AI in talent work
·
5 min read
Key findings
Assistance is strongest in development and weakest in analysis, which is the opposite of where most functions struggle.
A learning asset carries the organisation's authority, so a fabricated statistic inside it is the organisation's statement.
The tasks now automated are the ones junior designers learned on.
A function that halves its production time and skips the analysis has produced more of something nobody needed.
AI in instructional design works best as a drafting assistant. The machine produces objectives, scripts, storyboards and quiz items in minutes, and the practitioner decides whether the need is a learning need, checks every fact, and owns what is released.
A learning team that can draft faster still has to settle the analysis, the measure and the transfer plan, and no tool settles those.
The sections below set out the stages of a design project, the prompt sequences that work, the checks each asset needs and the governance line a sponsor can read in a minute.
Design Stages And Who Decides
Instructional design work divides into five stages, and generative AI helps each one by a different amount.
Analysis. The machine can summarise interview notes and survey comments. The practitioner decides whether the gap is capability at all
Design. The machine proposes objectives and a structure. The practitioner agrees the measure and the baseline with the sponsor
Development. The machine drafts scripts, storyboards, slides and quiz items. A named reviewer checks each one before release
Delivery. The machine can answer learner questions and reformat content. The facilitator still runs the live session
Evaluation. The machine can tabulate responses. The practitioner decides what the results do and do not show
The help is largest in development and smallest in analysis, which is the reverse of where most learning teams struggle.
Most teams already produce courses faster than their sponsors can use them, and their difficulty is knowing which course to build.
Put that to the sponsor before any claim about efficiency is made.
Analysis Before Generation
A request for training is usually a request for a solution, and the first task is to establish what performance the sponsor wants changed.
Performance consulting sorts causes into four levels: the business result, the performance on the job, the work environment and the capability of the people.
Cathy Moore's action mapping starts from the same place, with a measurable business goal and the actions that would reach it, and then asks whether each action is held back by knowledge, skill, motivation or environment.
Cavalier, who took a week-long performance consulting course in 1997, reports that these questions have stayed useful now that AI can process performance data in seconds.
Generative AI can write a course on any topic it is given, including a topic where training will change nothing.
The practitioner therefore settles four points before any prompt is written:
The business result the sponsor expects to move
The behaviour on the job that would move it
Whether the people already know how to do it, and lack the conditions or the incentive
The measure, its baseline and its owner
A machine-drafted course built on a wrong diagnosis arrives faster and fails in the same way.
Problem framing for HR covers the conversation that settles these points, and training needs analysis covers the method.
A learning asset carries the organisation's authority, and a fabricated statistic inside it is the organisation's statement.
What Does Not Move
Four parts of the work are unchanged and they are the parts that decide whether learning has any effect.
Establishing whether the gap is capability at all, rather than clarity, conditions, consequences or capacity.
Agreeing the measure, the baseline and the owner with the sponsor.
Designing the transfer, including the manager brief and the real assignment.
Deciding what to stop doing to make room for it.
A function that halves its production time and skips those four has produced more of something nobody needed.
Personalisation, Realistically
Adaptive and personalised learning is the claim most made in vendor demonstrations and least evidenced in practice.
Two versions are worth distinguishing.
Adapting the path through fixed content based on assessment, which is old technology with a new label and works reasonably.
Generating different content per learner, which is harder to quality-assure because no two learners see the same thing.
For anything carrying compliance or safety consequence, the second is not defensible without review of every variant.
Where the content is low stakes and the volume is high, it earns its place.
Assessment Items
Generating assessment questions is among the strongest uses and it has a specific failure.
Machine-written distractors are frequently obviously wrong, which makes the question easy without making the learner competent.
Three checks fix most of it.
Every distractor should be something a partially competent person would genuinely choose.
Every question should map to a stated objective, and questions that map to nothing get deleted.
Someone who has done the job should sit the assessment before it is released.
Translation And Localisation
Machine translation of learning content is now good enough to change what is affordable, and it introduces a risk that is easy to miss.
Examples, scenarios and names carry cultural assumptions that survive translation and then land wrongly.
Two practices handle it.
Translate the instruction and rebuild the examples locally.
Have a native speaker who does the job read the scenarios rather than the text.
A correctly translated scenario that nobody in the market recognises is worse than a clumsy one that describes their week.
What To Tell A Sponsor
Three statements keep the conversation honest.
Production time falls, and the analysis and transfer work does not.
Every asset carries a named human reviewer, and fabricated content is the risk being managed.
We will measure one real task before claiming a productivity figure.
That last commitment is what separates a function with a position from one repeating a vendor's number.
The Development Path Problem
A consequence sits closer to the talent function than to the technology.
The tasks now done faster are frequently the tasks junior designers learned on.
Storyboarding, first drafts and collation were the mechanism by which a new designer learned the craft.
Removing them without redesigning the early-career role breaks the development path rather than freeing capacity.
Two responses belong in the function's own workforce plan.
Name what a junior designer is now expected to learn, and where.
Give juniors the checking and the expert conversations rather than the producing, and supervise the reasoning rather than the output.
A Governance Line
Four rules cover most learning functions.
Every asset names its human reviewer
No statistic or citation ships unverified
Regulatory content is signed by its owner
Prompts and outputs are kept with the asset
The fourth is the one that makes the others auditable.
An organisation that can show what produced a piece of learning content is in a different position from one that cannot.
Where The Tools Sit
In Trusted Talent Advisor, the prompt discipline sits in The Prompt Library.
The diagnostic that establishes whether a capability gap exists at all is in component 01, The Core Practitioner Guide.
The Talent Template Pack includes the design brief and the impact measure canvas.
Related reading: AI in talent management, ChatGPT prompts for HR, what HR should learn about AI first.
Sources
1. Cavalier, J. Applying AI in Learning and Development. ATD Press.
2. Mollick, E. Co-Intelligence, on the uneven boundary of capability.
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
Where does AI help in instructional design?
Most in development, where assets are produced, and least in analysis, where the gap and population are established. That is the opposite of where most functions' difficulty sits.
What is a prompt workflow?
A chain rather than a single request. Cavalier's example runs from learning objectives to video script to storyboard, with each step checkable.
What are the risks of AI-generated learning content?
Fabricated information that sounds plausible. Statistics, citations, regulatory content and technical procedure all need verification before release.
Does AI change the role of the subject matter expert?
Yes. Their scarce contribution moves from drafting to verification and judgement. Bring them a draft and ask what a competent person would do differently.
Can AI write assessment questions?
Yes, with checks. Machine-written distractors are often obviously wrong, so every distractor should be something a partially competent person would genuinely choose.
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