Everyone is telling learning teams the same thing: you should be using AI. Fair enough. The useful question is the one that advice skips: how, exactly, should you use AI for learning? Not which logo to buy. Where AI belongs in the work, and where it does not. Three decisions answer it: who the AI accelerates, where it meets your program, and whether it operates as one system or a pile of parts.
Decision 1
Accelerate your designers, or replace them?
Accelerate. The designer owns the domain expertise, the organizational understanding, and the craft. AI puts an accelerant on those skills; it cannot supply them.
Decision 2
Where does AI meet your program? Two places, two trades.
Live, with your learners
Gain: free-form conversation. Cost: ongoing per-learner AI spend, experiences no one can monitor, limited analytics, generic content.
In the build, locked at release
Gain: a known, personalized experience from approved content, aggregatable analytics, no ongoing cost, no hallucinations. Cost: the free-form conversation.
Decision 3
One orchestrated platform, or a chain of AI tools?
Orchestrated: build fast, edit fast, one change carries through voice, video, and every language.
Tool chain: edit one word, then regenerate the voice, rebuild the video, and redo every translation.
Will AI accelerate your designers, or replace them?
This one comes first because it is personal. If you are an instructional designer or content developer, you have heard some version of it already, maybe from a vendor, maybe from your own leadership: AI can do what you do. The answer is no, and the reason is specific.
You own three things AI does not have: expertise in the domain being trained, an understanding of what is really going on inside your organization, and the craft to turn both into something compelling. A model cannot know that a product launch is stalling because reps avoid one specific objection, or that the last training rollout failed for reasons nobody put in a document. You know.
What AI does with those three things is act as an accelerant. The same designer builds more, builds faster, and gets more creative rather than less, because the mechanical work of production stops consuming the hours where creativity would happen. Replacement, by contrast, is limited by exactly what AI lacks: the domain nuances, the organizational context, and the creativity of instructional design itself.
The field evidence sits comfortably in a supporting role here. In a series of randomized field experiments with sales agents, Luo and colleagues found that the incremental benefit of an AI coach over human managers followed an inverted U: middle-ranked agents improved most, while bottom-ranked agents faced information overload from the AI's feedback and top-ranked agents showed the strongest aversion to it relative to a human coach. Their third experiment found that the AI and human coach combination outperformed either the AI or the human coach alone. Karl Kapp, relaying a longitudinal field study by Habel and colleagues on AI sales role plays, reports the same shape from a different angle: real salespeople improved by 7 to 35 percent, but the gains concentrated among reps with low prior performance, high goals, strong supervisors, and longer organizational tenure. His conclusion is that you cannot simply unleash an AI role play and expect results; the instructional and supervisory design around it does the heavy lifting. The people are not a bottleneck for AI to remove; they are the judgment it amplifies.
Where should AI meet your program: with your learners, or building their content?
There are two places AI can work in a learning program, and knowing the pros and cons of each is the second decision.
The first place is live, with your learners. The AI converses with each learner directly, and the pro is genuine: free-form conversation, available around the clock, with unlimited patience. That is exactly why live AI earns its place for how-to-say-it rehearsal, where a learner works on delivery, objection handling, and tone, and for generic tutoring in education, where the content is broadly established. The cons follow from the same architecture, and a program owner needs to know them going in. The cost is ongoing and linear: you pay for AI compute every time a learner talks to it, for every learner, for as long as the program runs. Every experience is unique, which sounds appealing until you have to manage a program: you cannot monitor what learners are actually experiencing, and the analytics stay limited because a thousand different conversations share no common frame. And the AI's responses cannot reliably carry nuanced, domain-specific content; what it generates in the moment is plausible, not authored.
The second place is in the build. AI accelerates the creation of the content itself, humans review and approve the result, and the experience locks at release. The gains stack up: every learner gets a known experience built from the same content. It is still a personalized experience, because each learner's choices drive their own path through it, but the personalization comes from the same approved content, and that shared frame is what makes the data aggregatable into genuinely insightful analytics. When a hundred learners face the same decision, the pattern in their choices means something. There is no ongoing AI cost, because no AI runs at delivery. And there are no hallucinations and no misrepresented content, ever, because nothing a learner sees was generated after approval.
What the second place gives up is the free-form conversation. That is the trade, stated plainly, and it is the right trade whenever the training is nuanced and specific to your business, which is most of what learning teams are asked to deliver.
| AI live with the learner | AI in the build, locked at release | |
|---|---|---|
| What you gain | Free-form conversation, always available | A known, approved experience for every learner |
| Cost profile | Ongoing and linear, paid per conversation | Development only, no AI cost at delivery |
| Personalization | Unique every time, unmonitorable at scale | Personalized by the learner's choices, from the same content |
| Analytics | Limited, no common frame to aggregate | Aggregatable and insightful, built on shared decisions |
| Domain-specific nuance | Not reliably carried by live generation | Authored, reviewed, and approved before release |
| What you give up | Monitoring, aggregation, and nuance | The free-form conversation |
Regulated environments add a further layer. Improvado's guide to pharmaceutical advertising compliance advises companies to "treat AI as a drafting tool, not an approval authority," with final compliance determination made by a human. In pharma, finance, or healthcare, an unreviewable live experience is a hard stop. But the point is bigger than compliance: even where no regulator is watching, content nobody approved is content nobody can stand behind, and an experience nobody can observe is an experience nobody can improve.
Is the AI a fragmented toolchain, or orchestrated inside one platform?
The third decision is about the build itself, and it is the one content developers feel in their calendars. The current reality of AI content creation is a chain of disconnected tools: AI video from one vendor, AI voice from another, AI images from a third, and then AI translation applied on top of all of it. Each tool is impressive alone. The chain is the problem.
Editing is where it breaks down. Change one word in a script and you are back through the whole chain: regenerate the voice for the changed line, rebuild the video around the new audio, redo the translation, and then repeat all of that in every language you deployed. A one-word correction from a subject matter expert becomes a production cycle, multiplied by your language count. The individual tools are fast; the pipeline built from them is slow, and the pipeline is where the work lives.
Orchestrate all of that AI inside one platform and the fragmented sprawl becomes what AI was supposed to be in the first place: an accelerant. You build quickly, because the pieces already work together. You edit quickly, because a change made once propagates instead of triggering a rebuild. And you control the whole process flow in one place, which is what makes review and approval workable rather than heroic.
What do the three decisions add up to?
One principle, applied at three points in the lifecycle: AI accelerates the creation, a human approves the result, and only the approved version ever ships, so nothing can drift. The first decision protects the people: the designers whose domain expertise, organizational understanding, and craft the AI is amplifying. The second protects the learners and the program: a known experience a human stood behind, with analytics that share a common frame. The third protects the work between them: when the AI is orchestrated in one platform, review cycles are cheap enough to actually run, so approval is a working step instead of a bottleneck.
AliveSim, Syandus's immersive learning platform, is a worked example of all three answers in one architecture. Its platform page describes it as an AI-enhanced platform for rapidly building nuanced, strategic, interactive scenarios: AI accelerates content creation while the authoring team maintains control, and all AI-generated content is locked down before publishing, so every learner gets a consistent, reliable experience. The company's medical communications page states the delivery consequence plainly: "Predetermined paths with no AI surprises." The AI work, including AI-powered voice and translation for multi-language deployment, is orchestrated inside the same platform where scenarios are authored, reviewed, and published, and the analytics track the decision patterns those locked scenarios make comparable across learners.
However you proceed, the three decisions travel well. Use AI to accelerate your designers, not to replace what only they know. Choose deliberately between AI that talks with your learners and AI that builds what they experience, with the pros and cons in view. And favor AI orchestrated in one platform over a toolchain you have to fight every time something changes. Get those three right and AI makes your program faster and better at the same time.
References
- Habel, J., Ahearne, M., Tirunillai, S., & Vandaveer Novak, A. (2025). Do AI role plays improve sales performance? Evidence from a longitudinal field study. SSRN. https://doi.org/10.2139/ssrn.5717822 (as summarized in Kapp, 2026)
- Improvado. (2026). Pharma ad compliance: FDA, FTC, fair balance, and ISI requirements. https://improvado.io/blog/pharma-ad-compliance-fda-ftc-fair-balance-and-isi-requirements
- Kapp, K. (2026, June 10). Do AI sales role plays work? L&D Easter Eggs newsletter, LinkedIn.
- Luo, X., Qin, M. S., Fang, Z., & Qu, Z. (2021). Artificial intelligence coaches for sales agents: Caveats and solutions. Journal of Marketing, 85(2), 14–32.
- Syandus. AliveSim platform. https://www.syandus.com/alivesim-platform
- Syandus. Medical communications. https://www.syandus.com/medical-communications
Related questions
Will AI replace instructional designers?
The replacement case runs into exactly what AI lacks. Instructional designers own three things: expertise in the domain being trained, an understanding of what is actually happening inside their organization, and the craft to build something compelling. AI has none of them, and its output is only as good as those inputs. What AI does well is act as an accelerant on those skills, taking over the mechanical production work so the same designer builds more, builds faster, and has more room to be creative. The field evidence points the same way: in randomized field experiments, the combination of AI and human coaches outperformed either working alone.
What are the pros and cons of AI talking directly with learners?
The pro is real: free-form conversation, available anytime, with unlimited patience. That is why live AI earns its place for how-to-say-it rehearsal and for generic tutoring in education. The cons follow from the same architecture. Cost is ongoing and linear: you pay for AI compute every time a learner talks to it, for as long as the program runs. Every conversation is unique, so there is no common frame for analytics and no way to monitor what learners are actually experiencing at scale. And a live model cannot reliably carry nuanced, domain-specific content, so the training stays generic. For nuanced, business-specific learning, the cons usually outweigh the pro.
What do you give up when AI-built content is locked at release?
The free-form conversation. That is the trade. In exchange, the program gains a great deal: AI accelerates the drafting, humans review and approve every element, and once published, the approved version is fixed. Every learner gets a known experience built from the same content. It is still a personalized experience, because each learner's choices drive their own path through it, and because those paths run through shared content, the data aggregates into analytics you can act on. There is no ongoing AI cost, because no AI runs at delivery, and nothing can be hallucinated or misrepresented, because nothing a learner sees was generated after approval.
Why does editing AI-generated content get complicated?
Because most AI content today is built through a chain of separate tools: AI video from one vendor, AI voice from another, AI images from a third, and AI translation applied on top. Building through the chain once is fast. Editing is where it breaks down. Change one word of a script and you regenerate the voice, rebuild the video around the new audio, and redo the translation, in every language you deployed. Each tool boundary is another step to repeat and another place for versions to fork. Orchestrating those same AI capabilities inside one platform collapses that loop: you make the change once and the platform carries it through.
Published July 17, 2026 · 8 min read