Insights · AI
AI in L&D: what to automate, what to leave to humans
The useful question is not whether AI belongs in learning design. It is which parts of the work are judgment and which parts are production. Automate production aggressively, protect judgment absolutely, and the same team ships several times more, at higher quality.
The split, in one line
AI is excellent at volume, variation and speed. It is unreliable at deciding what matters. Every sensible workflow follows from that.
Automate first, highest return, lowest risk
- First-draft content from an approved source document or SME transcript.
- Scenario and distractor variation. Producing ten plausible wrong answers is tedious for a human and trivial for a model.
- Localisation and reading-level adaptation across languages and audiences.
- Media production: voiceover, image generation, storyboard drafts, subtitles.
- Summarising SME interviews into structured learning objectives for a designer to edit.
- Content refresh sweeps, flagging modules whose source policy or product has changed.
Keep human, non-negotiable
- Deciding the business outcome and what behaviour must change. This is a conversation with leaders, not a prompt.
- Curriculum architecture and sequencing. Models produce plausible orderings, not pedagogically sound ones.
- Factual and regulatory sign-off. A confident wrong answer in compliance content is a liability event.
- Cultural and political judgment. What is a safe example in one region is a resignation in another.
- Assessment of real competence, especially where someone's role or licence depends on it.
- Facilitation and coaching. Practising a difficult conversation with a human who can read the room is the point.
The grey zone: use with a human in the loop
Adaptive learner pathing, AI role-play partners, automated feedback on written submissions, and analytics interpretation all work well, provided a human owns the thresholds and reviews a sample. Treat AI output here as a recommendation for a designer to accept or reject, never as a decision that ships unseen.
A workflow that holds up
- Human defines the outcome, audience and success measure.
- Human writes the objectives and module architecture.
- AI drafts content, scenarios, variants and media against that structure.
- Designer edits for accuracy, tone and instructional quality.
- SME signs off facts; legal signs off anything regulated.
- AI localises and produces variants of the approved master.
- Human reviews performance data and decides what changes next quarter.
Governance you need before scaling
- A written policy on which content types may be AI-drafted and which may not.
- Source control: models draft from approved documents, not from open-web recall.
- Data rules, no employee personal data or confidential material into public tools.
- An audit trail recording who reviewed and approved each asset.
- Disclosure to learners where AI generated substantive content or feedback.
What good looks like
In our engagements the effect is not fewer designers, it is a different mix of work. Production time drops sharply, and the reclaimed hours move to needs analysis, stakeholder work, scenario quality and measurement, the parts that determine whether the programme changes anything.