The AI-Native Learning Roadmap 2026–2031

How to read this roadmap

This report combines external market research with the author’s strategic interpretation and Mexty’s perspective on the evolution of AI in learning.

We have written this roadmap as a strategic foresight framework, not as a forecast with rigid dates. In practice, technologies overlap. Some pioneers are already testing capabilities that appear later in the roadmap, while other organizations will move more slowly. Adoption will vary by sector, context and level of maturity.

The roadmap is the main publication. Annexes A, B and C preserve the full original research documents on AI in Learning Market Research 2026, The ROI of Learning, and the AI-Native Learning Infrastructure Maturity Index.

Executive Summary

Over the past three years, I have watched artificial intelligence change learning at a remarkable pace. The first shift was easy to see: AI made content generation dramatically faster. Courses, quizzes, illustrations, scripts, translations, assessments and learning objectives that once took hours or days to produce can now be developed in a fraction of the time. Based on feedback from 30 Mexty customers, teams report time savings in the range of approximately 60-80%, depending on the complexity of the workflow.

What interests me now is the second shift already taking shape. AI is moving beyond content generation into the creation of actual learning experiences: scenarios, simulations, roleplays, games, adaptive exercises, custom interactions and AI-powered practice environments.

But I believe the more important transformation is still ahead. Over the next five years, AI will increasingly become part of the learning system itself. It will be able to observe what a learner does, interpret mistakes, connect those signals to competencies, recommend or trigger an intervention, reassess progress and, increasingly, connect learning evidence with real-world performance.

1. 2023-25

   GENERATE

   Available today in Mexty

   AI generates learning content.

2. 2025-27

   EXPERIENCE

   Available today in Mexty

   AI generates interactive learning experiences.

3. 2027-29

   ORCHESTRATE

   Partially available in Mexty, expanding

   AI continuously adapts and orchestrates learning.

4. 2029-31

   PERFORM

   Long-term vision

   Learning and work connect through capability loops.

Progression matters more than the exact dates: Content → Experience → Orchestration → Performance.

1. Why the AI-learning market is moving beyond content generation

The first wave of generative AI in L&D was, understandably, mostly about productivity. Ask an LLM to create learning objectives. Upload a PDF and turn it into a course. Generate assessment questions. Rewrite material for a different audience. Produce images, scripts, summaries, translations or feedback.

That was genuinely useful. Learning-content production has always consumed a great deal of time and resources, so adoption moved quickly. LinkedIn Learning reported in its 2025 Workplace Learning Report that 71% of L&D professionals were exploring, experimenting with or integrating AI into their work.

But there is a consequence we do not talk about enough: once everyone can produce content faster, speed itself stops being much of a differentiator. Content becomes abundant. The scarce value moves elsewhere - toward judgment, learning design, practice, measurement, orchestration, governance and evidence of performance.

2. Phase 1 - 2023-2025: GENERATE

AI makes learning content abundant.

StatusAvailable in Mexty

The first phase was characterized by AI acting primarily as a production assistant. The dominant question was: How can AI help us create learning faster?

AI could support instructional designers, trainers, teachers, and subject-matter experts with course structures, learning objectives, quizzes, scripts, summaries, translations, images, videos, assessment questions, rewriting, and first drafts.

This changed the economics of content production, but it did not fundamentally change the architecture of learning. A course was still largely a course. The learner still consumed it. The assessment still generally happened afterwards. The LMS still recorded completion and scores.

AI made the existing process faster. That was valuable. But it was not yet transformative.

3. Phase 2 - 2025-2027: EXPERIENCE

From generating content to generating learning experiences.

StatusAvailable in Mexty

AI is increasingly capable of producing not only information but experiences: branching scenarios, interactive activities, simulations, roleplays, games, adaptive exercises, conversational characters, custom interfaces, and contextual feedback.

This matters for a very practical reason: people do not develop capability simply by consuming information. They develop it through practice, decisions, feedback, retry, progression, repetition, judgment and application.

AI-assisted programming and vibe coding for learning reduce a long-standing constraint. Historically, instructional designers were partly limited by the capabilities of their authoring tools. If the desired interaction did not exist as a template, the creator often needed technical resources or had to simplify the pedagogical idea.

The starting question can increasingly become: What should the learner experience or practise to develop this capability? rather than: What interaction can my authoring software build?

And once interactive learning becomes easy to produce, that too becomes abundant. The value moves upward again - toward orchestration.

4. Phase 3 - 2027-2029: ORCHESTRATE

The architecture of learning changes.

StatusPartially available in Mexty, expanding

Today, most digital learning still follows a relatively linear model. A learning need is identified. A course is created. A learner is assigned the course. The learner completes it. An assessment produces a result. The LMS records the activity. Someone may later review a dashboard.

The system primarily records what happened. It knows whether the learner completed something, how long they spent, and perhaps what score they achieved. But in most implementations it does not continuously interpret what the learner is doing, understand why the learner is struggling, and change the experience while learning is occurring.

This is where, in my view, the architecture really starts to change. AI is no longer used mainly to create learning assets; it begins to participate continuously in the learning system itself.

1. Knowledge
2. Learner
3. Activity
4. Mistake
5. Feedback
6. Assessment
7. Competency
8. Intervention
9. Performance

If a learner repeatedly makes the same type of mistake, the system does not need to wait for the final assessment. It can identify a potential misconception, adapt the next activity, offer targeted practice, ask an AI tutor to challenge the learner's reasoning, return the learner to an appropriate trusted source, change the difficulty, test the same competency in another context, or recommend a human intervention.

The system moves from simply recording and reporting toward observing, interpreting, intervening and reassessing. That is a very different proposition from generating a course.

4.1 Architecture for AI-Native Learning Infrastructure

We use the term AI-Native Learning Infrastructure to describe a convergence already visible across existing categories: authoring, LMS/LXP, AI tutors and coaches, enterprise knowledge, AI agents, analytics, assessment, and business systems. The term describes the architecture created when these capabilities become connected, governed and able to work together across the learning lifecycle.

Trusted Source of Truth - AI should be grounded in approved knowledge: company procedures, policies, product documentation, research, curricula, standards, internal expertise, and validated learning materials. The question becomes: What knowledge is the AI authorized to use?

Full control of AI - AI-native cannot mean AI-controlled. Organizations need to determine which models are allowed, what information they may access, which actions they can perform, where human validation is required, what is logged, and which decisions remain human.

Multi-LLM architecture - Different models will continue to perform differently across reasoning, coding, generation, analysis, multilingual tasks, simulation, tutoring, and assessment. A multi-LLM layer preserves flexibility and avoids unnecessary dependence on a single provider.

MCP and connected systems - MCP and other integration mechanisms allow agents to interact with trusted knowledge, LMS, HRIS, CRM, collaboration tools, assessment systems, skills databases, and operational applications instead of operating in isolated conversations.

Persistent learner and competency context - Continuous adaptation requires more than identity. The system may need role, competency profile, previous activity, prior assessments, recurring mistakes, level of proficiency, learning objectives, permissions, and previous interventions.

Evidence layer - Activity is not the same as competency. AI-native learning needs stronger mechanisms that connect learner action to evidence, evidence to competency, and competency to intervention - especially for compliance, certification, and high-stakes assessment.

1. Source of Truth
2. Multi-LLM Intelligence
3. AI Agents
4. Learning Workflows
5. Evidence & Assessment
6. Analytics
7. Enterprise Systems

Across every layer: security, governance, traceability, and human control.

4.2 From personalization to continuous adaptation

We have discussed personalized learning for years, but much of what has historically been called personalization is actually segmentation. Beginner gets pathway A. Advanced learner gets pathway B. Manager gets one pathway. Employee gets another. Useful, but still predefined.

AI-native learning introduces something more dynamic: the system can adapt while learning is actually happening.

1. Observe
2. Interpret
3. Intervene
4. Practise
5. Reassess
6. Adapt

Suppose a learner makes a particular mistake during a simulation. Instead of simply recording a wrong answer, the system can try to determine what the behavior indicates. Does the learner lack knowledge? Understand the rule but struggle to apply it? Is the problem judgment, confidence, context, or a recurring misconception?

The system can then choose an intervention appropriate to that signal. Once the learner tries again, it can observe whether the intervention worked.

For me, this is the key change in question: not simply “Which course should this learner take?” but “What does this learner need next, based on what they have just demonstrated?”

That does not mean adaptation should always make learning easier. Sometimes the right intervention is to withhold the answer, challenge the learner, ask another question, require another attempt or deliberately preserve productive struggle.

4.3 Enterprise Use Case - Continuous Capability Development

Customer service training connected to operational performance.

Imagine a customer-service employee learning how to handle difficult complaints. In a conventional model, the employee might complete a course, watch a video, answer several questions, and receive 85%. The LMS records completion. But suppose the employee still unnecessarily escalates customer complaints.

In an AI-native environment, the employee enters a realistic conversation with an AI customer. The infrastructure knows the approved escalation policy, product rules, the learner's role, previous activity, and the competency being developed.

During the simulation, the employee escalates too quickly. Instead of simply recording an incorrect decision, the system detects a pattern. Perhaps the learner knows the policy but lacks confidence applying it when a customer becomes confrontational.

The environment can present targeted feedback, bring the learner back to the relevant approved procedure, ask them to explain their reasoning, generate another scenario concentrating on that judgment, let them retry, and observe whether the behavior changes. If the issue persists, the system can recommend manager coaching.

Eventually, the question becomes much more concrete: can this learning evidence be connected to operational data, and are unnecessary escalations actually decreasing?

1. Workplace objective
2. Simulation
3. Decision
4. Evidence
5. Targeted practice
6. Reassessment
7. Workplace performance

4.4 Education Use Case - Adaptive Learning Without Removing Productive Struggle

A high-school student learning quadratic equations.

In a conventional digital experience, the student watches an explanation, completes ten questions, and receives 6/10. We know the student struggled. We do not necessarily know why.

An AI-native learning environment can potentially detect the pattern behind the errors. Perhaps the student understands the equation once it is arranged correctly but repeatedly makes the same mistake when moving terms from one side to another.

The system might ask the student to explain their reasoning, provide a visual activity concentrating on that operation, introduce a simpler example, use an AI tutor that asks questions rather than giving the answer, generate targeted practice problems, allow the learner to retry, and later test the same concept in another context.

The teacher sees more than 6/10. They can see where the learner struggled, what support was provided, how many attempts were necessary, whether the misconception persisted, and whether mastery subsequently improved.

The goal is not for AI to solve the equation for the student. It is to create the conditions in which the student can think, struggle productively, receive feedback, try again and build real understanding.

1. Exercise
2. Mistake
3. Diagnosis
4. Targeted support
5. Practice
6. Retry
7. Mastery check

5. Phase 4 - 2029-2031: PERFORM

Learning and work begin to converge.

StatusLong term vision

Continuous adaptation is not the end of the story. The next logical step is to connect learning signals with performance signals.

Today, learning and work often exist in separate systems. The LMS may know that an employee completed negotiation training. The CRM knows whether negotiations subsequently improved. The learning system rarely connects the two continuously.

Phase 4 begins closing that gap. A salesperson repeatedly struggling with a specific objection, a technician repeatedly requesting help, a service employee generating an unusual escalation pattern, or a manager receiving recurring feedback can become capability signals.

The learning infrastructure can identify a potential gap, propose a targeted intervention, collect evidence of improvement, and later compare it with operational data. Learning begins to become part of the flow of work.

1. Work signal
2. Capability gap
3. Intervention
4. Practice
5. Evidence
6. Performance
7. Next intervention

6. The four-phase AI-Native Learning Roadmap

The dates in this roadmap should therefore be read as a direction of travel, not as universal adoption deadlines. Technology availability and organizational maturity are two different things.

7. How the L&D profession evolves

L&D already asks many of the right questions. What should people learn? Which capabilities really matter? Where are the gaps? What kind of practice will help? How do we know whether someone has improved? And, ultimately, how does that learning connect to performance?

The problem is that answering those questions is still often manual, fragmented and periodic. Learning teams interview people, analyze surveys, speak with managers, review LMS data, design programs, wait for completion and then try to work out whether anything meaningful changed.

An AI-native infrastructure can help connect signals that today live in separate systems. This changes where L&D creates value.

I do not see the instructional designer disappearing. Quite the opposite: I think the role becomes more strategic. Learning professionals will increasingly decide what AI should know, which sources it should trust, which model should perform which function, where learners need support, where they should struggle, when an AI agent should intervene, what counts as meaningful evidence, when a human should step in and how learning connects to business performance.

AI does not make L&D expertise less important. If anything, it makes good L&D judgment more valuable.

8. Measurement moves from completion toward performance

For years, we have tended to measure digital learning through what the LMS could easily give us: completion, attendance, time spent, assessment scores and satisfaction. These indicators are still useful, but they do not necessarily tell us whether the problem that triggered the learning intervention actually improved.

A stronger approach starts earlier. ROI should not be something L&D tries to reconstruct after the training has been delivered. Measurement needs to be designed into the learning system from the beginning.

If training is intended to reduce customer escalations, escalation rates should form part of the measurement strategy. If training is supposed to reduce safety incidents, incidents matter. If the objective is faster onboarding, time-to-competence matters. If the objective is improved sales capability, relevant performance indicators matter.

AI does not magically prove causality - and we should be careful not to pretend that it does. Human judgment and sound evaluation design remain necessary. What AI can do is help us combine evidence, detect patterns, identify gaps and monitor change more effectively.

1. Business objective
2. Performance indicator
3. Expected outcome
4. Learning intervention
5. Evidence of capability
6. Performance change
7. ROI

9. AI-native does not mean removing cognitive effort

If AI continuously adapts to learners, should it continuously make learning easier? No.

One of the questions I think future learning systems must handle particularly carefully is the difference between unnecessary friction and productive struggle. Searching, reasoning, hesitating, reformulating, making a mistake and trying again can all be essential parts of learning.

The goal, therefore, should not be maximum convenience. It should be better learning.

A good AI tutor will not always answer. Sometimes it should help, sometimes challenge, sometimes explain, sometimes ask another question - and sometimes deliberately stay out of the way.

10. Five conditions for AI-Native Learning Infrastructure

11. Enterprise and education will follow the same roadmap differently

12. Market roadmap vs organizational maturity

The roadmap describes where I believe the technology and the market are heading. It does not tell us where every organization sits today. That is a separate question.

This is why the AI-Native Learning Maturity Index complements the roadmap. The roadmap asks, “Where is the market going?” The maturity index asks, “Where are we today, and what do we need to build next?”

13. Where Mexty fits in this roadmap

Everything described so far is a market-level observation rather than a Mexty pitch. But it would be incomplete to describe where learning technology is heading without also explaining how we are positioning Mexty within that evolution.

We started, like much of the market, by helping people create learning faster. Over time, however, our ambition has moved well beyond creation.

Mexty is being developed as an AI-Native Learning Infrastructure that connects trusted Knowledge Bases and Sources of Truth, multi-LLM intelligence, authoring, interactive activities, simulations, assessments, learning paths, AI agents, learner support, analytics, human validation, versioning, governance and learning delivery.

The important point is not the number of features available. What matters is whether these capabilities work together as one connected system.

An assessment should influence what happens next. An AI agent should understand the trusted knowledge behind the learning experience. An interactive activity should generate useful evidence. A learning path should adapt to what the learner has demonstrated. Analytics should help us understand not only whether someone completed something, but whether capability is developing, where gaps remain and where an intervention may be needed.

Connections such as MCP, APIs and enterprise connectors can extend the learning environment into the broader enterprise ecosystem. Throughout this process, humans must remain able to understand, validate, govern and override what AI is doing.

AI-Native cannot mean AI-controlled.

This is not a slogan without evidence. Concretely, Mexty is ISO/IEC 27001:2022 certified and has been natively designed to support GDPR and EU AI Act requirements, with human validation, traceability, versioning and governance embedded into the platform's architecture, not added as a policy statement after the fact. When we say AI-native cannot mean AI-controlled, we mean it can be audited, not just asserted.

The learning environment we are moving toward must remain human-led, AI-enabled, grounded in trusted knowledge, secure, governed, measurable and connected.

How Mexty Connects the Existing Learning Ecosystem ?

Mexty can operate as a complete AI-Native Learning Infrastructure in its own right. Authoring, interactive activities, simulations, assessments, AI agents, LMS, analytics and learning delivery are all native capabilities.

At the same time, organizations should not have to abandon the systems, content and investments they already have in order to move toward a more AI-native model.

Mexty is therefore designed to support both approaches. Organizations can build directly within Mexty where that makes sense, or connect Mexty with an existing LMS, enterprise content, knowledge systems and business applications when replacement is unnecessary.

This also helps explain how organizations can move toward Level 5, AI-Native Learning Infrastructure, from very different starting points.

A company may currently be at Level 0, 1, 2, 3 or 4 of the Maturity Index. These levels describe where the organization is today, but they do not define a mandatory sequence that every company has to follow.

An organization at Level 1 or Level 2, for example, does not need to spend years moving step by step through each intermediate level. It can decide to design directly toward a Level 5 architecture if the right governance, infrastructure, data, integrations and operating model are put in place.

What matters is not the starting point. What matters is whether the organization can connect the essential components of the learning environment: trusted knowledge, creation, practice, assessment, AI agents, learner context, analytics, delivery and business systems.

Mexty is designed to support that transition from different levels of maturity. Some organizations may choose to build most of the environment natively within Mexty. Others may prefer to keep significant parts of their current stack and connect them progressively. In many cases, the best approach will be a combination of both.

The objective is not to force organizations through every maturity level. It is to help them define the right target architecture and move toward it in the way that best fits their current systems, constraints and priorities.

Level 5 is a destination, not a mandatory sequence of steps.

And the path to Level 5 is not about replacing everything. It is about making the elements that matter work together in a secure, governed and measurable way.

Conclusion: Learning becomes responsive

If I had to summarize the next five years in one sentence, it would be this: learning stops being something we only deliver and increasingly becomes something that responds.

It responds to a learner's mistakes, progress, changing knowledge, emerging capability gaps, assessment evidence, workplace performance and, eventually, to changes in the organization itself.

The course does not disappear. The LMS does not disappear. Instructional design certainly does not disappear. They become parts of something larger: a continuously operating learning environment in which human judgment sets the objectives and boundaries, trusted knowledge provides the foundation, and AI helps interpret signals, orchestrate experiences and respond at a scale that was previously difficult to achieve.

Content → Experience → Orchestration → Performance

The first phase made content abundant. The second is making interactive learning experiences easier to create. The third introduces continuous orchestration, where AI can interpret evidence, adapt interventions and support learners dynamically. The fourth connects learning more directly with capability, operational data and business performance.

This evolution does not mean that every organization will follow the same path. Some will start from traditional learning environments, others from AI-assisted content creation or connected learning workflows. What matters is not the starting point, nor whether one platform owns every component. What matters is whether the essential elements can work together as a secure, governed and evidence-producing system.

Organizations may choose to build more of this environment natively, connect their existing LMS, content, knowledge and enterprise systems, or combine both approaches. The goal is not replacement for its own sake. It is to connect what already exists with what AI now makes possible.

That leaves us with a question that, to me, is much more interesting than “How quickly can AI create a course?”:

How intelligently can we connect trusted knowledge, people, practice, assessment, evidence, AI and performance into a learning system that continuously helps people improve?

That is the transition from AI-assisted learning to AI-Native Learning Infrastructure.

It is not simply about adding more AI features. It is about creating a learning environment that is connected, measurable, governed and human-led, where AI can support learning across the full lifecycle while remaining within clearly defined boundaries.

Organizations can move toward that destination from very different levels of maturity.

Level 5 is not a mandatory sequence of steps. It is a target architecture.

I believe this shift toward AI-Native Learning Infrastructure will be one of the defining transformations in Learning & Development over the next five years.

References

LinkedIn Learning (2025). 2025 Workplace Learning Report.

https://business.linkedin.com/learn/resources/workplace-learning-report

OECD (2026). OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education. https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_062a7394-en.html

UNESCO (2025). AI and education: Protecting the rights of learners.

https://www.unesco.org/en/articles/ai-and-education-protecting-rights-learners

Gartner (2026). The Impact of AI on the Learning and Development Operating Model. https://www.gartner.com/en/documents/7880477

Gartner (2026). Guide to Redesigning the Instructional Designer (L&D) Job in the Age of AI. https://www.gartner.com/en/documents/7883477

About the author

Alexandre Maupas is CPO of Mexty, an AI-Native Learning Infrastructure designed to connect trusted knowledge, AI-assisted creation, interactive learning, assessments, AI agents, analytics, governance, and learning delivery.

His work focuses on how AI, interactive experiences, game mechanics, and connected learning systems can improve how people practise, learn, and develop capability.

This report is deliberately a point of view, not a claim that every organization will move at the same speed or in exactly the same way. Its purpose is to contribute to the discussion among L&D leaders, instructional designers, educators, technology providers and organizations deciding what their next stage of AI adoption should look like.

Research Annexes

Annex A - AI in Learning Market Research 2026

Annex B - The ROI of Learning

Annex C - AI-Native Learning Infrastructure Maturity Index

The following annexes preserve the original research documents provided for this publication. They give additional depth on the market, ROI, and maturity-index dimensions that support the AI-Native Learning Roadmap.

Annex A - AI in Learning Market Research 2026

Market Size

The estimates vary by scope, but all point to rapid growth:

• AI in Education (broad market, all segments):

Between $8-11B in 2026 depending on the research firm (Precedence Research, Grand View Research, Research and Markets), with projections reaching $57-137B by 2033-2035, at CAGRs ranging from 25% to 41%.

The most public conservative picture is given by MarketsandMarkets defines the global AI in Education market relatively narrowly and estimates it at $2.21B in 2024, growing to $5.82B by 2030, a 17.5% CAGR.

• AI-Powered Corporate Training (specific to enterprise training):

Mordor Intelligence states that the market is expected to grow from $6.27B in 2025 to $7.49B in 2026, and then reach $18.19B by 2031, representing a 19.43% CAGR from 2026 to 2031.

• AI in eLearning:

The Business Research Company, in its Artificial Intelligence (AI) in eLearning Global Market Report 2026, states that the global AI in eLearning market was $5.08 billion in 2025, is expected to reach $5.67 billion in 2026, and is forecast to grow to $8.76 billion by 2030, corresponding to a 11.5% CAGR over the 2026–2030 forecast period. This is looks a more modest pace since this segment is already mature.

These market estimates reflect different definitions and scopes and should not be added together or treated as directly comparable

Segmentation

• By component:

software solutions dominate with 70% market share in 2025, the rest split across associated services.

• By deployment:

cloud represents 57% market share in 2025, but on-premise is expected to grow faster, driven by data privacy and security concerns.

• By region:

North America leads with roughly 37.5% market share in 2025, while Asia-Pacific shows the fastest growth.

• By company size:

SMEs are expected to grow faster than large enterprises (21.93% CAGR), as vendors lower entry barriers through modular pricing, faster deployment, and lighter administration.

Key Market Trends Today — 2026

AI adoption is already mainstream in L&D, but the market remains at a relatively early stage of maturity. At the same time, several structural shifts are already visible: value is moving from content production toward practice, simulation and conversational learning are growing, AI tutors and coaches are emerging as a distinct category, corporate learning is moving from courses toward skills and workforce readiness, and AI literacy is becoming a new learning market.

Value is moving from content production toward practice

According to The Josh Bersin Company’s 2026 research, businesses spend approximately $400B on corporate training, content, L&D technology and related services. Yet the research concludes that traditional e-learning and video-based courseware are no longer sufficiently dynamic, personalized or comprehensive to meet rapidly changing workforce needs.The next wave is increasingly about practice and measurement, rather than content production alone.

Bersin, J. (2026). “New Research: How AI Transforms $400 Billion Of Corporate Learning.” https://joshbersin.com/2026/02/new-research-how-ai-transforms-400-billion-of-corporate-learning/

Roleplay, simulation and conversational learning are growing

RAIN Group and Allego found that 56% of organizations with highly effective sales training use online role-plays or simulations. Separately, Highspot’s 2025 State of Sales Enablement Report found that B2B organizations using AI in sales coaching are 20% more likely to report improved revenue outcomes.

This reflects a broader shift toward learning experiences based on practice, simulation, conversation and feedback rather than passive content consumption.

RAIN Group & Allego (2024). Continuous Learning in Sales.
https://www.rainsalestraining.com/sales-training-company/news/continuous-learning-research

Highspot (2025). State of Sales Enablement Report 2025.
https://www.highspot.com/resource/state-of-sales-enablement-report-2025

AI tutors and coaches emerging as a distinct category

This trend is particularly visible in education.

Khan Academy reported that Khanmigo reached 2.0 million students, educators and parents in SY24–25, representing 731% year-over-year growth, with 770,000 students using it through U.S. district partnerships. In 2026, Khan Academy also reported an average of 269,000 Khanmigo interactions on weekdays and more than 108 million interactions since its 2023 launch.

More important than the user numbers is the pedagogical approach. Khanmigo is deliberately designed not simply to provide answers. It asks questions and guides learners through reasoning.

Khan Academy (2025). Annual Report, School Year 2024–25.
https://annualreport.khanacademy.org/

Khan Academy (2026). “Learning in the Open: What AI Is (and Isn’t) Changing.”
https://blog.khanacademy.org/learning-in-the-open-what-ai-is-and-isnt-changing/

The OECD Digital Education Outlook 2026 highlights an important distinction: general-purpose GenAI can improve immediate task performance without necessarily producing learning gains, while GenAI designed or used with clear pedagogical intent can support sustained learning through tutoring, collaboration, feedback and guided assistance.

OECD (2026). OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education. OECD Publishing, Paris.

https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_062a7394-en.html

This represents a major market shift: from chatbot → pedagogical agent

Corporate learning is moving from courses toward skills and workforce readiness

The traditional LMS market itself remains substantial. According to Global Market Insights (2026), the global Learning Management Systems (LMS) market is estimated at $33.4B in 2026 and projected to reach $97.4B by 2035, representing a 12.6% CAGR.

Global Market Insights (2026). Learning Management Systems Market Size & Share, 2026–2035. Published August 2026.

https://www.gminsights.com/industry-analysis/learning-management-system-lms-market

But the important shift is not simply LMS growth. The architecture of corporate learning is changing from: Course management toward: Skills + Learning + AI + Knowledge + Performance

ISG's 2026 Learning Platforms Buyers Guide describes this evolution: enterprise buyers increasingly expect learning systems to combine compliance, personalized experiences, skills development, AI, job ontologies and integration with the wider HCM ecosystem.

This is a strategically important signal that the market is moving beyond the classic distinction: LMS vs LXP toward something much closer to: Learning infrastructure connected to the enterprise.

AI adoption is already mainstream in L&D

According to LinkedIn Learning’s 2025 Workplace Learning Report, 71% of L&D professionals are exploring, experimenting with, or integrating AI into their work.

LinkedIn Learning (2025). Workplace Learning Report 2025: The Rise of Career Champions.

https://business.linkedin.com/learn/resources/workplace-learning-report

However, this should not be interpreted as 71% having sophisticated AI-native learning environments.

Docebo's 2026 AI Readiness research shows that:

• 79% of learning teams say they already use AI for content, assessments or recommendations.
• 91% say their organizations have not yet fully redesigned their workflows around AI.
• 35% still describe themselves as being in the experimental stage.

This strongly supports the maturity model we have been developing for Mexty.

The market has largely reached: Experimentation → Content Generation but has not yet broadly reached: Connected AI Learning → AI-Native Learning Infrastructure

AI literacy is becoming a new learning market

There are actually two AI-learning markets emerging simultaneously:

Learning WITH AI and Learning ABOUT AI : Governments, schools and corporations increasingly need AI literacy.

The OECD and European Commission published an AI Literacy Framework for Primary and Secondary Education in June 2026, explicitly framing AI literacy around understanding AI, critically evaluating outputs, and using AI ethically and creatively.

Corporate demand is developing at the same time because organizations need employees to use AI safely and productively.

This creates a significant opportunity for platforms that support not only traditional learning, but also practical AI exercises and governed AI environments.

Key Structural Trends for 2026–2031

The trends visible in 2026 point toward a deeper transformation of the learning market over the next five years. AI will increasingly move beyond isolated content-generation features toward connected, adaptive and governed learning systems.

Content generation becomes a standard capability

AI-generated courses, quizzes, summaries, visuals and assessments will increasingly become standard functionality across authoring tools, LMS platforms and learning applications.

As this capability becomes widely available, content generation alone will become less differentiating.

The competitive advantage will move toward what happens after content is created: how learners practise, receive feedback, demonstrate capability and improve.

Interactive experience generation accelerates

The next wave of differentiation will increasingly come from the ability to create richer learning experiences: simulations, roleplay, branching scenarios, games, conversational practice, interactive applications and adaptive activities.

AI-assisted development and vibe coding will continue reducing the technical barrier to creating these experiences. The design question will progressively shift from: “What content should we produce?” toward: “What experience should the learner go through to develop the capability?”

AI moves from creation into learning delivery

Today, AI is still used primarily before the learner enters the experience: to generate content, assessments or learning assets.

Over the next five years, AI will increasingly participate during the learning process itself. It can act as: Tutor → Coach → Roleplay partner → Feedback agent → Assessor → Practice generator

This moves AI from being a production tool toward becoming an active component of the learning experience.

Personalization evolves into continuous adaptation

Much of today's personalization is still based on predefined segmentation. A beginner receives one path or an advanced learner receives another.

The next stage is more dynamic. AI-native systems can increasingly observe learner behavior, interpret what it means, intervene and then reassess whether the intervention worked.

The loop becomes: Observe → Interpret → Intervene → Practice → Reassess → Adapt

The key question changes from: “Which learning path should this person receive?” to: “What does this learner need next, based on what they just demonstrated?”

Skills and competency evidence become more important than completion

Learning platforms have historically focused heavily on: completion, attendance, time spent and scores.

These metrics will remain useful, but they are increasingly insufficient. Organizations will want to know:

• • What can this person actually do?
  • Where is the capability gap?
  • Has that capability improved?
  • Is learning affecting performance?

This will increase the importance of assessment, evidence, competency models and practical demonstrations of capability.

Learning, knowledge and business systems converge

Learning will increasingly stop operating as an isolated technology layer: Enterprise knowledge, learner data, skills, assessments, analytics and business systems will become more connected.

The architecture will move toward: Knowledge + Learning + AI + Skills + Performance

This means that learning systems will increasingly need access to trusted organizational knowledge and relevant enterprise context.

A learning agent should not operate only from generic model knowledge. It should understand the policies, procedures, product information, competencies and business context relevant to the learner.

Learning moves into the workflow

Learning will increasingly happen where work happens. Not only inside the LMS.

AI tutors, coaches, agents and performance-support systems will increasingly appear inside environments such as: Teams, Slack, CRM, HCM, ERP and other enterprise applications.

This will blur the boundary between formal learning and performance support.

Instead of employees leaving work to access learning, learning can increasingly appear at the moment of need.

Governance, human control and evaluation become infrastructure requirements

As AI moves from generating content toward tutoring, assessing, recommending and acting, governance becomes much more important.

Organizations will need to define: which models can be used, what data they can access, which sources they can trust, what actions they may perform, when human validation is required, how AI outputs are evaluated, and how decisions are traced.

Governance therefore moves from being a policy document to becoming an operational capability embedded in the learning architecture.

Human control remains essential. AI-Native should not mean AI-controlled.

Multi-LLM, agents, MCP and interoperability become architectural components

The future learning environment is unlikely to depend on a single model or isolated application. Different LLMs may be selected for different tasks such as: reasoning, generation, coding, simulation, multilingual support, assessment or analysis.

AI agents will increasingly perform specialized roles, while protocols such as MCP, together with APIs and connectors, will allow them to interact with trusted knowledge and enterprise systems.

This creates a more modular architecture: Trusted Knowledge → Multi-LLM Layer → AI Agents → Learning Workflows → Assessment → Analytics → Enterprise Systems

Interoperability becomes strategically important because organizations will want to remain independent from any single AI provider or learning platform.

Learning platforms evolve toward AI-Native Learning Infrastructure

The cumulative effect of these trends is a broader transformation of the learning technology category itself.

The market will progressively move beyond: Authoring tool, LMS, LXP, AI tutor, as isolated categories.

The next-generation architecture will increasingly connect: trusted knowledge, AI models, agents, creation, interactive practice, assessments, learner data, analytics, delivery and enterprise systems.

This is what we define as AI-Native Learning Infrastructure.

The objective is not simply to add AI features to an existing learning platform. It is to create a learning system in which AI can operate across the full learning lifecycle while remaining secure, governed, measurable and under human control.

The transition can therefore be summarized as:

AI-assisted learning tools → Connected AI Learning → AI-Native Learning Infrastructure

And this is likely to be one of the most important structural shifts in the learning market between 2026 and 2031.

Key Players

The AI and learning market increasingly spans several adjacent categories rather than one clearly bounded competitive set. We think that the most useful way to read the landscape is to look at the role each group plays in the learning ecosystem.

Conclusion

The AI and learning market is moving quickly from experimentation and content generation toward practice, adaptive learning, skills development, AI agents and measurable capability.

Over the next five years, the key shift will be from isolated AI features toward connected, governed learning systems where trusted knowledge, AI models, agents, assessments, analytics and enterprise systems work together.

This transition can be summarized as: AI-assisted learning tools → Connected AI Learning → AI-Native Learning Infrastructure and the long-term value will come less from generating more content and more from connecting learning, practice, evidence, AI and performance in a secure and human-controlled environment.

Annex B - The ROI of Learning

From Completion Metrics to Business Performance

Introduction

Learning & Development is under increasing pressure to demonstrate business value.

For years, learning teams have relied heavily on metrics such as completion, attendance, time spent, learner satisfaction and assessment scores. These indicators remain useful, but they do not answer the question executives increasingly care about:Did learning change performance, reduce risk, save money or create value?

This measurement gap is still significant.

ATD's 2025 research found that only 43% of organizations report strong alignment between learning goals and business goals, while only 30% say they are good at using learning-program data to support business decisions. The same research found that 90% regard isolating the impact of training on business results as a challenge.

LinkedIn Learning reaches a similar conclusion: engagement and retention remain the most common business-impact measures, but learning leaders increasingly need to demonstrate how initiatives contribute to money made, money saved or risk mitigated.

The next evolution in learning measurement therefore requires a move from: Learning activity → Learning metrics toward: Business Objective → Required Capability → ROI Objective / ROI Model → Learning Intervention → Evidence of Capability → Performance Change → Business Impact → Actual ROI → Continuous Improvement

AI and increasingly connected learning systems can make this much easier by linking learner activity, assessment evidence, skill development and operational data.

2. Why Learning ROI Is Difficult to Measure

Learning rarely produces a financial result directly because for example a sales course does not create revenue by itself, a safety program does not directly reduce accidents and a leadership course does not automatically improve employee retention.

However, Learning changes knowledge, behavior or capability, and those changes can subsequently influence business outcomes.

The challenge is therefore establishing a credible chain between:

Learning → Capability → Behavior → Performance → Business Outcome

This explains why many organizations stop at what is easiest to measure: completion, attendance, satisfaction, course ratings, quiz scores and time spent.

These metrics answer: Did people participate?

They do not necessarily answer: Did people improve?

And even less: Did that improvement create business value?

ATD's research illustrates this problem clearly. While 93% of organizations collect participant satisfaction data, relatively few are proficient at evaluating business impact. Only 40% of talent-development professionals in ATD's capability benchmark rated themselves proficient at evaluating impact.

3. ROI Should Start Before the Learning Program

One of the biggest mistakes in learning measurement is trying to calculate ROI after a training program has already been designed and delivered.

At that point, the organization may not have: a baseline, a clearly defined performance problem, access to operational data, a measurable business objective,or agreement about what success actually means.

A stronger methodology reverses the process.

Start with:

• Strategic Objective: What does the organization want to improve?
• Business / Performance Indicator: What measurable result should change?
• Required Capability: What must employees be able to do differently?
• Learning Objective: What knowledge, judgment or behavior must develop?
• Learning Experience: What practice will build that capability?
• Evidence: How will we know capability improved?
• Performance Measurement: Did behavior change at work?
• Business Impact: Did the relevant KPI improve?
• ROI: Was the value generated greater than the investment?

This changes learning design substantially.

The question is no longer: “What course should we build?”

It becomes: “What performance outcome are we trying to influence, and what evidence would show that learning contributed to it?”

3. The Different Levels of Learning Measurement

A useful ROI framework separates several different levels of evidence.

Level 1 — Participation and Experience

Typical metrics: completion rate, attendance, satisfaction, engagement, time spent.

These help evaluate whether people used and accepted the learning experience. They are useful operational metrics but they are not evidence of ROI.

Level 2 — Learning

Typical metrics: assessment scores, knowledge gains, demonstrated understanding, improvement between attempts.

This answers: Did the learner learn something?

Level 3 — Capability and Behavior

Typical metrics: performance in simulations, observed workplace behavior, manager assessments, practical demonstrations, task quality.

This answers: Can the learner apply what was learned?

Level 4 — Business Impact

Typical metrics might include: sales conversion, customer escalations, productivity;

error rates, safety incidents, quality, time-to-competence, employee retention, customer satisfaction, rework and regulatory risk.

This answers: Did the business outcome change?

Level 5 — Financial ROI

Only after business impact has been quantified financially should the traditional ROI formula be applied:

ROI (%) = (Financial Benefit − Learning Cost) / Learning Cost × 100

For example, if a learning initiative costs $100,000 and the financial benefit attributable to the initiative is estimated at $250,000:

Net benefit = $150,000

ROI = 150%

However, the credibility of this final percentage depends primarily on the model used to calculate the financial benefit.

This is why the ROI model should not be invented after the training has been delivered.

It should be designed before the learning intervention begins.

Organizations often define a learning objective and a measurable KPI in advance, but they do not always go far enough in defining the full model that will later convert performance improvement into financial or operational value.

A robust ROI model should therefore establish, before the intervention:

• which business outcome is expected to change;
• which capability is expected to influence that outcome;
• the relevant baseline;
• the target level of improvement;
• how the improvement will be translated into financial value;
• which costs and benefits will be included;
• how much of the observed improvement can reasonably be attributed to learning;
• which assumptions will be used;
• and what data will be collected to validate those assumptions.

For example, reducing customer escalations from 18% to 12% is a measurable KPI.

But that is not yet an ROI model.

The ROI model must also define:

• What does one avoided escalation cost?
• How many escalations are expected to be avoided?
• What portion of that improvement can reasonably be attributed to the learning intervention?
• What other factors may have influenced the result?

Only then can the financial benefit be estimated credibly.

So the key principle should be: A measurable KPI tells us whether performance changed. The ROI model tells us what that change is worth.

And that ROI model should be defined before the training is designed and delivered, not reconstructed afterwards to justify the investment.

This fits very well with your overall framework:

Business Objective → Required Capability → ROI Objective / ROI Model → Learning Intervention → Evidence of Capability → Performance Change → Business Impact → Actual ROI → Continuous Improvement

The important distinction is that ROI is designed upstream and measured downstream.

4. Most Learning Initiatives Can Be Connected to Financial Value

Not every learning initiative produces revenue directly, but that does not mean it cannot be connected to a financial ROI.

In many cases, the value is indirect and must be modeled.

For example, compliance or safety training may primarily aim to mitigate risk. The relevant financial question is not simply whether an incident occurred after the training, but: What is the expected cost if the risk is not mitigated and the event occurs?

A risk model can therefore consider: Probability of event × Financial impact of event

and compare the expected loss before and after the learning intervention.

If a compliance failure has a potential cost of $5M and the intervention is expected to reduce the probability of occurrence from 4% to 1%, the expected annual risk exposure moves from: $5M × 4% = $200K to: $5M × 1% = $50K

The modeled benefit of the intervention is therefore $150K of reduced expected risk, before considering the cost of the learning initiative.

The same logic applies to time-to-competence.

If onboarding reduces the average time required for an employee to become fully productive from 120 days to 90 days, the question becomes: What is the economic value of gaining 30 productive days per employee?

That value can be modeled using salary cost, productive output, revenue contribution, utilization or another relevant operational measure.

Similarly, productivity improvement can be translated into: time saved × number of employees × frequency × economic value of that time

Retention can be modeled through avoided recruitment, onboarding and productivity-loss costs.

Quality improvements can be translated into reduced rework, warranty or error costs.

Customer-service training can be linked to fewer escalations, shorter handling times or improved retention.

The key point is therefore:

The absence of an obvious revenue metric does not mean there is no financial value. It means the benefit model needs to be designed.

This reinforces why the ROI Objective / ROI Model must be created before the learning intervention.

The organization should determine in advance:

what benefit is expected, how that benefit will be calculated, which assumptions are being made, what data must be collected, and how the contribution of learning will be estimated.

This is fundamentally different from selecting a measurable KPI.

A KPI tells us: Did something change?

The ROI model tells us: What is that change economically worth?

I would therefore classify learning value into four broad economic mechanisms:

• Revenue creation: increased sales, conversion, customer value.
• Cost reduction / productivity gain: less time, fewer errors, faster onboarding, less rework.
• Risk mitigation: reduced probability and/or consequence of adverse events.
• Strategic capability creation: building capabilities needed for future performance, where the financial model may rely more heavily on assumptions and scenarios.

Even the fourth category can be modeled financially, although the further we move from directly observable performance, the more important it becomes to be transparent about assumptions and uncertainty.

So I would replace the comment believe that “Not Every Learning Initiative Needs Financial ROI” with “Financial ROI Can Often Be Modeled Even When the Benefit Is Indirect”

5. Examples of Learning ROI

5.1. Sales Training

Business problem: Conversion rate is too low.

Baseline: 20% conversion

Objective: 24% conversion

Learning intervention: AI roleplay, objection-handling practice, coaching and targeted feedback.

Learning evidence: improvement in simulation performance and objection-handling capability.

Operational measurement: conversion rates before and after training.

Business impact: incremental revenue generated.

The learning metric is not course completion.

The relevant chain is: Practice → Improved sales capability → Higher conversion → Revenue impact

5.2. Customer Service

Business problem: Too many unnecessary escalations.

Baseline: 18% escalation rate

Objective: 12%

Learning intervention: realistic customer simulations and targeted practice.

Evidence: ability to correctly handle difficult scenarios.

Operational KPI: escalation rate.

Financial impact may include: lower support costs, less manager intervention, faster resolution, higher customer satisfaction.

The chain becomes: Simulation → Judgment → Improved behavior → Fewer escalations → Cost reduction

5.3. Employee Onboarding

Business problem: New employees take too long to become productive.

Baseline: 120 days to competence

Objective: 90 days

Learning intervention: personalized onboarding, contextual practice and performance support.

Measurement: time to reach predefined competency thresholds.

Business impact: 30 days of additional productive capacity per employee.

This can often be translated relatively directly into economic value.

5.4. Safety Training

Business problem: Operational incidents.

Learning objective: improve hazard recognition and decision-making.

Learning intervention: interactive scenarios and simulations.

Measurement: hazard-identification accuracy, workplace observation and incident rate.

Business impact: reduced accidents, downtime, claims and risk exposure.

Here the value of learning may be predominantly:

risk mitigation, rather than revenue creation.

6. The Current Measurement Gap

The market data suggests that learning measurement has not yet caught up with the ambitions of modern L&D.

ATD found:

• 43% of organizations say learning and business goals are strongly aligned.
• Only 30% say they are good at using learning-program data to make business decisions.
• Only 4% rated themselves excellent at doing so.
• 66% expect pressure to demonstrate the financial benefits or profitability of talent-development initiatives to increase over the next two years.

LinkedIn similarly shows that organizations still rely heavily on indirect measures of career-development impact:

• employee engagement: 72%
• retention: 64%
• employees developing new skills: 55%
• promotions: 48%
• internal mobility: 32%

These are meaningful indicators but the next step is connecting them more systematically to business performance.

7. Why AI Changes the ROI Equation

AI does not automatically improve learning ROI but it can change both sides of the equation.

7.1. AI can reduce the cost of learning production

AI can accelerate, content creation, translation, assessment generation, interactive development, simulation creation, adaptation, maintenance and updates.

This reduces the investment side of ROI but that alone is not enough.

If AI simply makes it cheaper to produce ineffective learning, ROI has not necessarily improved.

The bigger opportunity is on the benefit side. AI can increase learning effectiveness as AI can potentially provide: more practice, personalized feedback, adaptive difficulty, AI tutoring, roleplay, targeted remediation and continuous support. This can improve capability development.

7.2. AI can improve measurement

Perhaps most importantly, AI can produce much richer evidence. Instead of only “completed / not completed”, the system can observe:

• what decisions learners make;
• where they repeatedly fail;
• how reasoning changes;
• how many attempts are needed;
• which competencies are weak;
• which interventions improve performance.

ATD's 2025 research found that 28% of respondents expect AI to have an extremely positive impact on learning measurement and evaluation over the following two years. This may become one of AI's most significant contributions to L&D.

7.3. AI-Driven Performance Learning Analytics

Traditional learning analytics largely describe what happened inside the learning platform: completion, scores, time spent, assessment results, practice history.

The next evolution is to connect those learning signals with what happens outside the learning environment.

AI is the key driver of this transformation. It can help interpret large volumes of learner behavior, assessment evidence and competency data, connect those signals with enterprise systems, identify patterns, surface correlations and support continuous adjustment of the learning experience.

The AI-Native layer needs data from both the learning system and the business systems in order to connect capability with performance. The architecture is closer to this:

Learning data
Assessment scores
Simulation behavior
Competency evidence
Practice history

+

Business performance data
CRM → sales, conversion
Customer service → escalations, resolution time
ERP → productivity, operational efficiency
Quality systems → errors, rework
Safety systems → incidents
HRIS → retention, mobility, performance

↓

AI Interpretation & Analytics

The AI analyzes the two sets of signals together:

Learning evidence + Business performance

to look for relationships such as:

Employees who demonstrate capability X in simulations subsequently show fewer customer escalations.

or:

Employees who reach competency level Y become productive 25 days faster.

or:

Teams showing stronger negotiation capability have higher conversion rates.

Then the system can:

Detect capability gaps → Connect them with performance → Recommend learning interventions → Measure what changes → Adapt the learning

This is where AI becomes particularly important. Without AI, organizations can already connect learning data and business KPIs manually. But it is difficult to do continuously across thousands of learners, multiple capabilities and large volumes of operational data.

AI can help identify patterns that would otherwise be very difficult to see.

A concrete example:

Imagine the business objective is to reduce customer escalations.

The learning environment records:

Simulation performance

• Did the employee identify the customer's real issue?
• Did they apply the correct policy?
• Did they attempt resolution before escalation?
• What mistakes did they make?

The customer-service system records:

Workplace performance

• escalation rate;
• first-contact resolution;
• handling time;
• customer satisfaction.

Now AI can analyze both.

It might discover that employees who repeatedly fail one particular step in the simulation (for example, applying the refund policy correctly) also have significantly higher escalation rates in real customer interactions.

That gives L&D something much more useful than: “The employee scored 72%.”

It gives: “This specific capability gap appears to be associated with this business-performance problem.”

The system can then prescribe additional targeted practice, reassess the employee, and subsequently observe whether the real escalation rate changes.

The loop becomes:

Capability gap detected → Targeted practice → Capability reassessed → Workplace behavior observed → Business KPI measured → Intervention adjusted

And this is probably the stronger message for your report:

AI does not simply improve learning analytics. It makes it possible to continuously connect learning evidence with performance evidence and use that relationship to improve both learning interventions and business outcomes.

One caution I would include at this stage: AI can detect relationships and patterns, but that does not automatically prove that the learning intervention caused the business improvement. That is where the ROI model and attribution methodology defined earlier remain essential.

7.4. AI Can Also Enable Continuous ROI Optimization

There is an even more interesting possibility.

Traditional ROI analysis is normally retrospective. Typically how it works: A program runs, months later someone evaluates whether it worked.

Connected AI learning systems could make measurements increasingly continuous.

Imagine that the system detects: Learners perform well on the course assessment, but workplace errors remain unchanged. That is a critical signal because perhaps the assessment is measuring the wrong capability, or one simulation improves performance substantially more than another. The system can surface that relationship.

Another possibility: A particular learner group needs three times more practice before achieving the same workplace result. That can inform subsequent learning design.

The learning system with AI begins to operate as: Design → Deliver → Observe → Measure → Improve rather than: Design → Deliver → Report

This is where AI can transform ROI from a reporting metric into an optimization mechanism.

8. The Role of AI-Native Learning Infrastructure

This is where AI-Native Learning Infrastructure becomes particularly relevant.

Measuring ROI requires information that usually sits in different systems: trusted organizational knowledge, learning activities, learner behavior, assessments, competency evidence, AI-agent interactions, LMS data, HR information and business KPIs.

If those systems remain disconnected, ROI measurement remains difficult and manual.

A connected AI-Native Learning Infrastructure can progressively bring these signals together. The architecture becomes:

Business Objective
Define the business outcome the organization wants to improve revenue, productivity, quality, safety, retention, compliance, customer satisfaction, cost reduction, or another strategic priority.

↓

Capability Model
Identify the capabilities, skills, behaviors, knowledge, and decision-making abilities people need in order to influence that business objective.

↓

Trusted Knowledge
Ground the learning system in approved Sources of Truth such as policies, procedures, product knowledge, expert content, standards, regulatory requirements, and internal documentation. AI should operate within this governed knowledge framework rather than rely only on what the underlying model may already know.

↓

Learning + Practice + AI Agents
Develop the required capabilities through courses, interactive activities, simulations, scenarios, roleplay, coaching, practice exercises, and AI agents.

AI can support learners dynamically by adapting practice, providing feedback, asking questions, challenging reasoning, and recommending the next intervention based on demonstrated needs.

↓

Assessment & Evidence
Capture evidence that the required capability is actually developing.

This can include assessment results, simulation behavior, decisions made during scenarios, practical tasks, attempts and retries, observed performance, feedback, and practice history.

The objective is to move beyond measuring course completion toward measuring what the learner can actually demonstrate.

↓

Competency Progression
Track how the learner progresses against the capability model.

The system should identify whether capability gaps are closing over time, for example:

Awareness → Understanding → Application → Proficiency → Mastery

This creates a continuous view of capability development rather than a single assessment score.

↓

Operational Data
Connect learning and competency evidence with relevant business systems and operational data.

Examples include:

CRM → sales, conversion, customer activity
Customer service systems → escalations, resolution time, satisfaction
ERP → productivity, efficiency, output
Quality systems → errors, defects, rework
Safety systems → incidents and near misses
HRIS → retention, mobility, performance, time-to-competence

↓

AI Interpretation & Performance Analytics
AI analyzes learning evidence and operational performance together to identify meaningful relationships.

For example:

• employees demonstrating stronger negotiation capability may show higher conversion rates;
• employees reaching competency faster may become productive earlier;
• employees struggling with a specific simulation behavior may also show higher customer escalation rates;
• teams demonstrating stronger procedural knowledge may generate fewer operational errors.

AI can therefore help identify patterns, capability gaps, performance relationships, and potential interventions at a scale that would be difficult to manage manually.

↓

Business Impact
Measure whether improved capability is associated with real business outcomes such as: higher sales, higher productivity, fewer escalations, fewer errors, faster time-to-competence, lower incident rates, better quality, higher retention, lower operating costs

↓

ROI Measurement
Translate the observed business impact into financial value using the ROI model that was defined before the learning intervention.

This is where the organization compares: Expected benefit → Actual benefit and Expected ROI → Actual ROI

↓

Continuous Improvement
Use both learning evidence and business-performance evidence to continuously improve the intervention.

The full loop becomes: Business Objective → Required Capability → ROI Objective / ROI Model → Trusted Knowledge → Learning & Practice → Evidence → Competency Progression → Operational Performance → Business Impact → Actual ROI → Continuous Improvement

The key change is that learning is no longer treated as an isolated activity. It becomes part of a continuous capability and performance system, with AI helping connect trusted knowledge, learning evidence, operational data, and business impact. This means ROI becomes part of the learning loop rather than a retrospective reporting exercise.

9. Human Judgment and Attribution Still Matter

There is an important limitation: correlation is not causation.

If sales increase after training, learning may have contributed but other factors may also have played a role. Prices may have changed, market conditions may have improved, marketing may have generated better leads, or the sales compensation system may have changed.

Organizations therefore need appropriate attribution methods. Depending on the importance and scale of the initiative, these might include control or comparison groups, pre/post measurements, manager and learner estimates, trend analysis, business-data analysis, or more rigorous experimental designs.

AI can help connect and analyze this evidence, identify patterns, and make attribution more systematic. But it cannot eliminate the need for human judgment and methodological discipline.

This remains one of the fundamental challenges of learning ROI: ATD reports that 90% of respondents see isolating the effect of learning on results as a challenge.

But imperfect attribution should not become an excuse for not measuring impact. The objective is not necessarily to prove with absolute certainty that learning caused a specific business result. It is to build a sufficiently credible body of evidence connecting learning, changes in behavior or performance, and business outcomes.

The result may require interpretation and moderation, but it still represents a considerable step forward from measuring only completions, attendance, satisfaction, or assessment scores. Over time, combining learning data with operational and business data can create a much stronger basis for estimating learning ROI, improving learning investments, and ultimately helping the organization perform better.

10. A Practical ROI Framework for L&D

The principles described throughout this report can be translated into a practical framework for designing and measuring learning ROI.

Importantly, the process should begin before the learning intervention is designed, not after it has been delivered.

A practical framework can be structured around nine questions:

The critical change is that ROI is not simply calculated at the end. It is designed at the beginning and measured throughout the learning and performance cycle.

The organization defines in advance what value it expects learning to create, what evidence will be required, which assumptions will be used, and how performance improvement will eventually be translated into financial value.

This creates a complete chain: Business Objective → Business KPI → Required Capability → ROI Model → Learning & Practice → Capability Evidence → Performance Change → Business Impact → Actual ROI → Continuous Improvement

The framework also recognizes that ROI does not always mean incremental revenue.

Depending on the business objective, value may come from:

1. Revenue creation: higher sales, conversion or customer value.
2. Cost reduction and productivity: less time, fewer errors, faster onboarding or reduced rework.
3. Risk mitigation: lower probability or impact of compliance, safety, quality or operational failures.
4. Strategic capability creation: developing capabilities required for future business performance.

The further the financial value is from directly observable performance, the more important it becomes to make assumptions, attribution methods and uncertainty explicit.

The objective is therefore not to manufacture a precise ROI percentage. It is to create the most credible evidence possible about whether learning contributed to business value and whether the investment should be continued, changed, scaled or stopped.

This better reflects the report's earlier distinction that a KPI tells us whether something changed, while the ROI model determines what that change is economically worth.

10. The L&D Role Changes Too

This evolution has an important consequence for the L&D profession.

As AI reduces the time and effort required to create, translate, adapt and maintain learning content, the value of L&D increasingly moves upstream toward business and capability analysis and downstream toward performance, measurement and continuous improvement.

The role progressively shifts:

• Less “How do we build this content?” to more: “What business problem are we trying to solve, and what capability needs to change?”
• Less “Did they complete the course?” to more “Can they demonstrate the required capability?”
• Less “Did they score well on the assessment?” to more: “Did that capability transfer into workplace performance?”
• Less “Did they like the training?” to more “Did performance and the relevant business KPI improve?”
• Less “How many courses did we produce?” to more “What measurable value did learning contribute?”
• Less “Can we calculate ROI after the program?” to more “What ROI model should we design before we invest?”

This means L&D increasingly becomes a capability and performance partner to the business, rather than primarily a producer and administrator of learning content.

That requires new competencies within L&D: understanding business KPIs, defining capability models, working with operational data, designing meaningful evidence, understanding attribution, collaborating with business leaders and using AI to identify relationships between learning and performance.

AI does not remove human judgment from this process. On the contrary, as learning systems become more capable of detecting patterns and correlations across large volumes of learning and business data, human judgment becomes increasingly important in determining what those relationships mean and what decisions should follow.

The opportunity for L&D is therefore significant.

AI can automate more of the production and analysis. L&D professionals can spend more of their time on the areas where human expertise creates the greatest value: business understanding, instructional judgment, capability development, interpretation, governance and performance improvement.

This also connects much better to your new “Human Judgment and Attribution” section, where you correctly emphasize that AI can analyze evidence but cannot eliminate methodological judgment.

11. Conclusion

From Learning Activity to Business Performance

Learning ROI should not begin with a financial formula. It should begin with a business objective: what needs to improve, what capabilities are required, and how learning can contribute to that change.

The chain becomes: Business Objective → Required Capability → ROI Model → Learning & Practice → Evidence → Performance → Business Impact → Actual ROI → Continuous Improvement

Completion, engagement and assessment scores remain useful, but they are signals, not the final evidence of value. The real question is whether learning improved capability, whether that changed performance, and what business value resulted.

That value may come from increased revenue, greater productivity, lower costs, faster time-to-competence, improved quality or reduced risk.

AI-native learning infrastructure can make this connection increasingly practical by linking learning and competency evidence with operational business data, helping organizations identify patterns, capability gaps and performance changes. But AI does not eliminate the need for human judgment: correlation is not causation, and attribution will rarely be perfect.

The goal is therefore not necessarily a mathematically perfect ROI figure, but a credible and continuously improving body of evidence connecting learning to business value.

This can transform ROI from a retrospective reporting exercise into a continuous optimization mechanism: Learning → Capability → Performance → Business Value

And ultimately shift the L&D question from: “How much learning did we produce?”

to: “What changed because people learned and what value did that change create?”

That is the real ROI of learning.

Annex C: AI-Native Learning Infrastructure Maturity Index

Purpose of the Index

Organizations increasingly say they are “using AI in learning,” but this can describe very different realities.

For some organizations, AI use is limited to occasional experimentation with tools such as ChatGPT, Claude or Gemini. Others are already using AI to generate learning content, support instructional design workflows, create simulations, deploy AI agents or connect learning with enterprise knowledge and analytics.

The purpose of the AI-Native Learning Infrastructure Maturity Index is to provide a structured way to assess where an organization currently stands in its AI adoption journey for learning.

The objective is not simply to ask:

Are we using AI?

but rather:

At what level of maturity are we today, what capability do we want to develop next, and what infrastructure, governance and measurement do we need to get there?

The Six Levels of AI-Native Learning Maturity

This is a strategic assessment framework designed to help organizations understand their current level of AI-learning maturity and identify the capabilities, governance and connections required for their next stage. It is not an independently validated scientific maturity model.

Level 0 — No AI

AI is not used in any structured way within the learning function.

Learning creation, delivery, assessment and administration rely primarily on traditional tools and processes.

Typical characteristics include:

• no formal AI tools used by L&D;
• no AI-related learning workflows;
• no AI governance framework;
• limited internal experimentation.

At this stage, the key question is generally whether, where and under what conditions AI should be introduced.

Level 1 — AI Experimentation

Individuals begin experimenting with general-purpose AI tools such as ChatGPT, Claude, Gemini or similar systems.

AI may be used occasionally to:

• generate ideas;
• summarize documents;
• draft learning objectives;
• create quiz questions;
• rewrite content;
• brainstorm activities.

However, usage remains largely individual and disconnected.

There are typically no standardized workflows, governance rules, shared knowledge bases or systematic measurement.

The organization is experimenting with AI, rather than operating an AI-enabled learning system.

Level 2 — AI Content Generation

AI becomes a regular part of learning-content production.

Typical use cases include:

• course outlines;
• learning sequences;
• quizzes and assessments;
• scenarios;
• training materials;
• scripts;
• summaries;
• interactive activities.

The primary value proposition at this stage is often speed and productivity.

Organizations can create more learning content, faster and at lower production cost.

However, AI is still largely used as a production tool.

Learning creation, delivery, assessment, learner support and analytics frequently remain separate.

The main question begins to shift from:

Can AI help us produce learning content?

to:

How do we use AI throughout the learning process?

Level 3 — AI-Assisted Workflows

AI starts supporting multiple stages of the learning and instructional-design process.

Rather than being used only for isolated content-generation tasks, AI becomes integrated into broader workflows.

Examples may include:

• content analysis;
• instructional design support;
• course creation;
• assessment generation;
• feedback;
• learner support;
• content adaptation;
• learning recommendations.

AI begins to support the professional workflow of instructional designers, trainers and L&D teams.

However, many systems and processes remain disconnected.

The organization has moved beyond isolated AI use, but AI is not yet operating as part of a fully connected learning infrastructure.

Level 4 — Connected AI Learning

AI begins connecting previously separate parts of the learning environment.

For example:

Trusted Knowledge → Creation → Practice → Assessment → Learner Support → Analytics

AI agents may support learners and creators.

Learning data can begin to inform interventions.

Assessments, simulations and practice activities provide evidence of capability development.

Enterprise knowledge can be connected to learning experiences.

Different systems may begin exchanging information through APIs, connectors or emerging interoperability frameworks.

At this level, AI starts becoming part of the learning architecture, rather than simply a feature within individual tools.

The main challenge becomes orchestration.

Level 5 — AI-Native Learning Infrastructure

At the highest level of maturity, AI is embedded across a secure, governed and connected learning infrastructure.

The organization connects:

• trusted knowledge;
• multiple AI models;
• AI agents;
• learning creation;
• interactive practice;
• assessments;
• competency evidence;
• analytics;
• learner support;
• business systems;
• performance data.

AI operates within defined rules, permissions and Sources of Truth.

Human validation remains central.

Governance includes areas such as:

• approved knowledge sources;
• model selection;
• data access;
• permissions;
• versioning;
• traceability;
• human approval;
• monitoring;
• security and compliance.

The learning system becomes capable of continuously connecting:

Business Objective → Capability → Learning → Practice → Evidence → Performance → Improvement

At this stage, the organization is no longer simply adding AI features to traditional learning tools.

It has built an AI-Native Learning Infrastructure.

Relationship to existing maturity frameworks

This Maturity Index draws on the general logic of established external frameworks rather than proposing an entirely new methodology. The Josh Bersin Company's 2026 Learning Maturity Model similarly describes a progression from static, compliance-driven training toward what it calls "dynamic enablement," where learning is embedded into the flow of work rather than delivered as a separate event. Gartner's AI Maturity Model applies a comparable five-stage logic more broadly across enterprise AI adoption, from Foundational experimentation to Transformational, where AI reshapes decision-making and operating models. Our Level 1 to Level 5 index adapts this general pattern specifically to learning technology, and should be read as a domain-specific interpretation of a well-established type of framework, not as an independent or unprecedented model. We reference these frameworks for context and credibility, not as an endorsement of Mexty by Bersin or Gartner, and the level descriptions used here are our own.

Sources: The Josh Bersin Company, "New Research: How AI Transforms $400 Billion of Corporate Learning" (Feb 2026), https://joshbersin.com/2026/02/new-research-how-ai-transforms-400-billion-of-corporate-learning/ ; Gartner, "AI Maturity Model and AI Roadmap Toolkit," https://www.gartner.com/en/chief-information-officer/research/ai-maturity-model-toolkit

The Maturity Journey

The maturity journey can be represented as:

No AI
↓
AI Experimentation
↓
AI Content Generation
↓
AI-Assisted Workflows
↓
Connected AI Learning
↓
AI-Native Learning Infrastructure

These levels describe increasing degrees of maturity, but they should not be interpreted as a mandatory sequence of implementation.

An organization may currently be at Level 0, 1, 2, 3 or 4 and decide to move directly toward a Level 5 architecture if the required governance, infrastructure, integrations and operating model can be put in place.

The real progression is therefore not simply about adopting more AI tools or passing through each level one by one.

It is about increasing the degree of connection, orchestration, governance, evidence, interoperability and integration with business performance.

The important question is not:

“Which level must we pass through next?”

but rather:

“What capabilities, controls and connections are missing today to reach the architecture we want?”

In that sense, Level 5 is a destination and a target architecture, not a compulsory sequence of steps.

Why the Index Matters

Many organizations are currently adopting AI rapidly, but adoption does not necessarily mean maturity.

An organization may use many AI tools while still operating primarily at Level 1 or Level 2.

The maturity question therefore becomes:

How deeply is AI integrated into the learning system itself?

The index helps organizations identify:

• where they are today;
• which capabilities are missing;
• what their next realistic step should be;
• which governance mechanisms are required;
• which systems need to become connected;
• and how AI use can eventually be linked to capability and business performance.

From AI Adoption to AI-Native Learning

The long-term evolution is therefore not simply:

Less AI → More AI

It is:

Isolated AI usage → Connected AI workflows → Governed AI learning infrastructure

This distinction is important.

The strategic objective is not to maximize the number of AI features being used.

It is to build an environment in which AI can support learning safely, continuously and measurably while humans remain in control.