The rise of the AI-augmented solopreneurs

Over the past couple of years, we’ve seen the quiet but quite general rise of AI-augmented solopreneurs or freelancers. This is not a brand new “species” per se, so much as the evolution of a category that already existed: the content creator, the independent consultant, the solo-developer, the one-person agency, the coach, or the small one-person trades business. This category had already expanded as the creator economy grew into a c.$250bn market of 200 million active creators and as hundreds of millions micro or small enterprises emerged worldwide with no or a handful number of employees.

All of this existed long before generative AI. What has changed is the toolset and the level of leverage a single person can now access, from setting up the business to upskilling themselves and finding new customers. This new AI-native or AI-augmented solopreneur is not a single, one-dimension persona. It can be a senior operator leaving a startup to start a one-person boutique practice, a content creator productizing their knowledge, a marketing generalist running a small portfolio of clients, a tradesperson or craftsperson selling directly to consumers. Their common trait however is that they work outside a traditional employer-employee model, are effectively their ‘own boss’, and leverage AI to build their own stack of tools and platform to produce, sell and run their operations. AI agents are now pushing this to a new level. This new phenomenon of the future of work naturally is giving rise to an entire generation of AI companies servicing this new worker, from portage and back-office platforms, to mission marketplaces, vertical CRMs, AI “copilots” for admin and finance, and a dense learning ecosystem.

Importantly enough, this is not happening in a vacuum. The rise of AI-augmented solopreneurs is the result of at least two deeply rooted macro drivers: a structurally imbalanced labour market, and the rapidly decreasing cost, both CAPEX to build and OPEX to operate, of starting and running a company.

The increasingly imbalanced labour market is driving people towards entrepreneurship

The job market is in a deep and ongoing imbalance. Over the past 5 or 6 years it has gone through a series of structural shifts that organisations and workers are still digesting. AI brought about an even bigger one. This transformation is the result of at least two long term paradigm changes.

The disrupted relationship to where and how we work

Covid forced a full-scale experiment in hybrid or remote work. Between 2019 and 2021, the share of workers that went full remote skyrocketed, especially in professional services and tech (see statistics by the US Census Bureau below).

Many firms now rolled it back entirely or at least attempted to roll it back. Some did so very publicly, like Jamie Dimon from JP Morgan (which now has a 5-day in office mandate). Adam Mosseri, from Instagram, sent out a companywide memo (called Building a Winning Culture in 2026) stated that to make Instagram “more nimble and creative”, employees from US officed with assigned desks were expected to be in office 5 days a week. But we’ve ended up a bit in an unstable middle (for now at least), where employees have tasted remote and hybrid culture with more flexibility and organisations are now trying to rebuild in-person culture. According to a Deloitte Survey from last year (see here), GenZ & Millenials “continue to value flexibility in where and when they work, and this drive for flexibility continue to increase the popularity of less traditional employment models”. In the US, 64% of employees would prefer hybrid or remote roles over being in the office every day according to a 2025 study on RTO policies. At the same time, younger cohorts entering the workforce are digital-native and, for the most recent, AI-native. They expect modern tools, transparency, and a real say over how they spend their time.

AI and the broken social mobility ladder

The economic engine inside firms looks less attractive. Wage mobility is low, in Europe especially, internal progression is viewed as constrained and opaque, and the skills obsolescence keeps accelerating. When you adjust for inflation, rising living costs, the traditional employer model is or at least is perceived as a lot less appealing.

On top of this structural stagnation, the last few years have brought repeated waves of layoffs and hiring freezes, especially in tech and professional services sectors. Teams that looked strategic one year can be shut down the next.

Most importantly entry-level jobs are now progressively impacted by AI. Recent analysis of AI-exposed occupations from Stanford researchers shows that employment for 22 to 25 year old has already declined in these roles by 16%, while employment for older workers in the same categories has increased, which suggests that automation is affecting the bottom of the pyramid first rather than the top.

On the one hand, it now becomes harder to start a career; on the other it remains difficult to advance in a career inside an organization. This combined effect reinforces the sense that the traditional bargain (“if I stay and perform, I’ll be rewarded and protected”) feels less and less credible. When wage progression is limited and mobility overall low, the feeling of stagnation acts as a psychological push. Trading perceived job security for higher-risk, higher-reward paths, in terms of both income and ownership, increasingly looks like a rational choice rather than an exception.

Cost to build and operate have significantly decreased with AI

If the labour market is the push factor behind the rise AI-augmented solopreneurs, AI itself is the main pull factor: it has broken the barriers to company creation, and the costs to start, build and even operate a company.

Infrastructure costs to launch a company decreased significantly. What used to require heavy capex and a lot of custom work (from legals to software development), is increasingly available off the shelf, via AI-tools. For most solo and very small businesses, “setting up the company” is no longer a multi-year, capital-intensive project. It is a small stack of tools that can be turned on in a week or less.

As a blog post from A16Z puts it “AI is already delivering 10x product experiences for a 1/10th of the cost”.

The second effect is on skills. Gen AI and AI-enabled no code tools allow non technical profiles to get to the MVP stage or an even fully working product via vibe coding and dedicated platforms. The overall skills requirements to startup are thus decreasing. Developers who do write code use AI assistants to move much faster.

AI also automates a large share of what used to be a startup’s first hires. The same person can now use AI to write code, design visual assets, produce video, draft contracts, handle a large share of customer support and manage accounting. For a solopreneur, opex have dropped. In a set of OECD countries, roughly 30–40% of SMEs use some form of AI, and about one quarter use generative AI specifically.

The aspirational narrative has also shifted. A decade ago, the reference point for ambitious talent was joining a hypergrowth company with hundreds of employees. The narrative now revolves around the “3-people unicorn”, a term first used by NfX.

All of this compresses time-to-market and cost-to-build. Launching something on the side, testing whether there is demand and iterating before making any big commitment has become a normal path rather than an exception.

The AI-enabled Solo Stack

This new demand of solopreneurs is serviced by a growing supply of companies making this shift viable. For the first time an individual can leverage AI to build their own operating system, across every critical function of a company: administration and legal, product building, go to market definition and monetization and learning and decision making. Taken together, these tools compress the gap between an individual and an organisation. They make one-person companies operationally real. Most importantly perhaps, the rise of agentic AI is increasingly changing the economics of launching a solopreneur company, turning the entrepreneur into an orchestrator of agents and software.

Admin & Management: removing the friction of independence.

This layer addresses the historical tax of going solo: administrative complexity and compliance risk. Banking and company-creation platforms increasingly bundle accounts, cards, invoicing and accounting integrations into “company-in-a-box” offerings. Legal and status platforms productise incorporation and regulatory compliance. Many of these product, that were not necessarily AI-first now embedded AI agents in their offering. They thus operate as an AI execution layer, running tasks end to end and increasingly taking responsibility for what is non core.

Product Building & Operations: compressing time-to-MVP and iteration cycles

This is probably where AI most visibly changes the equation. AI-native builders, AI IDEs and vibe coding platforms allow individuals to design, build and ship real products with minimal (to no) technical overhead.

On the operations side, automation tools orchestrate workflows that used to require coordination between multiple people. Analytics and feedback tools allow continuous optimisation without a dedicated data or product team.

This is one the most dynamic layer of the stack today, with companies emerging (and growing fast) across the entire product value chain, from UX and prototyping to coding, testing, deployment, and continuous feature iteration.

Solving the distribution and monetisation challenge

Finding initial demand is often one of the biggest challenges. For service-based companies, marketplaces have been aggregating demand for a long time helping freelancers or solopreneurs finding their first customers. Their core mission has not changed much but how they operate did with AI: matching is now smarter, AI can actually help solopreneurs broaden their offering or adjust their price point/profitability profile (doing more missions for less time and effort etc).

More generally, AI-powered sales and marketing tools are progressively demonstrating their ability to help scale operations (prospect identification and niche targeting, improved outreach, tailored messaging, autonomous follow up etc). Initial revenue gets generated much earlier and is less and less a function of how much money is spent on outsourced sales people and marketing talent. This is shown by a very interesting study by Gusto (available here): more than 75% of solopreneurs turn a profit in the first while the vast majority of them are self-funded.

Learning and deciding: reducing solo decision risk and closing skills gap faster

One of the most difficult parts of going solo is often linked to decision making. Solopreneurs have no managers to validate choices, no team to challenge assumptions and no cofounder to help on difficult trade-offs. In practice, many solopreneurs lose time in over-researching, and acting too slowly on important decisions (see below):

A new set of tools is emerging to make solo decision-making more efficient and more effective. Increasingly, AI acts as a decision-support layer by synthesizing data, summarizing feedback, and suggesting next actions or priorities.

The learning infrastructure is increasingly adapted to the learning habits of the solopreneurs: rather than long, one-off programs, there is an increasing amount of platforms designed to fill specific skill gaps on demand on key elements of building their business from pricing, product design and marketing to operational finance and how to use AI in daily operations. Training becomes a lightweight way to unlock the next step in execution, not a separate phase detached from day to day business. Solopreneurs can decide faster, with less uncertainty, and acquire the minimum set of skills needed to keep growing.

A new way of building a company

The emergence of AI solopreneurs is not necessarily a by-product of the creator economy as much as it’s a cultural shift. It’s the result of a much deeper alignment between demand and supply. On the demand side, AI is reshaping the labour market, compressing entry level jobs and breaking down the traditional career path. On the supply side AI is equipping solopreneurs with a full stack of tools covering all the key critical aspects of building a business.

The question now is what type of AI company is going to own the platform with the critical workflows that these solopreneurs have. For solopreneurs, the challenge is no longer whether independence is possible, but how far it can scale.

Equipping AI-augmented Solopreneurs was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

We’re more or less three years into ChatGPT (Nov-22), and one of the most visibly disrupted spaces is, unsurprisingly, education. It’s fair to argue that generative AI changed almost everything for edtech as a second-order consequence.

Since 2017, Educapital has been investing in and closely following the evolution of edtech, which we see as unfolding in three successive waves.

(1) the first wave was about digitization. Early edtech companies focused on getting the best content online, on digital platforms (from LMS to mobile apps). It broke down the technological barriers to access

(2) the second wave (the paradigm shift) saw edtech go global, with the first category leaders emerging and going public (e.g. Coursera, Duolingo, Udemy all did their IPO in 2021). The pandemic played a key role in this ‘acceleration of history’ breaking down the psychological barriers to adoption. Learning online with dedicated apps and products became a cultural norm.

(3) The widespread adoption of GenAI (with ChatGPT being the fastest consumer product reaching 1M users, in 2w, 12.5x faster than the second fastest one, Instagram) helped us enter into the third wave of edtech, which we called the “intelligence layer”. For edtech it enabled structural changes: content is getting increasingly commoditized, adaptive learning (one of the biggest promises of edtech) is now possible, and so is real-time feedback with AI tutors. As a result, a whole new generation of AI-native companies emerged and every existing company embedded AI in their product. The promise is now that the equity barriers to access will drop, as AI tutors will solve the 2-sigma problem for example, making tutoring more widely available.

Technological Progress Is Never Neutral

Technological progress, including AI, is rarely (if ever) neutral. This had been first formulated by M.Kranzberg back in 1986, in a paper on the history of technology and what he called the Six Laws of Technology.[1] The first one states that “technology is neither good, nor bad; nor is it neutral (…) the same technology can have quite different results when introduced into different contexts or under different circumstances”. This observation feels particularly relevant today. Much of the current debate around AI in education revolves not only around who has access to these tools, but also how they are built per se (what assumptions, datasets, or pedagogies they embed).

Even OpenAI’s new Head of Application, made the case in a blog post published in July 2025: “If we get this right, AI can give everyone more power than ever. But I also realize those opportunities won’t magically appear on their own. (…)That’s why we have to be intentional about how we build and share these technologies so they lead to greater opportunity and prosperity for more people.” And about the impact of AI on knowledge acquisition and learning, she added:

“It’s already working: people who use AI tutors learn twice as much as they do from human ones, and the gains are even bigger compared to learning in a traditional classroom. In a 2024 OpenAI study, 90% of users said ChatGPT helped them “understand complex ideas more easily.” Once we put a personalized AI tutor on every topic at everyone’s fingertips, AI will close the gap between people who have the resources to learn and people who have historically been left behind.”

This is a very exciting statement and observation but it does lead to a few questions. What actually happens to how we learn, when AI is always here to guide, or more importantly think for us?

Cognitive debt and offloading

Some of these questions were raised in a pre-print scientific article by MIT researchers, released in June 2025 (“Your Brain on ChatGPT”, N. Kosmyna et al, available here). Before getting into what this article actually shows, two small caveats: 1) it’s a pre-print, that has not been peer reviewed yet, and is based on a relatively small number of participants from a relatively similar background. 2) they are not trying to assess whether chatgpt makes us dumber, they look at neurocognitive activity to track what happens in the brain when we rely on AI to think or write, vs. when we use a search engine or no help at all.

In the study, 54 participants were divided into three groups: one using ChatGPT (LLM-only), one using a search engine (Search-only), and one writing independently (Brain-only). They used EEG to measure neural activity during essay tasks of these three groups.

All participants produced overall coherent work, but their brain activity differed sharply. ChatGPT when used without thinking first, homogenized outputs, making them smoother but more alike. Group 3 participants (Brain-only) produced essays with greater variability and originality (not necessarily better ones though).When participants used ChatGPT without thinking first, EEG scans revealed lower neural connectivity , a sign of what the authors call cognitive offloading, or delegating mental effort to AI (“ “a passive approach and diminished activation of critical thinking skills when the person later performs tasks alone”). Although ChatGPT improved surface-level performance, it also created what they describe as cognitive debt: users later showed weaker recall and reduced engagement when working alone.

The article does not draw dramatic conclusions or state at any point that AI makes us dumber. The study produces early results, based on a small group of participants, assessed on one task only. Most importantly perhaps, the last part of the study focuses a lot on sequence: the effects on brain activity depends less on using AI in itself than on when it’s used. The fourth group of the study (Brain-to-LLM) who started by thinking about the essay then used chatgpt showed some increased neural engagement.

What the MIT study ultimately suggests is not that AI harms learning, but that it transforms it, unsurprisingly. The challenge then becomes to channel that transformation and the tensions they create. They give way to three paradoxes on the consequences of AI on how we learn.

Three Paradoxes on AI & Learning

The Inverse Jevons Paradox of Learning

On our learning behaviour

Back in February 2025, everybody became obsessed with the Jevons paradox, which states that increasing efficiency of a resource (e.g. GPU), instead of leading to lower consumption of that resource, can actually lead to higher consumption. AI in education might lead to an inverse Jevons paradox, with potential adverse effects:

AI promise: AI will make learning more efficient, more personalized and more accessible.

Paradox: when AI can retrieve and summarize information for us, or even directly solve problems for us, our perceived need to learn might decrease, as we increasingly delegate more to AI and and retain less.

The Flattening Paradox

On efficacy

Teachers and researchers (see MIT study above) are increasingly noticing that AI homogenizes outputs, thereby flattening performance. We increasingly sound the same, using the same words, AI-polished syntactic structures and even ideas (see here, or here).

AI promise: AI will undoubtedly democratize access to knowledge and learning tools including AI-based tutoring, feedback, coaching. This should help students improve their performance.

Paradox: instead of leveling up students, AI may flatten performance, bringing the average level of students up to a certain ceiling, while capping excellence.

The best students may learn less deeply, thinking less independently, and becoming less differentiated.

The Back to Basics Paradox

On the meaning of why we learn

AI Promise: AI will enable a deep reinvention of how we learn at scale. Integration of AI allows for more precise or data-driven learning and teaching models.

Paradox: By redefining the modalities according to which we learn, AI is progressively pushing us to go back to fundamental questions on what learning means, and more basic question (how do we learn, why do we learn).

Product Design Matters in AI-Based Learning

What does that mean for the widespread adoption of AI in education?

First, that we’re still at the beginning of a structural shift and that there are still a lot of outstanding questions. ChatGPT is not making us “dumber”, and we had the same question about Google back in the early 2000s. But how we use these products has a direct impact on what and how well we learn.

Second, how the very products are designed affects greatly the use we make of them when learning. OpenAI and Gemini released a StudyMode this summer, more adapted to learning purposes. OpenAI’s Study Mode and Gemini’s Learning Mode, bring a socratic, step-by-step reasoning to guide learners instead of providing them with direct answers.

And finally, it reminds us that AI-based edtech goes (or needs to go) much further than this. Building an AI edtech product means having an adapted product & UX design suited for learners, pedagogical choices (spaced repetition, assessment and practice etc) and specific education-focused use cases. Knowunity, in our portfolio, has developed a full suite of AI-enabled features that cover all the needs a student might have: quizzes, mock exams, homework scanning and flashcards, social features etc.

The coming years will show whether AI can evolve from a tool of efficiency to a driver of deeper learning. For AI-based edtech, the challenge is not simply to embed large models, but to design products that leverage them through more sophisticated pedagogical approaches. The most promising companies will use AI to enhance cognitive engagement and learning outcomes, embedding strong pedagogy into thoughtful, age-adapted UX that turns AI’s potential into real learning impact.

Learning in the Age of Intelligence was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

Tutoring works but still fails to scale

When it comes to effective knowledge transmission in K12 education, personalized tutoring remains almost unparalleled. Bloom’s famous 2-sigma problem, demonstrated in the 1980s, showed that students receiving 1:1 tutoring perform two standard deviations above their peers in traditional classrooms, effectively placing them in the 98th percentile (source).

Despite its proven effectiveness, tutoring has long remained inaccessible to most students. High costs and logistical barriers have limited it largely to higher-income families. The online shift accelerated by COVID-19 removed geography as a constraint and helped decrease overall market prices for tutoring lessons, but it left the fundamental model unchanged: one-to-one, synchronous, expensive, and available mostly to higher-income families. In short, traditional tutoring has not scaled.

MyEdSpace Is a Game-Changer

MyEdSpace represents a genuine breakthrough by reimagining this paradigm. The platform combines live streaming, an engaging format popularized by the gaming industry, with top 1% UK teachers and pedagogy enhanced by AI. The result is an unmatched level of student engagement and interaction. MyEdSpace does not merely digitize traditional tutoring; it transforms educators into stars, placing them rightfully at the center of the educational experience. Teachers become true anchors of the learning journey, while AI tools and a robust mentoring system ensure stronger learning outcomes. Simply put, it works.

Since launching in 2022, MyEdSpace has supported more than 16,000 students. In 2025, 51% of its GCSE Maths students achieved grades 7 to 9, more than triple the national average (source). MyEdSpace has also built a following of over 4 million across social media, breaking down complex concepts in formats students already consume every day. Since our first investment in summer 2024, MyEdSpace has experienced exceptional traction, increasing accessibility and enrollment by at least a quartile.

Backing an Exceptional Team

Our conviction in MyEdSpace extends beyond their innovative product to their stellar founding team. The two co-founders Kharis Yanakidis and Sean Hirons are exceptional, execution-focused leaders. They have successfully cultivated a high-energy company culture, deeply rooted in a shared mission and driven by ambitious, scalable goals that align perfectly with our investment vision.

We are thrilled to reaffirm our commitment to MyEdSpace by participating in their latest $15m round, led by White Star Capital. The round will allow the company to hire additional teachers, including in the US where they started to expand, further build out their AI tools and continue to launch new topics. We’re very excited for what’s coming.

Investing in MyEdSpace was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

There’s been growing excitement over the past 18 months around agentic AI. The term AI agents in itself has evolved from technical discussions in research labs and academic research into mainstream conversations. It’s now commonly mentioned in media coverage, everyday discussions and corporate strategies. Recent web search data (see below) underscores this phenomenon: online searches for AI agents spiked dramatically over the past year alone by some 900%. Hundreds (if not thousands) of high quality publications have been released discussing their potential impact.

As always, part of it is market hype, pushing companies to rebrand as ‘agentic AI for X or Y’. Behind the noise though, there is a very strong chance that AI agents will have a profound and durable impact on companies, on how they operate from customer acquisition to recruitment, upskilling and customer support. In a blog post, Sam Altman suggested that “in 2025, we may see the first AI agents “join the workforce” and materially change the output of companies”. What does that concretely mean for how organisations do business?

A simple overview of how agentic AI works

Simply put, an AI agent is a system that performs tasks autonomously, on behalf of a user (or on behalf of another system). It’s an evolution of text-based LLM that designs its own workflow and leverages available tools (see IBM’s note). In other words, they’re a bit like a layer, on top of LLMs that enables them to break down complex objectives into subtasks to do in order to perform this goal. Although research on AI agents dates back at least to the 1990s, laying down its theoretical foundations (see here for a detailed history), the advent of LLMs has opened a new era for agentic AI. That being said what differentiates an AI assistant from an AI agent, and what an agent exactly is. A Techcrunch article summarizes it pretty well:

“But no one can seem to agree on what an AI agent is, exactly (…) the tech industry has boldly proclaimed that AI “agents” — the latest buzzword — are going to change everything.(…) But it also depends on how one defines “agents,” which is no easy task. Much like other AI-related jargon (e.g. “multimodal,” “AGI,” and “AI” itself), the terms “agent” and “agentic” are becoming diluted to the point of meaninglessness (…) Leher Pathak, OpenAI’s API product marketing lead, later said in a post on X that she understood the terms “assistants” and “agents” to be interchangeable — further muddying the waters.”

The framework of an AI agent & its core capabilities

So what makes them different from AI assistants?

1. They operate autonomously: unlike convetional AI assistants, that typically rely on explicit commands and directions, agentic AI, once set up, proactively identifies goals, plans out it actions and operate autonomously with minimal guidance
2. They can accomplish complex tasks: AI agents, in theory can handle and perform complex, multi-steps tasks, by breaking them down into subtasks.
3. To do so, they rely on reasoning capabilities: AI agents leverage chain-of-thought reasoning and are equipped with sophisticated frameworks to interpret information and choose the right path of action
4. They remember information and are more accurate: agentic AI integrate stateful memory management with real-time external retrieval of information (RAG), which means that they they remember the context to provide more up-to-date, relevant and accurate outputs
5. They improve continuously through learning: they leverage reinforcement learning (inc. RL with human in the loop, RLHF) to adjust and improve performance.

Building on these capabilities, the foundational types of AI agents evolved over time. Ever since the 1990s, scientific literature distinguishes between different types of agents, based on their deterministic/non-deterministic nature (i.e. whether these agents can operate in non-predictable environments). They are typically categorized as follows:

What are the macro & micro drivers fueling adoption

The impressive development of AI capabilities is the primary driver of the surge of AI agents. Their adoption by SMB and large corporates is still in its early days but it benefits from very strong positive tailwinds, both at the macro level and at the micro level.

A macroeconomic context of low GDP growth, stagnating (or even decreasing) productivity and lack of skills or talent availability for non-developmental, repetitive tasks, is a fertile ground for agentic AI adoption:

The case is just as strong at the company level as well:

1. Lower costs for low-value tasks: for a lot of repetitive, non-developmental tasks common in the knowledge economy, agents are less expensive
2. Reliability: chain-of-thought reasoning, built-in safeguards and the use of RAG on relevant data often result in fewer hallucinations and higher accuracy than standard LLM
3. Increased speed: often faster and more consistent than humans for repetitive or defined tasks; typically more accurate and reliable compared to general-purpose (agnostic) LLMs.
4. Ease of setup: AI agents can be rapidly deployed through low-code/no-code solutions when minimal company-specific data integration is required. Even with customisation, deployment remains relatively fast. The recent introduction of Model Context Protocol (MCP, released in Nov-24, see here for the release post by Anthropic), essentially a new way of connecting AI assistants/agents to various tools, data sources and content repositories of companies makes setting up agents even easier (see here for a deep dive on MCP & Agents from A16Z).

Having said that, there’s a lot of speculation and debate around how AI agents will affect the operational processes of organisations. What will this mean for the workforce? What skills will be most in demand? What about for the organisational design of companies? What would the AI-first organisation of the future look like?

The Optimistic View: AI-first Organisation of the Future

This is what an optimistic view of the rollout of AI agents in organisations would look like:

1. Increasingly decentralized decision: with AI agents handling repetitive tasks and operational decisions, teams would be able to focus more on (i) strategic oversight (ii) complex problem solving (iii) higher level management tasks, allowing for more agile organisations.
2. Organisational scalability: assuming that agents are widely adopted and easy to deploy, facilitated by growing agentic AI tooling and infrastructure, we could imagine a future where companies scale up or down operations depending of business activity or market conditions for instance. A complete reorganisation is always very difficult to orchestrate, the massive roll out of customizable agents would in that case make it a lot more easier.
3. Smaller, more flexible teams: as a result, teams could become smaller, and more flexible, helped by multi-agent systems designed to perform tasks autonomously and with a specific focus on the function they’re designed for (from software development to marketing for instance). At the company level, the integration of AI agents would mean more flexible organisational models, where humans interact and oversee AI agents depending on task complexity of strategic priorities.
4. New metrics of performance or productivity: again assuming agents become the norm, then the way we evaluate performance will probably have to be adjusted, to factor in the decreasing production costs of tasks. Pure output-based performance metrics would not allow to capture the value of work produced with the help of AI agents. This coincides with the rise of “service as a software”, whereby agentic AI performes services for users, effectively automating labour (see here for a great article by NfX on service as a software and the AI workforce).
5. New or redefined roles: this is probably the biggest shift people anticipate. As service as a software becomes gradually adopted by companies, software becomes the worker itself, able to execute and improve traditionally human delivered services. This might mean that roles will change, from execution to execution-supervision or oversight and that new roles will be created. In the (highly) optimistic view, these roles will be more fulfilling, less repetitive and more in line with what people want to do (see next part of the less optimistic view).

Most companies however are not fully positioned to capitalize on these agentic capabilities. Like most organisational transformation, they take time and “cultural” willingness. Achieving the level of autonomy and sophisticated decision-making offered by agentic AI probably requires moving through multiple stages of AI maturity, each characterized by progressively greater degrees of automation and integration. Drawing inspiration on another framework here, this is a schematic view of what these steps would be:

Beyond technological readiness, this transition thus demands significant shifts in organizational mindset, enhanced resilience to manage the complexities and uncertainties of AI-driven workflows, and the ability to effectively drive AI adoption across all levels of the company, a core challenge for many organisations today.

The more pessimistic view: a lot of outstanding questions

While the optimistic scenario for AI agents is pretty compelling, significant hurdles thus remain. Organizational maturity isn’t there yet. Adopting autonomous systems involves considerable execution risk: companies must manage internal resistance, handle operational complexity, and deal with skill displacement. Realistically, widespread AI-agent adoption will face meaningful friction, and the impact on workforce dynamics and organizational design will probably be more complex than initially anticipated. With that in mind, let’s now explore what a more pessimistic view might look like.

1. The Labour Market Question: AI agents, if widely adopted, will most probably disrupt, and displace jobs. The question is thus whether if at the general equilibrium agentic AI adoption would lead to great job destruction than job creation. Most reports point that the balance would be negative, especially in the short term as the labour market readjust to find a new equilibrium. In the long term, new roles will probably emerge but the transition period will be disruptive. Evidence suggests it might not favor workers, especially those in entry-level knowledge economy roles, who typically rely on tasks most susceptible to automation.
2. The Skills & Learning Question: This shift has two direct implications. First, what specific skills will be most valuable and necessary in an economy dominated by AI agents? Second, how can we effectively prepare and train people for these new roles, especially considering a potential contraction of traditional entry-level positions? This scenario likely demands entirely new methods of reskilling and upskilling, but it’s unclear whether our current educational and training infrastructures are equipped to manage this transition effectively.
3. The Efficacy Question: This question is the most obvious one. Do they work? And if yes, to what extent? AI agents currently face significant limitations when rolled out on complex, lengthy multi-steps processes. The more steps involved, the greater the risk of failure. Deepmind’s CEO compares it to inversed compound interest, at a recent event. Patronus AI built out a statistical model to assess the compounding risk of AI agents, suggesting that even an error rate of 1% per state can compound to a “63% chance of error by the 100th step in an agent operation” (see post here). As companies are often slower to adopt technology in general, it’s fair to anticipate that full scale roll out will take time, posing integration and security challenges along the way. As returns on investment get more evident (and that is true for GenAI in general), roll-outs of agentic AI will become more general (see below)

One of the biggest questions of course is who will be winning the market, horizontal or vertical agents.

Most companies are likely to adopt a hybrid strategy, combining vertical agents to automate specific industry or functional tasks with horizontal agents designed to streamline internal processes and integrate workflows. Ultimately, we anticipate successful outcomes for both models, similar to what occurred in the SaaS market, where verticalized solutions coexist and integrate effectively with broader, horizontal software layers within organisations.

AI Agents & the Future of Work was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

As we look back at 2024, the year tells the story of a cautious and moderate recovery for the edtech and future of work sectors. The year opened on a similar note to 2023, marked by a continued contraction in both deal volume and overall funding value. It had not been that hard to raise funds for an edtech company for a very long time. Throughout the year though, things improved and edtech and future of work companies ended raising overall more in 2024 than in 2023. Let’s take a closer look at how the year unfolded.

The big picture

Following the unprecedented (and arguably unsustainable) frenzy of the Covid19 years (2020–2022), the market underwent in 2023 and the beginning of 2024 a brutal recalibration. Overall, investment behaviors shifted back to fundamentals, much like in 2023: valuations — apart from certain high-profile AI companies — became more conservative, and deal sizes normalized. At the same time, funding processes became lengthier, terms harsher and investors continue to prioritize, for most companies at least, operational efficiency metrics over growth at all cost. However, as Generative AI continues to shake up the industry in an unprecedented manner, the market becomes harder to read, for investors and entrepreneurs alike — price discovery is challenging and a lot of people in the ecosystem found it difficult to right-size their round.

Nonetheless, edtech and future of work companies raised a total amount of c. $7bn in 2024, up 24.5% from the c.$5.6bn raised in 2023. Mega-rounds above $100m remained very rare this year, and mostly raised by AI-first future of work companies (not all though — see below!). A closer look at quarterly data reveals an optimistic view of the funding landscape, driven the growing clarity of genAI use cases and applications in edtech and future of work, and the early signs of a revitalized exit environment.

A closer look at the funding landscape in 2024

The edtech market in 2024 showed mixed signals, with a total funding increase of +24.5% compared to 2023, driven largely by growth in the Future of Work (+25%) and Lifelong Learning (+45%) segments. However, the number of deals across these segments saw only modest increase, suggesting that larger average deal sizes contributed to the growth (+14% and +68%). For core edtech (K12, Higher Ed, Lifelong Learning) this is mostly linked to a handful of larger rounds (inc. Physics Wallah’s $210m Series B, Eruditus’ $150m Series F and private equity led rounds like PSG’s $175m investments in AI-driven higher ed admission platform Element451). As mentioned, mega-rounds above $100m remain pretty rare but made a small comeback.

Other analysts, including HolonIQ — one of the most reliable platforms for edtech market analysis — estimate a significantly lower total of $2.4 billion raised in 2024 (compared to $3 billion in 2023). The delta between our data and theirs can be attributed to our probably broader definition of edtech (we include some subtopics like mental health, financial education, knowledge platforms, skills analytics etc). The Financial Times gives an estimate of the total amount raised of c.$3bn using Pitchbook data. Overall, excluding future of work, edtech companies raised between $2.4–3.1bn this year.

At any rate, the past two years have been challenging for edtech companies when it comes to raising funds. In Europe, however deal volume increased moderately (with 262 rounds announced this year) and average deal size grew by c.37% to reach $7m. In the US, the 27% increase in amount raised is explained by a value effect (+30%) rather than by a growth in the volume of deals announced (-2%). Throughout the year, things improved and more deals got done.

How edtech & future of work stocks performed

An overview of the performance edtech listed companies

For most of the listed Edtech companies, 2024 was a rough year of post-pandemic market correction. This correction stems from overall slowing consumer demand for some categories of products, AI-driven disruption and increased investor scrutiny over efficiency of growth — the share price of the two companies with >30% Ro40 (Duolingo and Stride) outperformed the Nasdaq (+36% LTM) over the period.

Among public EdTech stocks, Duolingo (NASDAQ: DUOL) was thus the clear outperformer, surging ~50% over the past 12 months. This was driven by record subscriber growth, family plan expansion, rollout of Duolingo Max (AI), and expanding verticals (Math, Music, and English Proficiency testing, or DET). By end of Q3, Duolingo had over 8.6M subscribers (a 48% YoY growth). Most importantly even, it saw its DAU/MAU ratio increase to 32% (up from 29%), while growing its overall number of MAU to 113M (up from 83M, +36%). When a lot of B2C models struggled with subscriber fatigue and generative AI competition, Duolingo clearly showed to the market the power of its very strong brand equity, its ability to diversify its revenue streams, and its the value of its strategic positioning at the crossroads between games and edtech. As a result the competition from AI-first language learning apps like Speak have not (yet?) affected the company’s market performance. Stride (NYSE: LRN), operating in K-12 virtual learning, also showed relative stability, benefiting from consistent demand and government contracts, and its share price grew significantly as a consequence.

On the contrary, Chegg (NYSE: CHGG) continued its dramatic decline, with shares collapsing over 90% from 2021 highs, as AI-powered alternatives like ChatGPT dismantled its core homework help business. Despite efforts to integrate AI features, revenue erosion and subscriber losses accelerated, forcing cost-cutting and layoffs. Revenue were down 24% this year, to c.$615M. Coursera (NYSE: COUR) and Udemy (NASDAQ: UDMY) also struggled, as consumer spending on standalone online courses softened. Coursera issued a downward revenue revision in April 2024 and reduced its workforce by c.10%, as part of broader cost-cutting measures. Over the past years, Udemy tried to grow its B2B segment (Udemy Business) but its enterprise adoption remained inconsistent, and required large upfront sales and marketing expenses. Their core consumer offering was also affected by the overall post pandemic slowdown in course purchases and higher sensitivity to price.

An overview of the performance of selected future of work listed companies

For future of work companies, 2024 brought mixed outcomes, amid heightened scrutiny around AI. Corporate training providers like Docebo (NASDAQ: DCBO) and Skillsoft (NYSE: SKIL) faced softer demand but doubled down on enterprise deals, pivoting to more efficient cost structures — Docebo maintained a c.14% top line growth, for a 15% EBITDA margin which allowed the company to trade at a pretty decent c.6x EV/Revenue multiple. Skillsoft’s share price recovered slightly over the year but is still much below its all time high of 2022. Workday (NASDAQ: WDAY) and ServiceNow (NYSE: NOW) remained relatively resilient, leveraging stable enterprise HR and workflow automation demand, even as AI features raised some questions about their product roadmaps. Atlassian continued to integrate AI in its collaboration tools, and delivered overall good and efficient growth (+16% revenue growth for a healthy 25% EBITDA margin) — one of the best performers in the category. In the marketplace for flexible work category, while the shift toward project-based and freelance recruitment is still carrying the market, Fiverr fared better, managing to leverage its “productized approach” to freelancing (with standardized offering etc) and to capitalize on the enhanced demand for AI experts on gig, to grow its revenue 9% (and its share price by c. 24%).

Read more about what to expect in 2025, in the second part of this article, accessible here.

Edtech & the Future of Work in 2024: Funding Recovers, AI Wins (1/2) — Market Overview was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

This is the second part of a two-part analysis. See here to read the first part.

The market sentiment in 2024 reflected lingering hesitation. Growth investors remained reticent to re-enter the edtech space at scale following the overinflated valuations and rounds during the pandemic era. However, there is a slow but notable return of generalist funds to the market, signaling renewed interest, albeit with a more measured approach. This has resulted in a marked flight to quality, with larger rounds heavily skewed in companies with strong fundamentals and teams. Let’s look at what themes attracted VC money this year.

What got funded in Edtech

Interestingly enough, K12/Higher Ed deals attracted a decent amount of VC funding. Physics Wallah’s massive round demonstrated that there is still appetite for new innovative online tutoring models (1:many massive live streaming in their case, with over 15m users). More generally funding in this segment went to companies addressing identified use cases with clear consumer or education institutions budget to tap into (e.g. supplemental education for test prep, administrative management for schools, learning toys, etc). This is overall true for other segments as well within edtech beyond K12/Higher Ed. In the lifelong learning segment, online language learning B2C apps made a strong comeback with Praktika raising $20m and Speak raised an additional $98m in a back to back Series B/Series C (it became a unicorn at the same time). Both companies prioritize improving fluency by enabling users to practice speaking: they leverage conversational AI, (speech recognition, dialog management etc), alongside AI-generated avatars to create a more immersive learning experience. As Speak’s founder puts it: “let’s teach people how to communicate in the language”. We also invested in two companies operating in this space with a very novel approach (Buddy.ai and one to be disclosed soon). Mental health and soft/hard skills development remained attractive themes this year.

What got funded in the future of work space

In 2024, against a backdrop of budget constraints, higher interest rates (now easing) and intense competition for market share driving a push for optimisation, VC funding in the future of work segment focused on themes that address workforce optimisation and productivity in an AI-driven era. Next-gen core HR companies continued to attract a large share of total amount raised (HCM, payroll management, employee experience platforms). The spotlight was most naturally on AI agents and productivity tools, across a wide range of either industry-specific use cases or task-specific use cases (e.g. agents for go-to-market teams, for software development, for IT support, data analysis or employee onboarding — like 11x.ai). The appeal of these tools lies in their ability to act autonomously within predefined parameters, adapting to user needs and organizational workflows. The gains in terms of cost optimisation and overall productivity are expected to be massive. Skills development and strategic workforce planning remained topics of interest (see here for a deep dive on this topic).

AI VC investments remained (of course) extremely significant in 2024

AI investments more generally continued to surge, reaching new highs. According to Pitchbook and Techcrunch, GenAI companies thus raised c.$56bn across around 900 deals, most of it going to US companies (across the AI application, foundation models and operations and infrastructure layers).

Interestingly enough for Edtech and future of work, AI-first companies focusing on voice and video generation continued to raise sizable amounts. ElevenLabs, a leader in voice AI, raised $80m in a Series B round in January, with its valuation exceeding $1.1bn to continue developing its AI voice cloning and text to speech tech. It then raised $250m on January 27th 2025 (almost exactly a year later), reaching $3bn. In the text-to-video space, Synthesia ($180m round reaching $2.1bn valuation in Jan-2025), Heygen ($60M reaching $500m valuation in June-2024) and Colossyan ($22m Series A) all raised in 2024 or early 2025. Their realistic AI-generated video content will pave the way to a complete overhaul of corporate and upskilling training content generation but use cases go beyond L&D (marketing, onboarding videos, recruitment). We expect multimodal AI applied to edtech & future of work to become even more critical as there is still a lot of things to do in the space.

Is hardware back?

Conversational AI was a major theme in 2024, coinciding with an interesting resurgence of hardware “crossing the chasm.” In early-age education, several players launched their conversational AI toys, such as Curio, which shipped its Grok toy in 2024. Hardware remains difficult to scale though and some efforts did not succeed — most notably Moxie, which ceased operations in December 2024.

Beyond early-age education and K-12, we anticipate both on-device AI and more broadly speaking hardware-based AI to significantly reshape usage across various domains in our sectors. In November 2024, Aravind Srinivas mentioned that Perplexity, the AI search engine, was considering entering the hardware space with a voice-to-voice search device:

Similarly, OpenAI is collaborating with former Apple design legend Jony Ive on an AI device. For now, Rabbit (with its R1 device) and Humane (that struggled significantly this year between safety issues, bad reviews and low commercial performance) are still leading the way.

This new wave of hardware innovation will probably unlock a broad range of applications in the future of work and edtech, introducing new ways of learning, enhancing productivity, enabling new usage patterns in the knowledge discovery process etc. However, many questions remain: Will people adopt yet another device? Can these AI-first devices offer a compelling enough value proposition to challenge smartphone dominance? More importantly, why would users commit to a standalone AI device — potentially locking themselves into a single AI platform or foundation model — when they could instead access AI through apps on their existing smartphones or an AI-native phone with an embedded OS and advanced conversational and image-processing AI?

What to expect in 2025

AI, AI, and more AI. Most experts predict that AI’s momentum will reach new heights, driven by (i) continued breakthroughs in open-source models (see DeepSeek’s recent progress — read more in this very good breakdown by the Algorithmic Bridge). While proprietary models might still lead in peak performance, open-source AI is closing the gap, offering competitive capabilities with lower licensing and running costs and greater flexibility. These models will continue to democratize access to foundational AI by reducing barriers to deployment, particularly when optimized for efficiency. Alongside smaller, fine-tuned models, this will enable edtech and future-of-work companies to integrate AI into their products faster and more affordably, accelerating experimentation and commercialization.

Similarly, (ii) multimodal AI that integrates text, images, voice, video, and other data is rapidly gaining traction both via proprietary and open-source models. Google’s Gemini 2.0. (dec-24), OpenAI’s GPT-4o (may-24), Meta’s LLama 3.2. (sept-24), and more recently Janus Pro by Deepseek (Jan-25) have significantly pushed multimodal reasoning, real-time input processing and generation of multimodal content. We look forward to what OpenAI’s o3 (announced in dec-24, and to be released soon, o3-mini having already been released) model will enable in that regard.

This will have a major impact on edtech and future of work and we expect to see a new generation of AI tutors emerge, which will no longer rely on text-based interactions, but also will be able to interpret a student’s facial expression for instance, or their tone of voice to assess engagement and comprehension more accurately. Next-generation AI tutors will emerge as personalized “agents,” guiding learners through language practice, skill-building, and productivity tasks, while adapting to individual progress and preferences in real time. Knowunity for instance, in our portfolio or buddy.ai are leading the way.

In the workplace, multimodal AI will impact a lot of the workflow from more intuitive onboarding and training, to deeper productivity gains via autonomous agents.

More generally, we expect to see more partnership between education actors (especially in higher ed) and AI leaders globally, on the model of the one concluded by the California State University and OpenAI, which will “allow CSU students to integrate AI into their studies, while faculty can use it to streamline administrative tasks (…) [and to] equip students with essential AI skills to succeed in an increasingly AI-literate workforce” (source).

Agentic AI (iii) will thus continue to be a major trend, as it opened an entirely new way of working paradigm (especially in countries that experienced slower producitivty growth, and thus contraction in GDP growth…). Will we move from AI assistants to purely autonomous agents not requiring inputs from workers to manage tasks end to end or will they coexist?

At any rate, by interacting with a company’s data and systems and making decisions under set constraints, agents will have a major impact on the way companies operate, paving the way for the continuous growth of a lot of these new actors automating repetitive processes. For research-grade knowledge generation, Deep Research Gemini (an AI research assistant/agent released in december 2024) and OpenAI’s Deep Research (launched in February 2, 2025) pave the way for new applications in AI for scientific research. OpenAI’s Deep Research is an agentic capability based on the o3 model, conducting, autonomously, multi-step research tasks. These developments might signal a shift towards AI systems that not only provide well-referenced and sourced responses but also offer more nuanced and opinionated insights, which is especially important for scientific research.

We remain exceptionally bullish on the transformative impact AI will have in the coming year — 2025 promises to be a pivotal moment for Edtech and the future of work, with unprecedented potential for innovation, growth, and (hopefully) real-world outcomes.

Edtech & the Future of Work in 2024: Funding Recovers, AI Wins (2/2) — Trends and What to Expect… was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

The debate about teachers’ working conditions is not new. For several years, if not decades, numerous research institutes, surveys, and NGOs have been issuing warnings about the deterioration of these working conditions. The general consensus is that teachers are overworked, underpaid in most countries and under-equipped to do their job. In the UK for instance, a report showed almost “three-quarters of teachers said their workload was unacceptable” (Source). Almost 25% of surveyed teachers declared working 60h/w or more. In the US, according to this study, teachers worked 53 hours per work, for an average salary that only 34% of them said was acceptable (vs. 61% of working adults). In the UK, this unsustainable situation is leading to 81% teacher considering leaving the profession.

A new path to better allocating teachers’ time

One solution is thus to raise the salaries of teachers. This increase should nevertheless be backed up by a policy to improve working conditions.

In particular, this means rethinking and optimising (i.e. reducing) the allocation of teachers’ time to tasks that they regard as repetitive on the one hand, and of little value in the knowledge transmission on the other. According to a McKinsey study teacher spend only 49% of their time in direct interaction with students:

Against a backdrop of a general decline in student level, as highlighted by the results of PISA 2023, and difficulties in recruiting teachers, this optimisation of time allocation seems all the more critical.

On the other side of the fence, students, and in particular the younger generations, experience a double frustration with the assessment process. Firstly, many believe that the rating and assessment process is subject to numerous biases. Scientific research has clearly established the frequency of these biases, as well as their nature — they can be name-based, background-based, or linked to the previous performance of the student in question, among other things [See here, here and here for instance]. At the same time, a number of assessment techniques and procedures have been put in place other the years to reduce these biases: anonymous exam papers and double marking, for example. One of the most effective methods, however, is to reduce the opportunities for biased decision making in the grading process.

Secondly, the grading process is time consuming in itself. This leads to visible inefficiencies, with students often waiting a very long time for their marks and feedback. This would not in itself be a problem if the assessment did not allow certain mistakes or limits in the student’s understanding of a topic or concept to be identified and corrected early enough to prevent students from repeating them. But one of the advantages of grading and assessment is precisely that it enables to assess a student’s level at a given moment, so that they can improve.

Not a new theme but new tools

Tech-enabled evaluation and grading is not a new theme in itself. The edtech sector has often seen this segment as a fertile ground for innovation: the technological and adoption barriers are relatively low, and the gains for users are immediate. Numerous tools have enabled teachers to assess their students using multiple choice questions or simple assessment formats. Since Kahoot, among others, gamified and collaborative assessment has become widespread.

Some subjects lent themselves to this more than others — those where answers could be assessed using a true/false system, for example. But the widespread access to generative AI has opened the way to new, more complex automated assessment practices that are better able to reflect the different ways in which students are assessed (open-ended questions, inverted classroom, practical exercises, etc.). In this respect, the assessment value chain is experiencing the emergence of new players while established players are launching new modules leveraging generative AI to reduce the workload of teachers and improve the assessment process as a whole.

The Assessment Value Chain & Landscape

1. Exam & Exercises Creation

The process of exam creation encompasses all the steps involved in crafting an assessment document. This category of products’ features range from assisting teachers in formalizing documents, quizzes by editing typos, questions, and scoring to generating AI-based documents that can autonomously create questions tailored to the course content. AI-enabled tools allow instructors to engage students more effectively by personalizing exams to suit individual learning styles and preferences. These advancements also facilitate increased collaboration among students and foster inclusivity through gamification features, resulting in a more dynamic and engaging learning experience. For example, Nolej allows teachers to introduce a variety of different exercises format (find & drag the word, open ended quizzes, flashcard exercises, question sets, crosswords) just like an enhanced authoring tool.

2.Exam Delivery & Proctoring

Online proctoring, which became a major theme during COVID-19, has emerged as a critical theme in Edtech for maintaining exam integrity in remote learning environments.

This sub-market is dominated by a handful of key actors (inc. Measure Learning — ex. ProctorU), and has been consolidating over the years. Key drivers of success remain the efficacy and the reliability of the tech rolled out. In recent years we saw these actors leverage advanced behavioral tracking, eye monitoring, and AI-enabled detection of writing styles, along with browser integrity monitoring. Innovation also extends to the pedagogical formats of proctoring (live, pre-recorded, and hybrid approaches) to cater to diverse educational needs and preferences.

3. Plagiarism Checker

Just as the proctoring sub-market is not new in itself, the plagiarism sub-market has existed for a number of years. With the emergence of generative AI and sophisticated writing assistants, the challenges have multiplied. The primary purpose of these tools is not to punish students; instead, their aim is to promote genuine understanding and knowledge. Nevertheless, research from Stanford indicates that this technological advancement has not led to an increase in cheating in high schools. As Victor Lee from Stanford GSE writes: “But I think it’s important to point out that, in Challenge Success’ most recent survey, students were also asked if and how they felt an AI chatbot like ChatGPT should be allowed for school-related tasks. Many said they thought it should be acceptable for “starter” purposes, like explaining a new concept or generating ideas for a paper. But the vast majority said that using a chatbot to write an entire paper should never be allowed. So this idea that students who’ve never cheated before are going to suddenly run amok and have AI write all of their papers appears unfounded.”

Two types of solutions dominate the market — plagiarism checkers for academic institutions, the leader being Turnitin, and reformulation and paper-writing support tools for students.

4. Grading, Assessment & Feedback

AI in edtech as written extensively (see here) serves as an assistant, facilitating tasks for both educators and students rather than replacing them outright. One prominent application is thus in grading and assessing papers, offering significant time-saving benefits. McKinsey reports that grading and feedback consume approximately 6.5 hours out of the available 50 hours per week for educators. Utilizing AI technologies can potentially reclaim three hours of this valuable time for other activities, including personalized feedback and student mentoring on specific tasks.

With the advent of GenAI we’re seeing a wide range of sophisticated point solutions for grading and assessment that allow for enhanced automated personalized feedback and to push the right exercises to the right student. As such, they act as a teacher multiplier of sorts, increasing the teacher-to-student ratio, while alleviating their workload. Simultaneously it allows to reduce the grading time and thus help students get access to feedback much more rapidly.

Automated grading is still in its infancy, and we expect to see many companies scaling up in this sub-segment of edtech. If you’re building in this space, drop us a line!

AI-enabled Assessment & Grading Landscape was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

Over the past few weeks, there has been much debate in France and elsewhere about the place of screens in children’s lives, both in and out of school. For many, screens, and technology with them, are at the root of falling academic standards, cognitive and learning disorders, whatever the use to which they are put. They not only criticize a certain passive use of screens (which is a fair argument), but the integration of screens and technology per se into children’s lives. Many of these arguments are sound and evidence-based. There’s indeed a strong rationale in favor of protecting the youngest children from overexposure to screens and what they imply (especially social networks), and of enabling them to develop in ways other than through the intermediary of a screen. This is especially true for children younger than 3, according to the WHO. But on this subject, as on others, there’s a considerable gap between media and political rhetoric and the actual findings of scientific research. Many empirical studies conclude that, while overexposure to screens is dangerous, after a certain age, it’s especially the way children use them that poses a problem (see [1], [2] for instance). Social networks adverse effects have for instance been widely documented, all the more so in the context of increasing screen time across age groups globally. While parental control is essential, especially at the youngest ages, ensuring that this screen time is used properly is even more important.

Passive consumption of low-quality content in particular is singled out for criticism. In this respect, many edtech products, whether screen-based or not, make it possible, on the contrary, to put technology to good use, to counterbalance this passive consumption with an active, engaging and instructional use.

Over the last 5–6 years, we’ve seen the emergence of a new category of products, ranging from connected learning-focused toys to screen-free storytellers, designed to stimulate children’s creativity and involve parents in the process. Scientific research in pedagogy and learning sciences has stressed the importance of stimulating children’s creativity from an early age (see [3] for example). The benefits are many and range from the development of soft and emotional skills, to thinking skills and logical reasoning ability. A creative mind is often synonymous with better interaction with peers, fostering empathy and collaboration. Creative stimulation can reduce learning disabilities by enabling children to discover new ways of understanding and assimilating the world around them.

The approaches of these products differ, as do their underlying value proposition and philosophy.

Mapping the Children Creativity Tech Landscape

Since the widespread availability of generative artificial intelligence, many solutions, mainly app-based, have sought to leverage this technology to enable children to express their creativity by giving birth to visual universes. Other approaches, such as the ones adopted screen free storytellers, by definition audio-based, appeal to children’s imagination and could be viewed as an extension of the children’s book. Most of the products in the mapping are consumer-first, although some are looking at the possibility of selling to schools or publishers (B2C2B) as an additional growth driver.

The 4 categories identified do not exhaustively cover a fast-growing media-like market. This market is at the intersection of several industries: the book for children market, the edutainment and video game industry, the toy market, and the animation industry. But these products thus face the challenges associated with each of these different markets:

• a difficult go to market, whereby parents are the final buyers but children are the users
• a challenging consumer retail and media market, in the context of high-inflation and reduced household spending
• a high sensitivity to the advertisement landscape, and rising cost of acquisition globally
• an increased competition for available free time, with incumbents, like Disney or Netflix being in an exceptionally strong position (see graph from Benedict Evans 2023 presentation below)
• a limited potential defensibility, intrinsically, and models with seemingly intangible success factors

What are thus the drivers of growth and success in the four categories identified?

1. Video Games for Children

As with most video games, products in this first category must above all be good games, both in terms of gameplay and the virtual world they develop. That in itself is a major challenge, given the growing expectations of gamers. This first challenge is therefore compounded by the game’s educational contribution. Many products fail to do both — children learn but do not have fun, or conversely they have fun but do not learn. PowerZ’s aim is to bring the two together, creating a Triple A game that is educational at the same time.

The graphic universe and IP developed by the game become elements of defensibility and user retention. These models are capex-intensive, because they require a depth of content that keeps players playing long enough to monetise them. This raises the question of appropriate distribution and the business model to adopt (free-to-play vs. one-off, platform-based publishing vs. indirect distribution, etc). Nevertheless, games that manage to develop within this balance can become real success stories, taping into an underserved and growing market.

2. Screen-free Storytellers

Operating between the toy, book, audiobook and educational content markets, screen-free storytellers have seen a major boom in recent years. These products are based on the production, often in-house, of content and are therefore IP-driven in themselves. The quality and density of the content is therefore essential to their success. The content itself must be able to compete with screen-based products. One of the keys to success (but also what makes this model difficult to scale) lies in the hardware through which the content is distributed. It must be seen as an essential toy for the child — and the quality of the materials and design plays a role. But at the same time, these companies need to maintain an attractive margin profile. Internationalisation is also sometimes difficult, requiring not only content translation but also adaptation to cultural habits.

One of Lunii’s successes has been its ability to move from a transactional, one-off model to a subscription model, based on the distribution of content, which serves as an upsell driver. Some products, like Edurino, use screens and add a hardware device on top of it, to enable engaging digital learning.

These products, as Matthew Ball puts it when analysing Disney, have to build a love brand, and monetize this love:

Businesses based around storytelling franchises (rather than “movies” or “media”) excel based on an intangible sort of operating leverage. Because it doesn’t actually “cost more” to make someone “love your content more”, but the “value of this love” is substantial, companies like Disney benefit from enormous “returns to marginal affinity”. And the more opportunities to exploit this love — and the better you are at doing so — the greater the leverage. Source

3. Gen-AI Based Content Studios

We have therefore seen the emergence of a new category of products, mainly app-based, that use generative AI to enable children to create interactive stories and give birth to entire universes. In a way, these products are new studio models. Some of these products rely on the commitment of parents, who are more or less in charge and involved in the process. The product category is therefore young, and it is not clear whether the companies will be able to establish themselves on the landscape in a sustainable way. Nevertheless, their potential is immense. It relies on the ability to produce stories smoothly, to leverage the LLMs intelligently and add value, and to create a user journey that goes beyond just a demo (where parents only use it once to test). Their monetisation paths are also diverse — direct to consumer like a typical B2C app or B2B2C through online publishers for instance.

4. Mental-health Focused Apps

The final category consists of solutions that use technology to support the development of children’s mental health. This trend, particularly important in North America, is beginning to emerge in Europe. Solutions such as Aumio, for example, develop sound universes to help parents help their children sleep better, dream better, relax better, etc. The key driver for success is therefore the actual outcome on the children’s health. The therapeutic approach is particularly interesting. However, it requires some sort of medical validation and it entails specificities in terms of narrative and go-to-market (should they go through physicians and doctors? do they position themselves like a wellness app? are they competing with doctors per se?).

In any case, this market, which is still in its infancy, bodes well for the future. If you’re building in the space, we’d love to talk!

Fostering Children Creativity with Edtech was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

It’s no secret that the global workforce is currently confronted to a monumental skills shortage crisis. The estimates vary but all point in the same direction. Most organizations are facing significant skills needs. According to a Deloitte report, “73% of business executives expect to continue to experience talent shortages over the next three years”. Another study by Korn Ferry finds that the talent shortage could cost up to $8.5tn to organisations globally by 2030. Employees need to adapt, update and acquire new skills faster than ever before. According to a 2021 Gartner report, 58% of the workforce needs to acquire fresh skills to effectively perform in their job. This amounts to a growing skills mismatch (i.e. the misalignment between the labour supply and the labour demand), which has been widely documented by empirical research.

Compounding this issue, the very nature of work is undergoing a structural shift. At an individual level, employees’ aspirations have undergone profound changes. Career mobility as well as the alignment of skills, work, and values are becoming increasingly important, particularly among younger generations. This gave birth to phenomena like the “Great Resignation” and more recently, “Quiet Quitting”. Amidst this evolving landscape, organizations are experiencing intensified competition for talent in an increasingly challenging economic environment.

Organisations find it difficult to source, retain and manage their talent pool. Worse, they face challenges in assessing their skills base. This difficulty leads to multiple strategic constraints — insufficient identification of future skills gaps, limited flexibility on the operationalization of projects in the context of changing business priorities, impossibility to efficiently design leadership succession career plans at the organisation-level etc. The importance of skills within organisation is thus growing. The emergence of skills-based organizations has become a major theme in business literature in recent years. The static view of work (where an individual equals standardized knowledge equals a task) is being replaced, at least in theory if not yet in practice, by an evolving perspective in which skills become the criterion by which companies evaluate and allocate their talent pool.

In response to these monumental challenges, a new wave of tech-enabled products has emerged, poised to facilitate the widespread adoption of skills-based strategies in talent management, recruitment, and resource allocation.

The Promise of Tech-Enabled Skills Mapping Products: Rethinking Talent Management at Scale

The emergence of skills-based organizations is prompting a comprehensive redesign of traditional processes across the entire talent lifecycle.

These tech-enabled skills mapping and strategic workforce planning tools aim to do so in 4 key areas.

Skills Assessment, Detection & Mapping

The value proposition of a lot of these products is to empower organizations to dynamically map their available skills. Unlike static taxonomies, dynamic ontologies theoretically evolve with the constantly changing skills landscape of the workforce. This enables real-time monitoring of the skills pool and facilitates detailed comparisons with peers, both at the individual and organizational levels. A lot of companies in the mapping below leverage AI to scale their skills inference — scrapping through employee data (annual performance reviews, CV, internal docs and projects they worked on) to extract their skillset (and even sometimes the proficiency levels of these employees in these skills, although this remains a challenge). This skills data layer finds application both at the functional level (within business units) and at the HR-level, enabling them to build out their skills architecture and competency mapping.

Talent Recruitment, Deployment & Allocation

This skills architecture then serves as a foundation and can be broken down into several use cases, the first of which being skills-based talent sourcing and candidate assessment. These products aim to place skills at the core basis of recruiting. It allows (i) for more unbiased processes (fostering diversity) and enables (ii) reduced cost-to-hire due to limited early mismatch between job description and pre-selected candidates.

A second use case relates to the allocation of talent, once recruited, within the company by placing skills, internal requirements, and individual employee aspirations at the forefront.. There are many applications for this, depending on the company’s value proposition: skills-based staffing, internal talent marketplace, functional or geographic mobility, etc.

Talent Development & Career Planning

Internal Talent Marketplaces and skills-based staffing within these products empower employees to actively chart their career trajectories within the organization, effectively reducing turnover. Data-driven upskilling and reskilling plans support employees in their career progression, ensuring alignment with the company’s evolving needs.

Many companies have set up competency-based career pathing schemes for their employees (that fuels into the design of key success profiles), in order to break the hire-fire-or-retire cycle and improve employee engagement. Conversely, by matching skills to roles and career paths, companies can maximize their talent potential.

“Skills-based practices can help employers upskill workers and provide learning opportunities to enable internal mobility and boost retention. Employers can design customized onboarding programs (adapting existing programs where possible) to meet new hires where they are and ensure they have the skills to succeed in the long term, provide on-the-job training and continuous-learning programs, and develop internal road maps to promotion from entry-level roles without requiring a degree.” (Source: McKinsey, 2022)

Organization-wide Strategic Workforce Management:

These products empower HR departments to adopt a more strategic approach to their workforce management. They assist in forecasting future skills requirements and gaps through succession planning, dynamic recruitment plans, and more precise business forecasts, thereby enabling HR to optimize their resources efficiently.

As Josh Bersin puts it, this is the premise: “Under the covers of this concept is the idea that we can “tag” or “assess” everyone’s skills with laser precision. And many of the AI tools (…) promise to do this today. How do they assess our skills? They use the magic of AI to look at our job history, performance, work product, and other sources to infer, model, and predict what we’re good at, what we’re exceptionally good at, and what we need to learn next.” But the reality is different, as he points out. (Source)

Challenges and the Road Ahead

While tech-enabled skills mapping products hold immense promise, several challenges must be navigated for their successful adoption:

Insufficient Tech Maturity: despite significant advances in AI-enabled technology for skills detection and inference, it remains in its early stages. The intense competition for AI talent has led to uneven progress among players in this space. Assessing proficiency levels, with no human intervention is still challenging for some products. So is inferring soft skills, that are more difficult to assess.

Cultural Reluctance & Employee Adoption: the transition towards skills-based strategies often necessitates deep-seated cultural changes within organizations. This cultural shift can be a protracted process, demanding commitment from all levels, especially HR and managers, to embrace the new paradigm. Similarly, achieving accurate skills mapping heavily relies on attaining high employee adoption rates. Convincing employees to actively engage in these processes remains an ongoing challenge. Some companies are exploring innovative strategies to encourage employee participation.

Go-to-Market Strategy & Market Readability: we’re seeing a lot of different go-to-market strategies in the space- bottom up via line managers to HR, top-down from corporate to business units. Identifying the dedicated budget and pinpointing the appropriate persona within organizations, be it Learning & Development or Talent departments, can be a complex decision fraught with uncertainty. The proliferation of players in this space, each with seemingly similar value propositions, has generated low market readability for HR buyers, and we expect consolidation in the years to come or integration from other players going up the value chain (e.g. 360learning buying eLamp).

Skills mapping landscape

As the workforce landscape continues to evolve at a rapid pace, skills mapping products are poised to become indispensable tools for organizations striving to navigate the monumental skills shortage crisis effectively. However, the organisations’ cultural adoption (at the HR/corporate level as well as at the functional business level), the go-to-market strategies of these players (L&D Top Down ? Bottom Up through business units? entering via one use case like recruitment vs. full packaged offering?) and the way the market will evolve (further consolidation? One use case-focused category leaders vs. all-in-one skills platform from hiring to offboarding?) are exciting market developments to monitor.

If you’re building in that space, do not hesitate to reach out to me :)

Skills Mapping & Strategic Workforce Planning Tools was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.

The pandemic-induced disruption alone cannot account for the unprecedented decline in performance

It’s widely acknowledged that the structural challenges the global education landscape faces extend far beyond the sole ramifications of the Covid-19 pandemic-induced disruptions. From resource constraints, unequal access to high-quality learning content, to outdated teaching methods, under-digitization and curriculum disparities, education systems around the world face significant hurdles in their quest to provide quality education to all. These challenges are now well-documented by scientific research. The 2022 PISA results provide further quantitative evidence of these issues in OECD countries.

Overall, the state of global education is worrying. The global trend the report highlights is that average student performance in reading, mathematics and science is decreasing:

“Some 25% of 15-year-olds in OECD member countries — representing 16 million children — are estimated to be low performers in maths, reading and science (…) they can struggle to do tasks such as use basic algorithms or interpret simple texts”

This drop in performance is unprecedented — students on average scored 10 points less in reading and 15 in maths vs. 2018. The graph below suggests nonetheless that mean student performance has been on a declining path long before the pandemic hit:

This does not mean of course that the pandemic did not negatively affect education systems across the world but it unequivocally highlights that “educational trajectories were negative well before the pandemic hit [indicating] that long term issues in education systems are also to blame for the drop in performance”.

It also provides interesting insights for policy makers when it comes to the benefits of edtech tools, when used appropriately, as pointed out in the infographic below:

What the 2022 PISA results tell us in numbers

An overview of the data in mathematics, reading and science

In most OECD countries, particularly across Europe, the prevailing trend in the fundamental subjects under examination (mathematics, reading, and science) is exhibiting a negative trajectory. The overall performance in PISA 2022 fell markedly below the average performance observed since the initiation of PISA assessments. Worse, some countries have experienced a “hump-shaped performance”, improving at first before engaging in a downward slope in recent years (Spain and Poland in Mathematics for instance, Germany and France in reading, Italy and Estonia in science)

In general, with the exception of Italy and Portugal, no European country has experienced an improvement in average mathematics performance. Below are the results in reading and in science.

Overall, the performance of European countries is negative. No country is on an upward trajectory across all subjects.

What explains the discrepancies in results between surveyed country?

Certainly, Covid-19 has contributed to the worsening of students’ average performance in all disciplines covered by the PISA assessment. On average, OECD students missed 98 days of class during the lockdowns. The short-term effects, both in terms of academic performance and mental health, have been extensively documented by scientific research (see here for instance). The medium and long-term effects, on the other hand, are just beginning to be felt and will be more directly observable in the years to come.

The PISA student confirms that « systems that spared more students from longer closures scored higher in mathematics and reported a greater sense of belonging at school.” (page 23,n Vol 2)

However, upon closer examination, the differences in school closure policies among countries (English students missed 88 days of classes, compared to 65 in Italy or 85 in Germany, and 67 in Spain), and the lack of exact correspondence with academic performance, seem to indicate that school closures alone are not enough to explain the drop in performance. France, in this regard, provides an interesting paradox — students’ performance between 2018 and 2022 declined more than in other OECD countries where the number of missed school days by students was higher.

Therefore, the reasons must be sought elsewhere and are of a structural nature. The inherent characteristics of educational systems plays a role in educational performance. Covid-19, which acted as a “stress test for resilience in education” has hence highlighted the “components for resilience”. They are numerous and varied— ranging from educational policies to the involvement of the teaching staff, and the mobilization of edtech tools.

PISA components of resilience :

Overall, there are thus various “components of resilience” and they can be categorized into two main dimensions: temporary policy choices and structural orientations.

Temporary policy choices refer to specific decisions made in response to certain circumstances during Covid19 — for instance, the duration of school closures, the attention given to students’ well-being during lockdowns, and efforts to maintain social connection and a sense of belonging despite school closures.

On the other hand, structural orientations are broader and relate to the overall design and approach of the educational system. One important aspect is the delivery of distance and online learning. The way in which these modes of learning are implemented and integrated into the educational system is crucial for building resilience. This includes the system’s capacity to prepare students and teachers to effectively utilize these tools and to develop “self-directed learning skills” that enable independent and autonomous learning. The PISA report indicates significant variance among OECD countries in this regard. In Europe, Estonia, Finland, Italy, Sweden, and Switzerland have a “solid foundation for remote and more autonomous learning.” The following table demonstrates this correlation:

In general, the difficulties in deploying remote learning are linked, on the one hand, to technical challenges (such as access to internet connection and the inability to find a suitable environment for studying) and, on the other hand, to pedagogical difficulties that could be described as instructional in nature (comprehension of assigned tasks, inability to find someone to provide assistance, etc). This second set of challenges suggests that one of the issues with remote learning is not so much related to the tools used but rather to how they are implemented and integrated into the pedagogical and learning framework. Once again, it becomes quite evident that this is not solely attributable to the Covid-19 situation:

Education systems in which fewer students reported problems with remote learning also had more positive 2018-2022 performance trends […] when analysed in relation to longer-term trends (i.e. “adjusted short term trends”), even though no significant relationship was observed to the 2018–2022 performance trends, when longer-term trends were not considered (i.e. “unadjusted short-term trends”). (page 81, Vol 2)

How edtech can help

The solutions to the challenges outlined in the PISA report will not solely come from remote learning. However, it is evident that edtech has a fundamental role to play in addressing these challenges. This role encompasses various aspects and for several reasons. According to the PISA report, it emerges that education expenditures only affect student performance up to a threshold of $75,000 per student. Beyond this threshold, there is no longer a correlation between extra investment and student performance. Therefore, the answer lies in the allocation of resources.

In this regard, by enabling pedagogy to scale, for example, through the use of AI like EvidenceB, edtech could optimize the allocation of resources. This becomes even more crucial when it comes to underperforming students who require adequate resources and support.

The PISA report highlights the value of greater personalization, the use of analytics, and the adaptive and interactive nature of learning experiences. These elements play a significant role in enhancing the effectiveness of education.

Personalization allows for tailoring learning experiences to individual students’ needs, strengths, and interests, promoting engagement and motivation. Analytics provide insights into students’ progress, strengths, and areas for improvement, enabling educators to make data-informed decisions and provide targeted support.

The adaptive and interactive nature of learning experiences allows for adjusting the content, pace, and level of difficulty to meet individual students’ learning requirements. This helps students stay challenged without feeling overwhelmed or bored, fostering a more effective learning process.

Game-based learning, mentioned in the PISA report, has gained a lot of traction in recent years — companies like PowerZ or complori being relevant examples of this.

Edtech can thus help personalize and target interventions, ensuring that these students receive the necessary assistance and resources to improve their performance, bridging the education equity gap.

Similarly, teachers, who play an essential role in learning, may not have the necessary time to assist students. The analysis of PISA highlights a paradox: the number of teachers per student has remained stable or increased in OECD countries, yet there is a strong sense that there are more teacher shortages than ever before. Tools like Nolej, an authoring tool that harnesses the power of AI to assist teachers in developing engaging educational content. By leveraging AI, teachers can streamline content creation processes, saving valuable time in their instructional preparation. This time-saving aspect allows teachers to allocate more of their precious time and attention to supporting and guiding their students. As the report points out, “the effectiveness of education can never exceed the quality of teaching and teacher support.”

Overall, and as always, it is the way edtech tools are used that matters — their impact varies depending on the pedagogical frameworks they support, their integration into the wider instructional setup, and the public policy choices made regarding their deployment.

Edtech & the 2022 PISA rankings was originally published in Educapital on Medium, where people are continuing the conversation by highlighting and responding to this story.