before anything is used.
Five reusable prompts follow, each with the check that applies to it.
Competing lists of HR prompts advertise a count, from a dozen to a thousand.
The test that matters comes later, when a candidate or an employee challenges a decision and someone asks what went into the process.
The Standard To Meet
Machine output is a source, and it meets the same evidence standard as any other source.
Three requirements follow.
The prompt is recorded with the output
A named person checked the output before use
Personal data stayed inside approved systems
An organisation that can produce those three for any decision is in a different position from one that cannot.
Roles, Never People
The most important rule concerns what goes into the prompt.
Name roles and not individuals.
A prompt that asks for development feedback for a head of operations who struggles with delegation is defensible.
The same prompt with a person's name attached has sent identifiable performance data to a third party.
Two further rules follow from the same principle.
Strip detail that would reconstruct the person, including team size, location and tenure in combination
Keep anything involving a named individual inside systems the organisation has approved for personal data
Employee data also carries legal duties.
Kakade and Vispute list informed consent and privacy regulation, including GDPR, among the issues HR has to address before using AI on employee data.
Anatomy Of A Reusable Prompt
Most practitioners write a prompt, read the answer and lose the prompt.
Cavalier reports that prompts perform best when they carry a role, a clearly defined task and detailed constraints.
A reusable prompt for HR work has five parts.
The role the model is to take
The task, stated as an output
The context it needs, and no more
The constraints, including what to avoid
The format, including length
The fourth part is where practitioner judgement sits.
A prompt that says to avoid generic competency language and to use the words a line manager would use produces different text from one that does not.
Five Prompts Worth Keeping
A short library beats a long one, and these five cover most of an internal practitioner's drafting.
Replace each bracketed item, and describe people by role.
Interview questions from a role brief.
"You are an HR adviser writing structured interview questions for [ROLE TITLE] in [FUNCTION]. Here is the role brief: [PASTE BRIEF, WITH NO NAMES]. Write eight behavioural questions, each tied to one requirement in the brief. For each, add one follow-up that tests what the candidate personally did. Avoid generic competency wording and use the words a line manager in this function would use. Return a numbered list."
The check is that every question traces to a requirement in the brief, and any question that screens on something the brief does not mention is removed.
A summary against three questions.
"You are an analyst summarising a long report for an HR director. Here is the report: [PASTE TEXT]. Answer these three questions using only the report: [QUESTION 1], [QUESTION 2], [QUESTION 3]. For each answer, quote the passage it rests on. If the report does not answer a question, say so. Do not add information from outside the report."
The check is to find each quoted passage in the source, because an invented quotation is the usual failure.
Three framings of a problem.
"A business leader has asked for [THE REQUEST, IN THEIR WORDS]. Write three different statements of the underlying problem, each assuming a different cause. For each, name one fact that would show it is true and one that would show it is false. Do not recommend a solution."
The practitioner decides which framing the evidence supports, as problem framing for HR describes, and the machine only widens the options.
A register where nothing was ever changed after checking describes a process where nobody checked.
Free-text comments coded into themes.
"Here are [NUMBER] anonymous survey comments from [ROLE GROUP]: [PASTE COMMENTS]. Group them into no more than six themes. For each theme, give a short name, the number of comments and three comments quoted exactly. List separately any comment that fits no theme."
The check is a recount of a sample, because Cavalier notes that language models are trained mainly on text and can produce plausible but inaccurate figures.
What follows an engagement survey covers what to do with the themes once they are checked.
A stakeholder briefing at two lengths.
"You are an HR business partner writing to [ROLE OF RECIPIENT]. The decision needed is [DECISION]. The facts are [FACTS]. Write the briefing twice, once in 100 words and once in 300. Put the decision and its date in the first sentence. Use only the facts given, and mark any statement you are inferring."
The check is that every number in the output matches the facts supplied.
Where The Output Needs Checking Hardest
Cavalier warns that generative systems can produce information that sounds plausible and is entirely made up.
He says fact-checking should be a critical part of any AI-integrated workflow.
Four categories deserve a check every time.
Any statistic, which is traced to a source or removed
Any citation, because plausible-looking references are sometimes invented
Any legal or regulatory statement, which belongs with the people who own it
Any claim attributed to research
Any count or total produced by the model is recounted, for the reason given under the comment-coding prompt.
The Judgement Problem
Dell'Acqua gave 181 professional recruiters different levels of AI support for evaluating 44 job applications.
Recruiters with higher-quality assistance performed worse than those with lower-quality assistance.
They spent less time and effort on each application, followed the recommendations without challenge, and did not improve over time.
The design of the workflow has to allow for that effect, and where AI belongs in talent work sets out the evidence and the countermeasures in full.
Three practices from it apply to every prompt above.
Write one line of your own view before opening the tool, on anything concerning a person
Sample a share of outputs for independent review
Report the disagreement rate between the machine and the reviewer
What Not To Delegate
Four pieces of work fail the standard, and no prompt repairs them.
Ranking candidates. Strong assistance degrades the judgement it is meant to support
Writing talent review evidence. The record is valuable because a named person committed to a claim
Drafting words for a hard conversation. A manager reciting unfamiliar phrases loses the listener
Inferring potential from performance data. A model trained on past decisions reproduces them
Checking Received Documents
Practitioners now receive documents drafted with these tools, from vendors, colleagues and candidates.
Three checks are quick enough to make routine.
Verify every number and every citation before repeating either
Read for specificity, because the common failure is fluent text that could describe any organisation
Ask the sender which parts they wrote, which is a reasonable question and shows where to look hardest
Writing The Policy Line
Most organisations asking whether HR needs an AI policy need three sentences before they need a document.
The machine drafts and the practitioner decides, and the practitioner is named on the output.
Prompts name roles and never individuals, and no personal data leaves approved systems.
Machine output meets the same evidence standard as any other source, which means it is attributed and checked.
Those three can be audited, and they are short enough for people who will never read a policy to remember.
Cavalier notes that organisations now also have to work with AI-specific regulation alongside data protection law, so the legal owner should review the wording.
The Register
The record that matters when a decision is challenged is a register and not a policy.
Four columns are enough.
The decision, and its date
Where machine input was used
Who checked it, by name
What was changed after checking
The fourth column shows that judgement was applied.
A register where nothing was ever changed after checking describes a process where nobody checked.
What To Measure
Licence counts and usage rates say nothing about value or risk.
Three measures say more.
Time saved on a named task, measured once properly
Disagreement rate between the machine and the reviewer
Decisions taken with machine input, by type
Run one measured comparison on one real task before claiming a productivity effect.
Cavalier reports a client that cut training development time by 60 per cent with generative AI, which is one result and not a benchmark.
Where The Tools Sit
Trusted Talent Advisor ships with The Prompt Library, which holds a prompt for each of the 51 tools and five more for the diagnostic lenses.
Every prompt describes people by role and not by name, and the library states what not to paste into an external tool.
Related reading: AI in instructional design, AI agents and accountability.
Sources
1. Cavalier, J. Applying AI in Learning and Development. ATD Press.
2. Mollick, E. Co-Intelligence: Living and Working with AI. Portfolio.
3. Kakade, K. S. and Vispute, S. A. AI Dynamics in HRM. Apple Academic Press and CRC Press.
From the system
Trusted Talent Advisor sets out the full method: seven components, 51 tools, The Talent Template Pack and The Prompt Library.
See the system
Questions this answers
What are the best ChatGPT prompts for HR?
The best ones are specific, reusable and defensible: interview questions from a role brief, a summary against named questions, three framings of a problem, comments coded into themes, and a briefing at two lengths. Each names roles and is checked.
Is it safe to put employee data into ChatGPT?
Not personal data. Describe people by role, strip detail that would identify them, and keep anything about a named individual inside systems the organisation has approved. Check your employer's position first, and assume that anything typed may be retained.
How should HR write an AI prompt?
State the role the model takes, the task as an output, only the context it needs, the constraints including what to avoid, and the format and length. Record the prompt with the output so that it can be shown later.
Do we need an AI policy for HR?
Start with three sentences: the machine drafts and the practitioner decides, prompts name roles and no personal data leaves approved systems, and machine output meets the same evidence standard as any source. The legal owner then reviews the wording.
Can AI write a performance review?
It can organise a manager's own dated notes into a first draft. The rating and the example behind it stay with the manager, who checks the draft against the notes before anyone sees it.
How do you check AI output for accuracy?
Trace every statistic to a source, confirm each citation exists, send legal statements to their owners, test claims about research, and recount any figures the model produced. Record what changed after checking.
Keep reading
Measuring HR impact
Internal Mobility Metrics Without A Platform
Seven measures an HR adviser can build from existing records, each with a stated definition, a base and a way to read it.
7 min read
Workforce and future skills
Job Rotation As A Talent Gap Strategy
A job rotation programme closes a talent gap when the role, the person, the cover and the evidence are agreed with both managers before it starts.
8 min read
Talent reviews and succession
Seeing Talent Below The Top Team
Visibility of leadership potential drops below the top team. A cross-unit forum, run on one page per unit and a written evidence standard, is the method for seeing it.
8 min read