The debate around artificial intelligence in training is no longer about whether to adopt it. That conversation closed two years ago. The debate in 2026 is a different one — far more demanding: how to integrate it without ending up with something worse than what we had before. And this is where the industry is splitting into two camps, with very uneven results.
On one side, solutions that delegate the entire process to AI — from the script to the assessment — trusting the model to produce valid training content on its own. On the other, teams that keep instructional design firmly in human hands and use AI only as a tactical acceleration tool in very specific parts of the process. The difference between the two approaches shows up in something very concrete: completion rates, on-the-job transfer, and — above all — what the learner actually retains.
At CAE, after 45 years supporting companies and educational institutions through their pedagogical evolution, we have witnessed many waves of technology. The current one is no exception: AI delivers a genuine efficiency leap, but only when integrated with sound judgment within a solid pedagogical framework. This article explains exactly where we draw that line and why.
The Temptation to Delegate Everything to AI (and Why It’s a Mistake)
The promise is tempting. “Upload your syllabus and AI will generate the complete course in hours.” And technically it works: there are tools that produce modules, videos, voiceovers, questions, and summaries at astonishing speed. The problem is not the speed — it’s what gets lost along the way.
When AI produces content without pedagogical supervision, four things happen that the buyer won’t spot on first viewing but the learner will feel from the very first hour. The content becomes generic: examples are universal, not specific to the sector or the company. Cognitive progression breaks down: concepts are presented without the sequencing an instructional designer would build to ensure consolidation. Assessments measure recall, not competence: questions verify that the learner read the content, not that they can apply it. And the connection to business objectives disappears: the course teaches about a topic, not how to solve a specific on-the-job problem.
The result is something we see more and more often: broad, visually appealing training catalogues produced at speed, with low completion rates and near-zero transfer to the job. This is precisely the problem we analysed in 8 Signs Your Online Training Catalogue Is Failing: training that looks good on paper but generates no real change in day-to-day practice.
What AI Does Well and What the Human Expert Does Well
The productive discussion is not human vs. AI. It is which tasks each one excels at within the instructional design process. When this is separated clearly, the result is far superior to what either could achieve alone.
AI brings speed to producing first drafts, generating large question banks, professional voiceovers, translations, conversational role-plays, and adaptive feedback in practical exercises. It excels when working from a clear brief, within a well-defined domain, and against defined validation criteria. AI does not tire, does not improvise, and produces at near-zero marginal cost.
The human expert brings judgment to pedagogical architecture, cognitive sequencing, disciplinary rigour validation, alignment with business or academic objectives, adaptation to the real learner profile, and compliance assurance. The human expert does not produce at AI speed, but makes the decisions AI cannot — because they are not generation decisions, they are design decisions.
The trap is confusing the two functions. When AI is asked to design, it fails. When the human expert is asked to produce all content without assistance, they lose cost and timeline competitiveness. The model that works recognises that these are distinct functions and distributes them accordingly.
The Hybrid Model: Where CAE Draws the Line
At CAE we apply a model refined across hundreds of training projects. We call it the hybrid model, and it is built on a simple but decisive distinction: design is human, production is AI-assisted.
The instructional design phase — needs analysis, definition of learning objectives, competency architecture, pedagogical sequencing, assessment model — is always led by a pedagogical expert from our team. AI can contribute inputs (sector trend analysis, activity suggestions, examples), but design decisions are made by a human. Without exception.
The content production phase — development of detailed scripts, voiceovers, illustrations, question bank generation, conversational role-plays, adaptive feedback — is carried out with intensive AI assistance. The pedagogical team reviews, adjusts, and elevates the outputs, but AI accelerates the process dramatically: what used to take weeks is now completed in days.
The pedagogical validation phase returns entirely to human hands. Every module produced is reviewed at two levels: a pedagogical review that verifies instructional coherence, accessibility, and packaging; and a disciplinary review that verifies academic rigour and alignment with project objectives. Only when both sign off does content move to final production.
This architecture is the operational translation of what at CAE we call the balance between POWER (the power of technology) and HUMAN (specialist professionals) — the two central pillars of our methodology.
The Seven Moments Where Human Validation Is Non-Negotiable
There are moments in the process where delegating to AI is directly dangerous for academic quality. These are the seven we never give up.
The definition of learning objectives and their alignment with the real learner profile. AI can suggest generic objectives; only a pedagogue knows the specific institutional or corporate context.
The cognitive sequencing of content: in what order concepts are presented so that learning consolidates. AI tends to present content in the order it appears in the source material, not in the optimal order for learning.
The selection of authentic examples from the sector or company. AI generates plausible but generic examples; the expert incorporates the real examples that connect with the learner’s experience.
The construction of the assessment rubric and its alignment with learning outcomes. AI generates questions; the expert decides what is assessed, with what weight, and against what criteria.
The review of disciplinary rigour. AI can contain subtle factual errors that only a subject-matter expert will detect. In regulatory, healthcare, technical, or legal content, this level is absolutely non-negotiable.
The cultural and linguistic adaptation. AI translates; the expert localises. The difference between the two operations is immediately felt in how the learner receives the material.
The integration with the client’s strategic objectives. A training course is not an end in itself — it is a tool for change. Only a pedagogue with a consultative mindset connects the content to the real business objective.
Five Best Practices for Designing AI-Powered Training Without Sacrificing Rigour
Drawing on experience accumulated across real projects, these are the five practices that distinguish well-designed AI-powered training from training that merely looks well-designed.
Define the brief before touching AI. The brief includes the learning objective, learner profile, target level in the European Qualifications Framework, target duration, context of use, and success criteria. Without a brief, AI produces aimless content.
Validate every AI output before moving forward. This is not about reviewing the course at the end — it is about reviewing each block as it is produced. Detecting a pedagogical error in phase 6 that was introduced in phase 2 multiplies the correction cost.
Maintain a record of human decisions. Every design decision made by the pedagogical team must be documented. This allows the rigour of the process to be defended in audits, with corporate clients, or before academic committees.
Combine AI with proprietary institutional data. The best results are achieved when AI works with the company’s or university’s specific material, not generic content. Personalisation using your own data is the difference between a useful course and an applicable one.
Measure what matters, not what is easy to measure. The real quality metrics are on-the-job transfer, completion with genuine learning, and observable behaviour change. Vanity metrics are clicks, logins, and progress percentages.
Three Bad Practices That Are Damaging the Sector’s Reputation
It is worth naming them explicitly because they are undermining the perception that companies and institutions have of AI-powered e-learning as a whole.
Mass-generating courses with no subsequent pedagogical review. This is what some low-cost platforms offer as a selling point. The result is broad but generic catalogues, with no differentiation and extremely low rates of on-the-job application.
Replacing the instructor with chatbots lacking pedagogical conversational design. An effective virtual tutor requires conversational architecture, validated scripts, anticipated scenarios, and ongoing supervision. A chatbot plugged into a general-purpose model is not a virtual tutor — it is a reputational risk for the institution that deploys it.
Confusing personalisation with trivial segmentation. Inserting the learner’s name on a screen is not personalisation. Real personalisation adapts the itinerary, examples, difficulty level, and feedback to the individual learner’s profile. This is the difference between well-applied AI and cosmetic AI.
Academic Quality Indicators in AI-Powered Training
How to tell a well-designed AI course from one that only looks well-designed. Four observable indicators that any buyer can verify before approving a project.
Diversity of activity types: if the course consists only of linear videos and multiple-choice questions, it was likely produced without solid instructional design. A well-designed course combines at least four or five activity types.
Depth of feedback: if exercise corrections are of the “correct/incorrect” type, pedagogical judgement is missing. A good course explains why an answer is wrong and connects back to the content to reinforce learning.
Coherence between objectives and assessment: every learning objective declared at the start must be assessed at the end. If stated objectives do not appear in the assessment, the design is weak.
Adaptation to the learner profile: the course must demonstrate that it knows who it is addressing. If examples are universal and scenarios generic, there is no real personalisation.
The CAE Methodology Applied to Instructional Design with AI
At CAE, we have integrated AI into all our design processes since 2023, but always within the methodological framework that has guided us for more than four decades: powerful technology supported by specialist human teams. Our R&D, linguistics, pedagogy, and technical support teams work in an integrated way so that every course, every SCORM content package, and every platform meets the academic standards that our clients — international companies and prestigious universities — expect from us.
If your organisation is considering integrating AI into its instructional design processes and wants to do so with genuine quality assurances, talk to our team and we will show you how we work.
Frequently Asked Questions
What is instructional design with AI?
Instructional design with AI is the process of creating training content that combines the involvement of pedagogical experts with generative artificial intelligence tools. AI accelerates content production, while human experts make the design decisions, validate academic quality, and ensure alignment with project objectives.
Can AI completely replace the instructional designer?
No. AI is excellent at generating content from a clear brief, but it cannot replace the design decisions that require pedagogical judgement, knowledge of the institutional context, and disciplinary rigour validation. The effective model combines both: AI for accelerated production, human experts for design and validation.
How is the academic quality of an AI-produced course guaranteed?
Academic quality is guaranteed through a pedagogical validation process at at least two levels: instructional review (coherence, accessibility, packaging) and disciplinary review (academic rigour, alignment with objectives). Without these human filters, AI-based production loses reliability.
What time savings does AI bring to instructional design?
Generative AI reduces content production times by between 40% and 60%, provided human pedagogical supervision is maintained. The reduction is concentrated in production phases (scripts, voiceovers, question banks), not in the design or validation phases, which retain their full human workload.
Is AI compatible with standards such as SCORM or LTI?
Yes. Content produced with AI assistance is packaged in standard SCORM or LTI formats in exactly the same way as traditionally produced content. Compatibility with any LMS is guaranteed as long as the provider follows the corresponding technical specifications.
What risks does poorly supervised AI carry in training?
The main risks are: subtle factual errors that go undetected; generic content with no connection to the learner’s real context; assessments that measure memory rather than competence; and loss of coherence between declared objectives and outcomes that are actually achievable. All of these risks are mitigated through human pedagogical validation at every stage of the process.