The Autonomous Learning Ecosystem

The question I hear most often when I describe a course-less L&D department is not about technology or strategy. It is: what happens to us? It is a fair question, and it deserves a real answer rather than simple reassurance. The work does not disappear; however, it does morph. It moves upstream, becomes more consequential, and demands a fundamentally different set of skills. This piece is an attempt to map what that actually looks like in practice.

AI adoption in L&D, for a lot of organisations, has amounted to ADDIE with a jet pack. Faster content production, cheaper translation, and auto-generated quizzes. It is not nothing, but not really something. It does not address the gaps that existed before the tools arrived. We still largely lack personalisation at scale. Practice and experimentation are still bolted on as afterthoughts, if they appear at all. And measurement still tends to stop at completion, which is the digital equivalent of “if a tree falls in a forest, does anybody hear?”

I have a line I use with clients: "AI won't solve your business problem if you don't know what the problem is. It'll just automate your confusion at scale." Simply doing ADDIE faster is not a transformation. The Autonomous Learning Ecosystem I have been developing is less a dismantling of previous operating models, and more an optimisation of the new technology available to us. And yes, there is still a Human-In-The-Loop (HITL) as the necessary glue to make things work.

This Is a Supply Chain Problem

In The Eight Levers of EdTech Transformation, I introduced the idea that skills management should borrow its operating logic from supply chain management. Not because people are products, though I understand why that framing makes people uncomfortable, but because supply chain thinking forces the precision that L&D has historically avoided.

If you have followed me at all, then you have certainly heard this analogy: Imagine you run a pizza factory. Your most critical ingredient is tomatoes. You need to decide whether to grow them yourself, buy them from suppliers, or use canned tomatoes for the things where freshness does not actually matter. Each option has different costs, lead times, quality implications, and shelf lives. Now replace tomatoes with skills, and you have the beginning of a completely different way to think about what a learning function is for. And yes, I do like pineapple on pizza. We can talk about that later.

So, this is what I mean by Rightskilling: the deliberate, evidence-based process of developing exactly the skills that are needed, in exactly the people who are positioned to develop them, through exactly the mechanism that the economics and the evidence support. This is a methodology we have successfully implemented for a number of our clients over the years. Only now do we have tools available to enable it at scale.

To do this requires a symbiotic ecosystem. In my mind, it has two loops. The first is the larger strategic one. The second is the continuous improvement engine that runs once learning is in motion. It looks a bit like this:

The Outer Loop: What Has to Happen Before L&D Gets Involved

Step 1: Sense Demand. The first input into the system is way upstream from L&D. In fact, L&D does not come on stage until Step 4. Instead, we start with business signals dissected for insights. These would include strategic plans, OKR cycles, performance data, market shifts, workforce planning outputs, signals from systems like Workday, SAP SuccessFactors, etc. All of this data feeds into a picture of what skills the organisation will need, when, and where. If you are not reading these signals and translating them into a skills forecast, everything that follows is guesswork dressed up as planning. Basically, it is giving you way more information than a biased conversation with a business partner during an annual planning meeting. This is real-time, pulse checking of the health of the corporation.

Step 2: Skills Intelligence Layer. Once you know what you need, you need to know what you have. Yes, I know. You want to measure tasks instead of skills. Honestly, you could measure pounds, litres, cubits or stones. The important part is that you are measuring and taking action against something to tell you what is in your workforce wallet. In this case, we will work with skills intelligence platforms. These are building inventories from work evidence in ways that were not practically possible until recently. The output of this step is a living, self-healing, skills map of the organisation, not a competency framework that lives in a SharePoint folder and gets reviewed annually if you are lucky (as per above).

For both of these steps, tech already exists. I do not want to make this an inadvertent advertorial for anyone, but a friendly chat with your favourite LLM should give you some results.

Step 3: Gap and Adjacency Modelling, and the Buy-Build-Borrow-Bot Decision. Here is where it gets interesting, and I do a little jumpy clap.

The gap is simple enough: the delta between what you need and what you have needs to be closed. So, build a course! Not so fast, little grasshopper. Rather than immediately routing everyone with a skill gap into a development program, you model who already holds a neighbouring skill close enough that targeted upskilling could close the gap faster and at lower cost than starting from scratch. I was part of a restructuring scenario where customer service agents had deep product intelligence and brand positioning skills that made them naturals for social media management roles. By following the adjacency rather than the gap, we retained roughly 70% of employees initially at risk of redundancy. That thinking IMHO is both more commercially sound and considerably more humane than the alternative (the-don’t-puppy-shop,-adopt, equivalent of the HR World).

Once you have the gap and adjacency picture, you make a deliberate sourcing decision. Build the skill internally? Buy it through hiring? Borrow it through a contractor, partner, or secondment? Or…Bot it. Meaning, automate it away entirely, because the company only needs the outcome. Annee Bayeux, one whip-smart cookie, added that fourth option to the traditional equation. chef’s kiss.

If we think back to our pizza factory, automating a skill is like having canned tomatoes on hand. They might not be as good quality as fresh, but in some cases, it is all you need to get the job done.

Finally, before anything touches L&D (I know you are twitchy; be patient), there is also a calculation to make. I walk clients through a version of this that considers five things: what it would cost to build the skill in-house (including the hidden costs most people forget, such as subject matter expert time and productivity lost while people are learning); the potential ROI if the skill is developed; the arithmetic of return minus cost; the cost of doing nothing, which is usually the most revealing number and the most frequently ignored; and finally, what it would cost to hire the skill on the market instead. More importantly, how many people need the skill should factor into the calculus. This prevents wasting time in sheep-dip programmes (the sheep do not like them)

Only when that calculation points clearly toward build does it make sense to route a skill gap to L&D for development.

Now, I know I am painting a slick picture, and most of this is conceptual. There would most certainly need to be a HITL here. Bias and defensibility of the AI models used will be paramount. But much of this technology and data exists in disparate locations and solutions, and could be within reach. Now, clearly not all gaps are due to lack of skill; they could also be situational, environmental, etc. People are also not widgets to be programmed and deprogrammed. That said, this scanning, sensing, and decisioning allows for L&D to do an electric slide into the process at precisely the right time.

The Inner Loop - Steps 4-7

Step 4: Generate Learning and Practice.

Here’s where it gets a little more untraditional. What I described above is really just good performance consulting on steroids and Ozempic. Typically, this would result in a course. However, one-size-fits-all courses are not effective. They worked for a period, but now, we have so many possibilities. We can retire the woman-holding-clipboard and click-next-to-continue. Behold, the bidirectional AI tutor.

Soon, and already out on the interwebz, people are learning via a prompt bar entry point. This is a massive and fabulous shift. However, if left without any bidirectionality, we have essentially recreated the effectiveness of a massive course library buffet. You can learn anything, but a candlemaker is not going to succeed at a basket-weaving company.

This AI Tutor knows the KPIs of the business. It knows which skills are tied to which outcomes. Basically, with all of the insights from Steps 1-3, it actively guides people toward the capabilities the organisation needs. More importantly, the learning is delivered through interaction, practice, and feedback, not passive consumption, and it has memory. It is not a one-time event but rather paced and repeated over time. It is personalised to the specific gap, the specific learner's context, and their workflow.

This is not science fiction. Google DeepMind's LearnLM, now infused into Gemini, takes this further. In a 2025 exploratory randomised controlled trial conducted with Eedi across five UK secondary school classrooms, the human-and-AI tutoring team proved as effective as expert human tutors at correcting mistakes in real time, and produced knowledge transfer gains 5.5 percentage points higher than human tutors working alone. The evidence base for this model of delivery is building.

If L&D is struggling with its new role and purpose, I would posit that it is here: building, training, refining, orchestrating, this tutor. Sure, there might be one to procure off the shelf. However, it will require care and feeding, burping and soothing, so to speak. This is where L&D professionals who embrace a future-forward mindset will thrive: creating the architecture for targeted learning experiences.

So, what happens to all of our courses? Well, regulatory content will probably need to stay the same until compliance requirements mature. As for existing libraries, the best investment would be a SCORM scraper. This tool can extract all of the content locked in SCORM so it can be ingested into an LLM. If you want me to recommend a few, HMU on InMail.

Step 5: Deliver

The AI Tutor described in Step 4 is only as effective as the ecosystem it lives inside. And this is where a great deal of the current EdTech vendor market is making a strategic miscalculation.

Building a standalone AI Tutor and asking people to come to it is, functionally, building a better portal. Adoption will follow the same arc it always has: initial enthusiasm, declining engagement, eventual abandonment. The ghost towns of enterprise learning libraries did not fail because the content was bad (well, some of it was pretty dire). They failed because they required a behaviour change that most people, under most conditions, were not willing to make.

The vendors who understand this are not building destinations. They are building integrations. Coursera became the first online learning platform embedded directly inside ChatGPT, not because their content got better, but because they understood where hundreds of millions of people were already spending their time. The logic is not complicated: if learning has to compete for attention against the tools people use to do their actual jobs, learning will lose. So, stop competing. Be where the work is. I know you want the biggest piece of the pie. It could backfire.

getAbstract drew the same conclusion from the enterprise side. Their Copilot Connector threads verified business knowledge directly into Microsoft 365 Copilot, surfacing curated summaries inside Copilot Chat, Outlook, Word, Teams, and PowerPoint, with no app-switching required. The learning does not ask to be found. It appears that the work is already happening. Mic drop.

My toonie for thought (a Canadian $2 coin, adjusted for inflation): Any vendor whose current roadmap ends at a proprietary interface should be asking itself a hard question: why would someone leave the tool they are already in to come to yours? In the absence of a compelling answer, the integration strategy is not optional. It is existential. Closed-loop systems are always at risk of implosion - think AOL, Blackberry, and Google+

For organisations, this reframes the procurement dilemma entirely. The right question is no longer which learning platform has the best content library. It is the learning ecosystem that connects most seamlessly to where your people already are, and how does your AI Tutor become part of that fabric rather than a separate thread? To EdTech CEOs, it might not be palatable to be middleware; however, I would anticipate delivery via Copilot or Gemini.

Two more steps to go.

Step 6: Measure

Measuring the impact of learning in a corporate environment is genuinely hard. Not because we lack ambition, but because we lack the conditions that make clean measurement possible. A randomised controlled trial in a secondary school classroom can isolate the learning variable with a precision that most corporate environments simply cannot replicate. Two plus two is four. You can either apply the quadratic equation or not (I cannot, but I can recite the lyrics of the entire musical canon of The Clash).

In our world, people are managed differently, motivated differently, and operating under conditions that shift constantly and unpredictably. When performance improves, learning may be one of twenty contributing factors. When it does not improve, learning may have worked perfectly while everything around it failed.

So, what can we actually measure? Behaviour change. Are people doing things differently after the intervention? Are they applying the capability in context? Are the leading indicators that the business cares about moving in the right direction? To be explicitly clear, these are imperfect proxies, but they are connected to outcomes a P&L owner can read, and they are infinitely more meaningful than completion rates.

We will not achieve the clean causality of a controlled trial. But ecosystems generate patterns, and patterns are actionable. The goal is not to prove that learning caused the outcome. The goal is to locate the signal reliably enough to act on it, adjust, and improve. That is a different standard of evidence than academia demands, and it is the right one for our context. It requires instrumentation, patience, and the organisational will to look honestly at what the data is actually saying rather than what we hoped it would say.

And this is where it gets uncomfortable (sorry, people)

What happens when the data challenges something we believe in? What if the models, which many of us have built entire operating frameworks around, do not hold up against behavioural signals at scale? This is another area where L&D can transform. This data can be crunched by AI, but the insights and subsequent actions will need learning professionals who have the skills to dissect how people are learning.

Step 7: Optimise

This brings us to the final step in the loop.

To understand what Optimise actually means, it helps to understand what machine learning actually does. If you are already au fait with the concept, scroll on, Comrade! At its most basic, machine learning is pattern recognition at scale. You feed a system data. It looks for what works. It adjusts. You feed it more data. It adjusts again. Over time, the system does not just process information; it gets better at processing information. It learns from its own outputs. It improves with use.

Now apply that to this autonomous learning ecosystem. Every time a learner interacts with the AI Tutor, that interaction is data. Every time a capability gap closes, or fails to close, that is data. Every time a behavioural signal moves in the right direction, or does not, that is data. If this is built correctly, a properly instrumented optimising system ingests all of it, identifies what is working, weights it more heavily, identifies what is not, and recalibrates.

This is already happening in academia and K-12. Systems like Khanmigo from Khan Academy, and Socratic by Google have been doing a version of this for years. They maintain internal learner models, update them continuously, and shift their pedagogical approach based on what those models reveal. It is like a teacher who gains experience from being in the classroom for years. I was once a high school teacher, and trust me, I can spot a kid lying about bunking off school in a nanosecond.

However, the Optimise step only has integrity if we are willing to follow the data somewhere inconvenient. A system that learns but whose outputs are filtered through our prior convictions is not an optimising system. It is a confirmation engine with better tooling. The fortitude this model demands is not technical. It is the willingness to let go of a principle we have held for years because the evidence, gathered at scale, across real people doing real work, is pointing somewhere else. Can we really prove that listing learning objectives at the start of a module really contributes to outcomes? What about that drag and drop?

The mantra of 8Levers for 2026 is: “To keep people curious is noble. To keep people employed is divine”. If the data tells us that what we have been doing serves the first goal but not the second, we do not get to look away. The people on the other side of our decisions are trying to stay relevant in a labour market that is moving faster than any curriculum we have ever designed. That is the responsibility sitting underneath this step. And it is bigger than any pedagogical principle we might have to surrender. My apologies in advance to Kolb, Gange, and Ebbinghaus.

After all of this, the fancy footing of Steps 1-7, we should land in a place where we actually have the acquisition of a skill. This could take an hour or months. But when it is demonstrated and validated, we zip back up to Step 1 to update the inventory of skills. This cycle will be on continuous repeat, constantly calibrating and rebalancing.

Conclusion

Phew. That was a lot to write. I know I missed pieces, and others I am deliberately being obtuse about. There is more thinking around agentic and orchestration layers. Governance and maintenance are critical, and so on. However, “why buy the cow when you can have the milk for free?” So, if this type of thinking intrigues you, then you know where to find us (psst - here)

The question at the beginning was: if we do not build courses, what do we do? After seven steps, the answer is this. We sense, model, decide, guide, deliver, measure, and optimise. We build the conditions for skill acquisition rather than the content of it. We become, finally, the function that the business has always needed us to be.

The workforce does not have the luxury of waiting for L&D to find its footing. The decisions we make about how to develop capability have consequences that land on real careers.