The professional learning industry invests billions of dollars every year in online training. The returns are startlingly low. A landmark Cochrane review of 215 studies, involving more than 28,000 healthcare professionals, found that educational meetings produce a median improvement of just 6% in professional practice and 3% in patient outcomes. Over 80% of continuing education still relies on passive methods: read, watch, click next. Methods that decades of research have shown to have little or no effect on real-world behavior.
The problem is not a lack of content. It is a lack of practice and feedback.
This is the case we make in our new whitepaper, Practice, Not Content: The Evidence for AI Coaching in Learning and Development. Below is the short version: why passive learning fails, what AI coaching actually is, and what learners told us when we embedded it in real, production eLearning programs at Iowa State University, The France Foundation, and GardaWorld.
In summary
- Organizations spend billions on online training, yet passive content drives a median improvement of just 6% in professional practice. The gap is not content. It is practice and feedback.
- AI coaching gives every learner personalized, rubric-based feedback on what they actually write, embedded directly inside the course. It requires production, not recognition.
- Across three real deployments, the signal was consistent: 87% of learners were satisfied with the AI practice, 94.6% reported confidence in the target skill, and 60% spontaneously named the AI practice as the most valuable part of the course, without being asked.
- It works because of design and discipline, not the AI itself. The eLearning industry was built on content. The next era will be built on practice.
The broken promise of online learning
Enterprise learning is a massive global category. Corporate training alone is projected to expand from $361.5 billion in 2023 to $805.6 billion by 2035. The outcomes do not match the investment.
The reason is that most eLearning is built on the “information deficit model,” the assumption that if people simply knew the right thing to do, they would do it. Behavioral science has comprehensively debunked this. Three findings tell the story:
The forgetting curve is steep.
People forget a large share of newly learned information quickly without reinforcement, often with the steepest drop in the first day. Passively watching a video or clicking through slides does not create the neural pathways required for durable memory.
The knowing-doing gap is wide.
Intentions often fail to translate into action. Most people who intend to change behavior do not follow through unless they form specific if-then plans.
Self-assessment is unreliable.
Systematic reviews of physician self-assessment found that self-ratings often do not correlate with observed competence. Learners cannot be trusted to accurately identify their own skill gaps.
Then there is the checkbox trap. Most eLearning optimizes for completion, not competence. Multiple-choice questions test recognition (identifying a correct answer from a list) rather than production (generating an appropriate response from memory). Learners click through modules, select pre-written answers, and earn a completion certificate without ever doing the cognitive work that produces real learning.
The eLearning industry has spent two decades optimizing content delivery. The research says the problem was never content. It was the absence of practice and feedback.
What AI coaching actually is
When we say AI coaching, we mean something specific: AI-powered, personalized feedback on learner-generated responses, embedded within the learning experience and evaluated against expert-designed rubrics.
The mechanism is straightforward. A learner encounters a scenario: a student in distress, a patient with a complex medical situation, a customer with an objection. Rather than selecting from pre-written options, the learner types what they would actually say, in their own words. An AI system evaluates that response against a rubric designed by subject matter experts and returns specific feedback: what the learner did well, what they missed, and what they might say differently. Then the learner tries again.
This is not a chatbot, and it is not a stand-in for a human coach. It takes the best part of live coaching, real-time feedback while someone is actually practicing, and delivers it at eLearning cost. Here is how it differs from what most teams use today:
Versus multiple choice: AI coaching requires production, not recognition. Learners must think and write, not scan and select.
Versus branching scenarios: Branching offers pre-written paths. AI coaching responds to what the learner actually wrote. There are no pre-determined branches.
Versus open models like ChatGPT: AI coaching operates in a closed sandbox with expert-designed rubrics, references only approved content, and needs no prompt engineering. There is no risk of hallucinated advice.
Versus human coaching: AI coaching offers unlimited practice at any hour without scheduling or cost. It fills the gap between what learners need and what most organizations can afford.
The behavioral science backs this up. A meta-analysis of more than 200 studies and 2.1 million participants found that well-designed behavioral interventions produce a meaningful effect (Cohen’s d of 0.43). When learners are required to articulate specific plans for applying what they learned, exactly what AI coaching does, the effect size rises to d=0.65, a medium-to-large effect.
The industry is already moving
Two major industry surveys from 2025 and 2026 paint a consistent picture. The annual AI in L&D survey by Donald H. Taylor and Egle Vinauskaite found that 54% of L&D practitioners now actively use AI, up from 40% in 2024. The authors call this the “Implementation Inflexion.” The Synthesia AI in L&D Report found even higher adoption, with 87% of teams now using AI and only 2% having no plans to adopt.
More telling is where the frontier has shifted. Content creation is still the most common use of AI, but the fastest-growing applications are the ones that touch the learner directly. AI role plays went mainstream in 2025. Dynamic, AI-generated feedback that adapts to learner input is now a distinct modality. The question is no longer whether to use AI, but where it adds the most value for learners.
The evidence: what learners told us
We looked at primary learner data from three real-world deployments where AI coaching was embedded directly inside production eLearning, using AIReady technology, with no external tools or prompt engineering required:
Iowa State University deployed a university-wide program, Cyclone Support: Creating a Culture of Care, teaching staff a conversation framework for supporting students in distress.
The France Foundation, an accredited medical education company, built an AI Case Coach to train clinicians in shared decision-making and motivational interviewing for patients with obesity.
GardaWorld, a global security and operations organization, used AI coaching to give frontline workers and supervisors safe, repeatable practice in customer service, de-escalation, and SOP adherence.
The Iowa State deployment produced one of the larger real-world datasets of learner feedback on embedded AI coaching, with 473 post-course survey responses. The headline findings:
87%
of learners were satisfied with the AI-powered practice and feedback (52% extremely satisfied).
94.6%
reported confidence in the target skill after completing the course.
96%
rated the course at least moderately effective.
92%
found the course easy to navigate. The AI did not create friction.
“We’re seeing strong learner engagement with the AI interactions within the Cyclone Support Training and appreciate the partnership in developing and implementing this approach.”
Leif Olsen, Student Success and Retention Specialist, Office of the Senior Vice President and Provost, Iowa State University
The unprompted signal
The most important finding was not a Likert scale. It was what happened in the open-ended comments. Of the 248 learners who answered “What is one part of the course you appreciated most, and why?”, a question that never mentioned AI, 60% spontaneously named the AI practice or personalized feedback as the most valuable part of the course. That is 150 out of 248 comments: qualitative evidence at quantitative scale.
Three themes came through again and again.
Theme 1
Better than multiple choice.
“It was harder to come up with the right words than just to select the right words, but it was more realistic.”
“So much better to make me think of a response than to choose from a list.”
“It also wasn’t a fake choice… that feels like the ‘choices’ I give my literal toddler.”
The same shift away from multiple choice showed up in enterprise training contexts:
“The AI agents provide immediate clarification on SOPs and guide real-time decision-making. This reduces hesitation and helps ensure actions remain aligned with documented procedures when time matters most.”
Training Director, GardaWorld
Theme 2
The feedback felt real.
A common assumption is that AI feedback will feel generic or robotic. The data challenges that directly, including among healthcare professionals who hold high standards for clinical accuracy.
“The feedback felt genuine and was helpful.”
“The AI feedback was specific to my response.”
“It allowed me to practice patient-centered communication in a safe, simulated environment, helping me refine how I elicit patient motivations, address barriers, and collaboratively set realistic goals.”
That perception of feedback as actionable and contextually relevant is echoed by the teams who build these coaches:
“The Artha team turned around a working AI coach for us at remarkable speed. The rubric is working great, and our review team thought the feedback length and tone were spot on.”
Ailene Cantelmi, Director, Educational Development, The France Foundation
Theme 3
I didn't expect this from training.
Learners came expecting standard compliance training and found something different.
“I was surprised how effective the AI component was. It made me really consider what precisely I might say to a student.”
“I’m curious how I could create a similar module for class instruction. Very cool, and made it fun.”
The pattern holds across very different domains, learner populations, and skill types. As GardaWorld’s L&D Manager Charlie Piche put it:
“AI coaches developed through Artha provide an opportunity not only to practice, but to practice repeatedly, refining the critical skills that drive our business… we are moving beyond multiple-choice decision making. Instead, we are enabling learners to actively practice, while freeing up trainers to focus on foundational development.”
The value, in other words, is not domain-specific. It is design-specific.
“With AI, we can deliver the kind of personalized, real-time coaching that traditional training just can’t match. My goal is to use AI to give learners the safe, adaptive practice they need to master tough skills like negotiation. The Claims coach we created at Liberty Mutual with Artha achieves just that, without losing sight of human accountability.”
Marlie Cardiff, Senior Instructional Designer Specialist (Retired), Liberty Mutual
Being honest about the limits
This is post-course survey data, not longitudinal behavior-change measurement. Confidence is self-reported, not observed, and there was no control group. These were production deployments, not randomized controlled trials. The converging evidence from academic research, industry data, and learner feedback gives a strong triangulated foundation, but controlled studies measuring on-the-job behavior change would strengthen it further.
It is also worth addressing the resistance head on. At Iowa State, only 11 of 473 learners (2.3%) expressed explicit discomfort with the AI component, and most of those objected to AI as a concept rather than to the learning experience. On hallucination and safety, AI coaching in a closed sandbox is categorically different from open-ended ChatGPT: the AI references only approved, expert-validated content, and learners are asked to practice a skill and receive feedback, not to trust the AI as a source of truth.
It is not the AI. It is the design.
AI coaching works, but it does not work automatically. The technology is just a delivery mechanism. The design and the discipline behind it determine whether it produces real learning and whether it can withstand scrutiny from regulators, accreditors, and learners. These are the principles we have developed across deployments in higher education, healthcare, and enterprise:
Expert-informed rubrics. Subject matter experts define what “good” looks like before any AI is involved. Without that, AI coaching is just a text box with a chatbot.
Closed sandbox, not open internet. The AI references only approved, vetted content. This is what makes it safe for regulated environments.
Production over recognition. Force learners to write, not select.
Immediate, specific feedback. Not “good job,” but “you acknowledged the concern but did not connect them to a specific resource.”
Retry opportunity. Learners try again after feedback, creating the practice-feedback loop that drives skill development.
Embedded, not bolted on. The coaching lives inside the existing course and LMS. Learners do not leave or learn a new interface.
Behind the scenes, four more disciplines separate a coach that holds up from one that breaks down in the wild: heavy upfront investment in tone and voice, designing for the unexpected (off-topic answers, disclosures, one-word responses), multi-method testing before launch, and continuous audit after production. This convergence is showing up in regulation too: in early 2026 the ACCME published its most detailed guidance yet on AI in education, spanning source integrity, human oversight, transparent disclosure, and learner data protection.
“An AI coach isn’t always the right answer. But when it is, and when it’s designed well, it works beautifully… This is a new type of learning, and L&D needs to develop the craft to create it well.”
Garima Gupta, Founder and CEO, Artha Learning Inc. and AIReady
The shift is already happening
The evidence converges from three independent directions. Academic research says passive learning fails. Industry data shows the profession is moving. And primary learner data shows that when passive content is replaced with AI-powered practice and feedback, learners do not just complete the training. They engage with it, value it, and report confidence in the skill.
The eLearning industry built its foundation on content. The next era will be built on practice. The shift is already happening, and our learners are expecting it.
Read the full white paper. Practice, Not Content: The Evidence for AI Coaching in Learning and Development is free and ungated — read it online or download the PDF.
Not sure where your team stands? The five-minute AI roleplay readiness self-assessment scores you across six dimensions and points to a sensible next step.
Want to see what AI coaching looks like in a real course? Explore the AIReady demos. To learn more about how Artha Learning builds responsible, evidence-based AI coaching, visit arthalearning.com.
