AI in talent work
·
7 min read
Key findings
Korn Ferry reports that 40 per cent of CHROs name insufficient AI skills in HR teams as the biggest obstacle.
Five skills carry across tools: framing a task, checking evidence, spotting fabricated sources, protecting data and keeping decisions with a person.
A work sample with planted faults shows where each role stands, where a self-rating records only confidence.
A four-week sequence on real work, with a named checker for every output, gives a before and after comparison.
Recruiters given better AI support spent less effort and followed its recommendations without challenge, so training has to reward checking.
A self-rating records confidence, and the risk this plan guards against is confidence that runs ahead of checking.
The AI skills for HR professionals that matter first are five working habits. They are framing a task for a model, checking an output against evidence, spotting a fabricated source, handling confidential data and knowing which decisions a person must keep.
A four-week sequence on real work, with a named checker for every output, builds them.
Tool training covers one product and ages within months, while these five habits carry across any product an HR team is sold.
The Gap In Numbers
Korn Ferry reports that 40 per cent of CHROs name insufficient AI-related knowledge and skills within HR teams as the biggest obstacle to integrating AI into talent management.
The same page says that only 5 per cent of HR teams feel fully prepared to implement AI effectively.
It names no survey base for either figure, so both show the direction of opinion among HR leaders and nothing finer.
Gartner's survey of CHROs, built on 750 responses collected in May 2026, ranks HR technology and AI strategy as the second priority, behind leader and manager development.
Within that priority, lack of skills is the top challenge, cited by 45 per cent of CHROs.
Use has spread quickly. Gartner surveys of 197 HR leaders, in June 2023 and January 2025, found organisations piloting or implementing generative AI rising from 19 per cent to 61 per cent.
These figures come from different surveys with different bases, so they cannot be added together.
Read together, they suggest use spreading faster than the skill to check what comes back.
Skills Before Tools
Five skills transfer from one product to the next, and each can be practised on work an HR team already does.
Framing a task. State the output wanted, the context the model needs and no more, the constraints and the format, so that a colleague could rerun the instruction.
Checking against evidence. Set each claim in an output beside the policy, the data extract or the source document it should rest on.
Spotting a fabricated source. Open every citation and every named study before repeating it, because these systems produce plausible references that do not exist.
Handling confidential data. Know which items may enter an approved tool, which may not, and which need advice first.
Keeping decisions with a person. Know which calls about people stay with a named practitioner, whatever the tool offers.
Mollick records that large language models readily make up plausible but wrong citations and facts, and that a system can justify an answer it has already committed to.
He once had to search for the studies cited by an AI version of himself to confirm that they were fake, because they seemed plausible enough to be his own.
Cavalier names the same hazard for learning content, where unvetted fabrication can mislead learners.
He adds that employees who put sensitive information into public models may share corporate secrets without noticing.
Mollick calls the edge of AI capability a jagged frontier, which is invisible and on which tasks of similar apparent difficulty sit on opposite sides.
He argues that experimenting on everyday tasks a person knows well is how that person finds where the frontier lies.
Assessing Each Role
A self-rating records confidence, and the risk this plan guards against is confidence that runs ahead of checking.
Recruiters in Dell'Acqua's study who had better AI support spent less time and effort on each application, followed its recommendations without challenge and did not improve.
A work sample gives evidence instead. Each person completes three tasks in about an hour, set by the HR leader or the learning lead from material the function already holds.
Brief. Write the instruction for a task from your own desk, scored one point each for output, context, constraints and format.
Fault finding. Review an output with three planted faults: a clause that contradicts the policy, a figure that does not follow from the data supplied, and a citation that does not exist.
Data sort. Sort twenty items, from a job description to a named grievance file, into permitted, not permitted and needs advice.
A colleague scores each task as not yet, with a checker, or alone, and the grid of roles by tasks shows where to begin.
First Priorities By Role
Each role starts with the skill its work exposes most, on a task that already recurs.
HR business partner. Checking against evidence, on the briefing for a leader meeting, with every figure traced to the system extract it came from.
Talent partner. Keeping decisions with a person, on interview questions drafted from a role brief. Ranking candidates and inferring potential stay with people, for the reasons in where AI belongs in talent work.
Learning and development. Spotting fabricated sources, on draft learning material. Each statistic and research citation is traced before use, as set out in AI in learning design.
Reward. Handling confidential data, on job family descriptions and published survey summaries that contain no individual salary, with all arithmetic recalculated by hand.
HR operations. Framing a task, on letters and answers to employee queries drafted from approved policy text. A checker sees a fault in one reading, and legal statements go to the people who own them.
Each role starts with the skill its work exposes most, on a task that already recurs.
Four Weeks Of Practice
Coyle sets out Ericsson's definition of deliberate practice as working on technique, seeking constant critical feedback and focusing on weaknesses.
He advises choosing a goal just beyond present abilities, and quotes Bjork on an optimal gap between what a person knows and what the person is trying to do.
The named checker supplies the feedback, and the planted faults and harder tasks supply the gap.
Take an HR team of eight: a director, two business partners, a talent partner, a learning lead, a reward analyst and two operations specialists.
Pair each person with a checker who could have produced the work: a business partner who has used the same pay data can check the reward analyst.
The sequence asks for about ninety minutes a week inside work that would be done anyway.
Week one. Each person picks one recurring task, writes the brief and runs it. The checker logs every fault found.
Week two. The same task under a harder standard, with every number and citation traced to a source the person has opened, and the checker re-tracing a sample.
Week three. On any task touching a person, the practitioner writes a one-line view before opening the tool, then records where the output disagreed.
Week four. Each person does one task without assistance, then repeats the work sample with a different set of planted faults, and the pair agrees the next task.
The checker's log has four columns: the claim checked, the source opened, whether it held, and the minutes taken.
Faults that recur across the team become the planted faults for the week-five work sample.
Skills For The Second Cycle
Four further skills follow once the first five are steady.
Data literacy. Read an AI-produced summary of survey or people data and recalculate one figure by hand.
Bias testing. Compare outcomes across groups for any tool that screens or scores people.
Employment rules. Rules on automated decisions vary by jurisdiction and belong with employment counsel.
Tool selection. Put the vendor questions below to every product before purchase.
Each depends on the first five, because a test run on an unchecked output inherits its faults.
Measuring The Result
Four measures come straight out of the sequence, and each can be reported to a leadership team.
Faults found in the work sample at week five against week one
Disagreement between the practitioner's first view and the output
Outputs used with the prompt recorded and a named checker
Confidential data incidents, including use of tools not approved, reported by the team itself
Time saved counts only on outputs that passed the check, and satisfaction scores are left out because they measure confidence.
The measure, the baseline and the owner are agreed with the sponsor before week one, as described in training evaluation.
The leadership team receives the week-five comparison and the list of decisions that stay with a person.
Its question is whether the function can be trusted with AI on employee matters.
Confidence Without Checking
Mollick reports that consultants using AI on a task outside its frontier got it right 60 to 70 per cent of the time, against 84 per cent without the tool.
Dell'Acqua, in the recruiter study, named the pattern falling asleep at the wheel: when the AI is very good, people stop working hard and stop improving.
A programme that rewards completions, speed and satisfaction rewards that behaviour. Three design rules reverse it.
Include a planted fault in every exercise
Score the checker as well as the producer
Keep unassisted work in rotation so the skill stays in use
The full evidence is in where AI belongs in talent work, and recording the prompt with each output is covered in prompts a practitioner can defend.
Questions For Vendors
Vendor training teaches the vendor's product. Six questions test whether it also teaches the five skills.
Is anything entered used to train the vendor's models, and where is it stored
Which outputs show their sources, and can a person open each one
Can a person override a recommendation, and is the override recorded
Which decisions does the vendor's own documentation say need a person
Does the training include faulty outputs, and does it assess the checker
What result has the vendor measured at a comparable customer, and on what base
Cavalier recommends a weighted scoring model when selecting a platform, which keeps the decision on the function's needs and off the vendor's promises.
The weights are set before the first demonstration.
First Step
Build the three-task work sample this week and run it with the team.
Its results decide which role starts with which skill, and they give the week-five comparison its baseline.
Where The Tools Sit
Trusted Talent Advisor ships with The Prompt Library, which holds a prompt for each of the 51 tools and states what not to paste into an external tool.
Each person can write the week-one brief against one of those prompts, with people described by role and not by name.
Sources
1. Korn Ferry. HR Trends to Watch in 2026. Korn Ferry.
2. Gartner C-level Communities. Top 3 Priorities for CHROs in 2026, Leadership Perspective Survey, May 2026. Evanta.
3. Gartner. Gartner Announces Top Trends Shaping HR Priorities in Australia in 2026, press release of 17 November 2025. Gartner.
4. Mollick, E. Co-Intelligence: Living and Working with AI. Portfolio.
5. Dell'Acqua, F. Falling Asleep at the Wheel, cited in Mollick, on recruiter performance with AI assistance. Harvard Business School working paper.
6. Cavalier, J. Applying AI in Learning and Development. ATD Press.
7. Coyle, D. The Talent Code: Greatness Isn't Born. It's Grown. Here's How. Bantam Books.
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 AI skills do HR professionals need?
Five skills carry across tools: framing a task for a model, checking an output against evidence, spotting a fabricated source, handling confidential data, and knowing which decisions a person must keep.
How can HR teams upskill in AI?
Run four weeks of practice on real work. Each person writes a brief for a recurring task, a named checker logs every fault, and the sequence ends with a repeat work sample on a new set of planted faults.
How do you assess AI skills in an HR team?
Use a work sample instead of a self-rating. Each person writes a brief, reviews an output with three planted faults and sorts twenty items by data permission. A colleague scores each task as not yet, with a checker, or alone.
What AI skills should HR learn first?
It depends on the role. An HR business partner starts with checking against evidence, a talent partner with keeping decisions with a person, learning and development with fabricated sources, reward with confidential data, and HR operations with framing a task.
Do HR professionals need technical AI skills?
The five skills described here need no programming. They need an approved tool, tasks the function already does, and a checker who could have produced the output without the tool.
How long does it take to train an HR team in AI?
The sequence here takes four weeks at about ninety minutes a week, inside work that would be done anyway. It gives a first measured comparison, and a further four-week cycle can follow on a harder task.
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