What AI cannot do in training
Current AI systems cannot assess skills that require high-quality human judgement, handle complex emotional situations, guarantee the reliability of what they generate without expert validation, or compensate for failing organisational conditions. These limits are not temporary defects awaiting a fix: they are structural. Ignoring them leads to deployments that create disillusionment and remediation costs.
Why identifying limits is a condition for success
Enthusiasm for AI in training produces two opposite errors. Rejection without evaluation: "human relationships are irreplaceable, AI cannot train." Indiscriminate over-expectation: "AI will revolutionise everything, trainers will disappear."
Both views ignore the same fact: AI performs well on certain types of task and poorly on others, in predictable, documented ways. Precisely identifying what it can and cannot do lets you use it where it adds real value and keep human mechanisms where they are irreplaceable.
What AI cannot do
Assess skills that require high-quality human judgement.
Public speaking, managerial posture, conflict management, creativity in context, active listening in a consulting relationship: these skills are assessed through human observation and cannot be reduced to automatable indicators. An AI system can assess whether a learner has retained the steps of a commercial negotiation. It cannot assess whether that learner builds the trust needed for a negotiation to succeed.
This limit has a concrete implication: training for behavioural and relational skills requires retaining human assessment mechanisms, whether observation by a trainer, role-plays with human feedback, or peer assessment.
Replace human support for complex learning.
An expert trainer brings contextual judgement on the learner's specific situation, an ability to detect what the learner does not articulate, and a relationship that motivates persistence. A line manager giving feedback in a real situation produces learning that AI cannot replicate.
Handle complex emotional situations.
AI does not perceive distress, demotivation rooted in personal context, or relational tensions that block learning. These signals require human listening and response.
Guarantee reliability without expert validation.
Language models produce fluent, confident answers that may be factually incorrect (hallucinations). Any AI-generated content used for training must be validated by a subject-matter expert before deployment.
Compensate for failing organisational conditions.
AI cannot fix a lack of time allocated to learning, a manager who does not support development, or an organisation that does not allow skills to be practised. Optimising the training system does not create the conditions in which training can have an effect.
What AI does well
Generate tailored content quickly from a skills framework. Personalise a learning path based on measured gaps. Provide 24/7 answers to comprehension questions. Measure acquisition skill by skill and trigger spaced revision. These are powerful but bounded capabilities.
Operating principles
Use AI where it is strong (generation, personalisation, measurement, availability), keep humans where they are irreplaceable (judgement, relationship, emotional support, contextualisation). Design hand-offs: the AI handles the first level, the human takes over for complexity.