Roughly a quarter of the text I consume every day was written for an audience of exactly one: me. And most heavy users of AI probably are experiencing the same. That’s an entirely new phenomenon in human history. What does this mean? How does it change how we think and act?

One of the most fascinating books I have ever read is Marshall McLuhan’s “The Gutenberg Galaxy”. McLuhan is best known for his concept of “the global village” and his often quoted (and frequently misunderstood) statement “The medium is the message”. In the book, McLuhan explores how the invention of the printing press changed not only how information was disseminated, but fundamentally transformed how humans were thinking and shaping their world.

We have been experiencing similar shifts with the Internet and social media, and with AI we might be facing the most fundamental change to human thinking since the invention of print.

AI changes the dominant unit of thought substantially: from a fixed, publicly accessible text (or video, or podcast) to a private, adaptive dialogue. That makes thinking more fluid and personalized, but also less shared and increasingly shaped by our own questions, the context of our own little bubble.

In a way, AI may take us full circle to an information environment that has more in common with the conversational, individualized world of our ancestors than the industrialized, standardized world of Gutenberg. But on another level, we are entering entirely uncharted territory.

It’s worth looking at this shift with a bit of historical context.

From oral cultures to print

For most of its existence, humanity was shaped by oral communication. People interacted face to face, debated and told each other stories. The fastest information could travel was at the speed of human movement, and the precision of information transfer was very limited. Humans forget things, make up new stuff and include their own agenda in what they tell each other. This limited the scale of human endeavor significantly.

The invention of writing and most of all the printed word changed all of that. It didn’t just make the distribution of information cheaper and easier. Print changed the cognitive environment in which people lived by making information more reliable and precise. It encouraged linear and sequential reasoning, privileging the visual over the oral. It standardized languages, established the idea of a fixed, universally valid text, and made private reading a normal part of life.

Print also helped create larger shared realities. Millions of people could read the same texts, use increasingly standardized languages, and participate in common political, cultural and religious narratives. Modern institutions, nation states, mass education, global organized religions and, maybe most importantly, scientific progress would be unthinkable without print.

Intellectual interaction was transformed from ephemeral oral debates to an exchange of commonly accessible written texts. A scientist might react to another scientist’s findings with their own detailed paper, quoting many other sources. A politically engaged citizen might react to an essay in a newspaper with a long letter to the editor. Discourse centered around verifiable, shared texts. Even passive observers were able to follow these chains of arguments, leading to an explosion in the effectiveness of human thought.

Electronic mass media (TV and radio) changed the nature of information dissemination from the textual to the oral and visual and reduced the options for interaction. But one crucial aspect remained: Shared reality. When people tuned in to the evening news, they all saw the same show with the same content.

The Internet: From linear streams to fragmented, interconnected ideas

The World Wide Web was invented in 1989 and started becoming a mass medium in the mid-nineties. It fundamentally changed how people access, consume and react to information.

The most important element of the web is the hyperlink. It enabled much more instant interaction with text: On the web, we can’t just read and then maybe write our own text as a reflection, we can click from page to page, finding our own journey through a topic. Anybody who ever got lost in a Wikipedia rabbit hole has a deep intuition for how different this process is compared to linearly reading a book.

In addition, the Internet brought various ways in which people were able to interact online and react to each other’s content. The early newsgroups and web-based discussion forums were environments for global information exchange and debate on countless topics. The blogosphere, which experienced its peak in the late 2000s, was for a while a lively environment for discourse, often with significant depth.

But compared to the world of print, the Internet was a very different place for information and debate: It was faster, much more fragmented and associative, more individualistic and far less permanent. This certainly sped up the momentum of many fields, including science and technology, but it also led to fragmentation: There was no easily identifiable common ground for discourse anymore.

Social media: A relentless stream of attention signals

Then Twitter, Facebook, Youtube, LinkedIn and Instagram fractured the landscape even more: The unit of discourse went from a blog post of maybe a few hundred words to a tweet of 140 characters, a Youtube video of a few minutes or a picture on Instagram. The dominant unit of interaction went from a blog comment to pressing a “Like” button.

These short bursts of easily consumable information and social attention have a strong effect on the reward centers of the human brain. Cheap dopamine hits from social media became the dominant force shaping our information environment. Content is abundant, but attention is scarcer then ever in relation, raising noise levels and shortening attention spans.

To be clear: Of course not all of this is negative. It’s obvious that the lower barriers of content production and dissemination have enabled countless people to express their creativity and, in many cases, actually contribute truly important content. Tech Twitter (sorry, I’m still not calling it X) is currently the liveliest place of debate about AI, and its very fragmentation might contribute to an incredible velocity of information sharing and collective concept discovery.

But social media has changed how attention is allocated and what the central unit of discourse is: The personality of creators now often trumps the relevance of the format or the substance of the content. Institutions, such as media companies, have clearly lost their importance. The most successful channels on YouTube are run by people like MrBeast or Marques Brownlee, not by impersonal media brands. Podcasts are typically anchored around individuals, and so are most Substack newsletters. Influencer culture on Instagram and TikTok is highly personality-driven anyway.

You can even see this trend in business: Tweets from celebrity CEOs get much more attention than corporate PR statements or even mainstream media articles. Tech companies and investors increasingly “go direct”, trying to reach an audience without the help of legacy media, and the key element are recognizable personalities with strong opinions.

What does this do to thinking? Social media shifted the nature of the discourse away from the linearity of print and the abstract network of ideas of the web to something that feels a bit more like oral debate — people immediately reacting to utterings from other people.

But the interaction is asymmetrical: Algorithmic feeds decide which creator content we see, and we can react only with emoji buttons or short comments. It’s a very hierarchical form of interaction, superficial in the moment, but maybe with potential for more depth through its iterative nature. And it’s also interesting that we have seen a move back to longer-form content in the last few years, such as long e-mail newsletters and multi-hour podcasts.

The bottom line is: The web and social media have led to a wide variety of content types, interaction patterns and distribution channels. The noise level is incredibly high, but it’s still possible for talented individuals to break through and find a large audience, even with ambitious long-form content. The main losers are traditional editorial media channels. They are clearly losing their gatekeeper function.

And the dominant pattern of thinking and debate is short-form, iterative and highly conversational, but (in most cases) still publicly visible.

The (hopefully temporary) age of AI slop

The rise of generative AI has led to many new effects. Most immediate, and quite aggravating, is the omnipresence of AI slop — AI-generated content of questionably quality. Some studies suggest that the majority of LinkedIn posts is now mostly AI-generated, and so are clearly many comments. Platforms like TikTok, Instagram and even Spotify are flooded by AI-generated content. The number of new smartphone apps has increased dramatically, thanks to AI-based vibe coding.

Much more content, most of it not very good, is trying to capture a finite amount of human attention. And of course AI-based algorithmic feeds help us sift through all the noise. AI now makes (much of) the content, and AI decides what we should pay attention to.

Of course this likely will have the effect that trusted humans and their unique viewpoints will become even more important as a source of quality and authenticity.

But that’s just the most immediate effect of AI, and it might be quit temporary before we find a new equilibrium.

AI: Debates with a machine

Much more importantly, AI enables us to create content for an audience of one: ourselves. I would estimate that currently about 20–30% of text I consume daily was specifically written just for me by an AI tool. I get customized news summaries, meeting prep documents, market briefings, and of course just-in-time responses to all kinds of questions.

I can ask my AI agent questions, ask for more details or different angles, or have it turn the results into other form factors. I can teach it skills that help me be even more efficient with information. It can act proactively or on demand. It is, for many topics, the smartest and most thorough thought partner I could wish for.

Yes, almost all of it is directly or indirectly based on human-generated content, be it in training data or from online sources. This thing I’m interacting with has read most of the Internet, countless books and scientific papers. It can retrieve and analyze additional information in seconds, from both public and my private sources. The AI serves as an editor that picks, prioritizes, summarizes and rewrites information according to my needs, preferences and questions.

Interacting with such an entity is a very new era in human thinking. In many ways, it emulates interactions with other humans in its different forms: I can have a dynamic voice chat with AI, or I can exchange long written statements. I can discuss abstract topics, solve technical problems collaboratively or have very personal conversations. Over time, it learns more about me and can customize its responses.

The AI emulates much of what is great about communicating with a human being, but avoids many of the downsides. The AI is never tired or grumpy; it doesn’t (for the most part) have its own agenda and biases; there’s no ego or emotional backdrop involved. And most of all, it has an incredible wealth of knowledge and the ability to process new information at superhuman speed.

Interestingly, AI models seem to have “personalities” that resonate with different people in different situations. For example, many still mourn the now defunct GPT-4o model, which was particularly empathetic (or sycophantic, say the critics). But for the most part, AI is a neutral canvas for thought, supercharged by knowledge.

When the question is becoming more important than the answer

Many people with programming skills, myself included, had a crisis of identity in early 2026. Suddenly, AI coding agents became so capable that they were better at generating code than most humans. An AI model just won AtCoder World Tour, a global programming competition, beating all human programmers (the elite of the elite of competitive programming) by a huge margin. Many top computer scientists are saying that they have stopped writing code manually. They now just instruct an AI to make software.

Similar things are happening in many other fields of knowledge work. The rate of improvement of these AI tools is breathtaking.

And that leaves us humans with the question: What are we still contributing here? In software creation it’s currently quite clear: We have to decide what the AI should code, and we have to supervise it. Yes, the agent can in principle create any piece of software you can imagine, but it doesn’t have the context and understanding to know what software product would actually be useful and how it will fit with the needs of its human users.

Similarly, my AI agents can now write any market analysis I could wish for, it can draft my blog posts or presentation decks, it can analyze financial statements. It can give me helpful hints about my health, finances and hobbies. But what of this is worth doing in the first place? What does a useful result look like? I still have to tell them.

But this might be a temporary phase. You can already ask the more advanced agent platforms for suggestions about what you could do better in your work, what might be worthy goals to pursue or what might be valuable to build. If you let them, the agents will watch you work and can come up with surprising recommendations. Of course this is still limited to a somewhat tactical level, and broader context is often missing. But it’s probably a question of time until our AIs can help us shape our lives consistently for the better. At least we can hope so.

How can we still share and refine our thinking?

The most consequential change is not merely that AI can answer our questions, suggest actions and provide analysis. It is that a growing part of our thinking probably no longer happens around a shared object.

In the world of print, two people might disagree passionately about a book, a scientific paper or a newspaper article, but at least they were reacting to the same words. The text was fixed, publicly accessible and available for inspection.

With AI, two people can ask the same question and receive different explanations, different supporting sources and very likely different conclusions, depending on the setup and memory of their AI agent. The conversation may remain entirely private. There is no original article that both sides have read and no open chain of argument that an observer can retrace.

AI therefore brings us back to something resembling oral culture, but in a very strange form. It is conversational, fluid and responsive, yet it is not shared. It is based on the accumulated textual knowledge of humanity, but individualized at Internet scale. We are entering an era of private intellectual universes, with potentially much depth and a wide scope, but very few interfaces to the outside world, particularly to other humans.

The upside is enormous. A patient tutor, editor and thought partner can meet every person at exactly the right level of complexity.

But the same personalization can quietly remove the friction from other people that often improves our thinking. A human conversation partner occasionally misunderstands us, disagrees with us or refuses to follow the premise of our question. An AI optimized to be helpful may instead become an increasingly sophisticated mirror, reducing our thinking to our own little bubble, limited by the possibilities of the AI.

Since frontier AI models are trained on similar data and optimized in similar ways, AI is very good at giving us somewhat mediocre, middle-of-the road ideas. Models are trained on best practices and consensus perspectives, but struggle to come up with original viewpoints. They lack the unexpected, slightly chaotic spark often found in strong human thinking. So while it feels that we’re having highly individualized conversations with AI, the underlying substance is standardized. This can be partially alleviated by smart, opinionated prompting, but that’s a conscious act of pushing the AI harder towards original thinking.

That is why asking good questions will matter enormously. But even that may not remain the final human advantage. AI will learn to suggest better questions too. What remains is the responsibility to decide which questions deserve our attention, which answers we should distrust and what goals are worth pursuing in the first place. Judgment cannot be delegated completely without also delegating essential human agency.

Most importantly, we will need to decide when we want to subject our private thinking to the opinionated, biased, messy feedback of other humans. Yes, other people have not read the entire Internet, most don’t have the 130 IQ equivalent intelligence of current frontier models, and they are more driven by their own goals than intellectual purity. But that messy input is often what is needed to achieve the very best outcomes.

Print taught us to think in lines. The web taught us to think in networks of ideas. Social media taught us to think in streams and iterative interactions. AI may teach us to think in constant dialogue — with a partner that is always available, knows more than any individual human and increasingly understands our personal context. The medium is no longer merely carrying the message. It is becoming a participant in the process of thought itself. The promise is extraordinary, and so is the risk if we get locked into our own AI-enabled bubbles and ignore human input.

(Em-dashes are mine)

Originally published at https://innospective.net on July 25, 2026.

Coding agents have revolutionized how software is created. The new default assumption in technology is that when software becomes easier to create, we will get many more applications. You can already see the effect in app stores and on GitHub.

Agentic coding makes it much cheaper to produce internal tools, prototypes, dashboards, and simple websites. Many things that were previously too small, too specific, or too temporary to justify engineering time can now be built by non-technical users. And that’s great.

Do we really need all that software?

But I suspect the framing “easier coding = more apps” misses a much more interesting change.

The real question will not be: how much more traditional software can we create now? It will be: how much traditional software do we still need?

18 months ago I wrote about the future of the software industry. One possible future outcome I explored was the “obsolescence of traditional software”. What was largely speculation back then has now come into increasingly sharp focus.

General-purpose agent platforms like OpenAI’s Codex and Anthrophic’s Claude Cowork are being adopted very rapidly. People aren’t using agent harnesses just for coding, but increasingly for any kind of knowledge work. And the agents they create aren’t traditional software, they are something amorphous, more flexible and more malleable.

There is an older precedent for this. It is the spreadsheet.

Spreadsheets are probably the most successful form of user-created software in history. Most people do not think of them that way, because a spreadsheet does not look like an application. It has no polished interface, no fixed workflow, no engineering process behind it. But in practice, spreadsheets implement an enormous amount of business logic, including many mission-critical workflows. They track sales pipelines, manage inventory, run budgets, plan events, manage hiring processes, report KPIs.

The tech industry tends to forget it, but normal people actually don’t enjoy using and paying for software products. They just want to get work done. Often a structured app is the best way to do that, but quite frequently the flexibility of a spreadsheet is the right (and much cheaper) solution.

Spreadsheets gave users a few powerful primitives: cells, formulas, tables, lookups, charts and integrations. With those building blocks, people could solve a problem without having to turn it into a formal application.

The four capability pillars of AI agents

AI agents may be the same kind of thing, but for a much broader and deeper class of work. And that means that they will frequently be a better solution for a given problem than formal software.

A contemporary AI agent is not just a chat interface. Over the past few months, the architecture of a general-purpose AI agent has crystallized around the underlying foundation models with four important pillars on top:

1. Tool and computer access: The agent can use software the way a person does. It can open files, search, browse, call APIs, edit documents, query databases, send messages, and operate across many systems. In principle, anything that can be done on a computer can become part of the agent’s capabilities.

2. Persistent memory. The agent can remember facts, instructions, decisions and past outputs. In companies, it may also have access to shared context: document stores, CRM records, chat history, meeting notes, code repositories, and so on. This means the agent does not have to start from scratch every time. It can build on what is already there, and it leaves its own work traces.

3. Skills. The agent can use pre-defined procedures that package more than just a prompt. A useful skill may include workflow steps, data connections, acceptance criteria, formatting rules, and escalation rules. This is important because most valuable work is not a single instruction. It is a procedure with judgment points. Skills make agents less like free-form improvisers and more like operators who have been trained on a specific task.

4. Autonomy. The agent can act without being manually steered through every step. It can run on a schedule, react to an incoming event, monitor a condition, or carry a task through until it needs help from a human. The user is delegating work to a system that uses apps on their behalf.

None of these ingredients are entirely new. Traditional software had integrations, settings, automation. It had plugins, macros and scheduled jobs. RPA tools could operate user interfaces. Enterprise applications have always embedded deep business logic.

But the combination is new in depth and generality. Tool access gives the agent rich capabilities in the digital realm. Memory gives it continuity. Skills give it repeatability. Autonomy gives it independence. Together, these create something that is a clear departure from traditional software categories. It is at the same time more flexible and more powerful than conventional software.

A new playing field for computation

Many tasks are are still best solved in traditional apps. CRMs, payroll software, accounting, core banking or medical record management benefit from the characteristics of conventional applications. They need high structure, strong permissions, repeatable interfaces, and clear audit trails.

But a huge amount of work sits outside that zone: Anything that deals with unstructured text, information sources that are spread over several systems and have to be pulled together, the combination of internal and external information, and most importantly workflows that are somewhat unpredictable and need just-in-time judgement and adaptation. These tasks have some structure, but they are too varied and fluid to fit into a normal application.

Historically, these tasks lived in the manually maintained gaps between applications. People copied information from one system into another. They maintained spreadsheets, created templates and used checklists. They relied on memory and asked colleagues. They improvised.

Agents are so revolutionary because they can operate in this complex, foggy space.

They do not require the entire workflow to be formalized upfront. They can work with partial structure and high-level goals, navigate ambiguity and can ask for clarification when needed. Agents can use a spreadsheet, a CRM, an email inbox, a web browser, and a document repository in the same task. They can produce an output, check it, revise it, and remember what changed for next time.

Solving problems instead of producing code

This also suggests a different way to think about the future of software. The naive view is that AI coding expands the long tail of custom applications. Every company can make more internal tools. Every team creates its own micro-apps. Every user vibe-codes their own solution. Some of that will happen. But today’s cool agent-coded application is tomorrow’s maintenance nightmare, especially for rapidly moving requirements.

That’s why I believe that agentic workflows will absorb many things that would otherwise have become applications.

For example, instead of building a fixed analytics dashboard, you may ask an agent to investigate metrics and produce the right view for the current business question. Next month the question might be different, and the agent can just run an adapted process, building on its experience from last time.

Instead of turning every workflow into a SaaS product, you may let agents operate across existing systems with enough memory, permissions, and procedural skills to get the job done.

The boundary will not be whether something is “important”. Spreadsheets already showed that important work often happens in informal software. The boundary will be risk, reliability and regulations.

When a process needs strict consistency, formal governance, high scalability, or a clean and highly predictable customer-facing experience, traditional applications still win. There is no reason to replace a well-designed transactional system with an agent wandering through a UI. That would be a worse architecture.

But when a process is variable, judgment-heavy, context-dependent, or changes faster than an application can be designed, agentic software will be the better fit.

Doing stuff that previously wasn’t possible or too expensive

This could make agents the equivalent of spreadsheets for the AI era, just many times more powerful. A computer turns from something that you use into something that does work for you and in the process helps you do your work better and more efficiently.

This includes many things that are useful, but previously would have been much too complicated or expensive to solve. Heavy users of agents are already experiencing the benefits.

An example from my own work as a VC: I have an agent that reads my calendar, figures out which new startups I’m scheduled to talk to, then pulls together every piece of information about these startups that it can get, and then writes a detailed preparation report with a market overview, competitive differentiation drill-down, and SWOT analysis. As a result, I’m much better prepared when I talk to a startup team for the first time, and the following steps in the process are expedited. It would be prohibitively expensive to do these things with human labor. But the agent both saves time and improves quality, at negligible cost.

I did not have to write any code to set this up, but it wasn’t an immediate success. I needed a few weeks of iterations to get everything right. It was a bit like training a junior employee while at the same time improving the underlying process, and now the agent executes flawlessly, consistently and fully autonomously.

This is a deep shift. Coding agents let us create more formalized software. But agentic patterns may reduce the need to create so much traditional software in the first place. They allow us to solve computation problems in a dynamic, lightweight, yet very powerful way.

What this means for the software industry

The obvious conclusion would be that this is bad news for software companies. I think that is too simple.

Traditional software will not disappear, but it will have to focus. In many areas it may become even more valuable. Systems of record, transactional infrastructure, compliance-heavy workflows, and core enterprise processes still need structure, security and auditability. These are not good places for improvisation and rapid iteration. If anything, agents make the need for trustworthy underlying systems more obvious.

But the role of many software products will change. A large part of SaaS was built around owning the user interface in tight combination with the underlying data model for a particular set of workflows. In an agentic world, that pattern is far less relevant. The agent may become the primary interface, while the SaaS product becomes a reliable data source or business logic repository behind the scenes. Software is increasingly moving from screens to capabilities, from apps to skills, from fixed workflows to adaptive procedures.

That is the biggest implication for software founders and investors: the best opportunities will not necessarily look like “software companies” in the familiar sense. They could look like agents, services, infrastructure, or strange full-stack hybrids.

And there might be a new class of problem solver that thrives in this environment. The industry currently talks a lot about “AI-first builders”: People who use AI to build software products in the traditional sense. There are hackathons with demo days were these products are being shown off. And that’s great. There is a lot of room for a new class of applications that look and feel like products.

But it seems like we’re starting to see a new kind of entrepreneur and startup, one that doesn’t think primarily in productization terms, but focuses on solving customer problems at scale with the help of agents. The currently so popular “forward-deployed engineer” might be a first blueprint for this. “Service as software” is another popular buzzword. But I think these are just first prototypes that show a glimpse of the future.

It’s perfectly normal in technology waves that the dominant pattern of the future emerges slowly and iteratively. As I have written before, the now so familiar SaaS playbook needed well over a decade to be defined.

AI in many ways is a much more profound shift, so we shouldn’t expect to have clarity so early in the process. But it seems like agents are now at a stage where we can at least see them as a clear milestone pattern that points to the future.

Originally published at https://innospective.net on June 6, 2026.

There is a theory we hear a lot these days: AI will make many more people entrepreneurs.

I hope that is true. As a venture investor, I obviously want more people to start companies. More importantly, I think many more people can start companies now. The tools are better, the cost of building is lower, and a single person with good taste, domain knowledge, and persistence can do things today that would have required a small team only a few years ago.

But I am skeptical of the simple version of the argument. AI may make it easier to build a prototype, write code, research a market, run marketing campaigns, or learn useful skills quickly. These are all substantial changes. They lower the cost of starting.

They do not lower the cost of enduring.

Entrepreneurship is not mainly hard because building prototypes is hard. It is hard because it comes with tons of uncertainty. You do not know whether customers care, if the market is real, whether you are too early or too late, or if you are simply wrong. AI can help you move faster, but it cannot tell you with certainty if you’re going into the right direction.

AI is eroding traditional career arcs

For most of recent history, ambitious people were not really choosing careers in isolation. They were choosing arcs. A good career was not just a good job, it was a story about the future. You joined a bank, a consulting firm, a law firm, or a large technology company. You knew roughly what the next ten, twenty, thirty years could look like. You knew what counted as progress and it took to get promoted. You knew what your parents could tell their friends.

At the University of St.Gallen (HSG), where I studied, this kind of career logic is very visible. Most of the smartest students want a path with a clear ladder: consulting, banking, corporate roles, professional services. These are not unambitious choices. Quite the opposite. They are highly competitive paths chosen by highly capable people. But they are also legible. They come with a map. Entrepreneurship has become much more important in recent years, but it’s still the choice of a minority.

MIT, my other alma mater, has a different culture. It is more technical, more experimental, more comfortable with building things. But even at MIT, most graduates do not become founders. Many go into engineering roles, graduate school, research, or large technology companies. Even among very talented people, the desire for a clear arc is strong, often stronger than the desire for independence.

For a long time, that desire was rational. A good career was predictable. You could work hard in your twenties, specialize in your thirties, then manage people and reasonably expect that the institution around you would still understand your value until retirement.

AI is now making that deal less obvious, far less certain.

It is not only that certain tasks will be automated. The deeper issue is that specialization cycles are getting shorter. The thing that made you valuable five years ago may become ordinary very quickly. The skill that looked like a durable foundation may turn into a plugin in an agent platform.

This is disorienting for young people, but it may be even more disorienting for people in their thirties and forties. A twenty-three-year-old can say: maybe I should learn a different stack, join a different kind of company, take more risk. That is hard, but it is an advantage of being young. It is harder when you are thirty-nine, have spent fifteen years becoming excellent at something, have a mortgage, a family, a professional identity, and suddenly realize that the career arc you chose is being disrupted while you are already halfway through it.

This is why the “everyone will become an entrepreneur” story is both attractive and misleading. It is attractive because it seemingly offers a way out. If institutions are less reliable, build your own thing. If ladders are moving, stop climbing and start creating. If AI gives everyone leverage, maybe the best move is to bet on yourself.

There is truth in that. But there is also a romanticism that is, frankly, a bit naive.

It’s cool to be a founder. Is that a good thing?

Entrepreneurship has become very high status. For many young people, “founder” is now one of the most desirable identities. In some circles, it has replaced the banker, consultant, or big tech engineer as the default symbol of ambition. That is a remarkable and positive change.

When I co-founded my first company in 1995, entrepreneurship in Europe was not glamorous. It was exotic, even strange. There was almost no infrastructure, very few role models, and barely any capital. The safe path was much easier to explain. Why would you start a company if you could easily get a job at UBS or McKinsey?

Nowadays, starting a company is culturally legible. There is an entire ecosystem built around the idea that starting a company is an admirable thing to do. I am glad this changed. But the psychological reality of entrepreneurship has not changed nearly as much as the social status around it.

Today’s mature ecosystem suggests that there is a predictable arc to entrepreneurship, almost like a career in consulting. You “ideate”, get into an incubator program, raise some angel money, get first customers, raise a seed round and then series A, B, C, D, until a glamorous exit.

But the daily reality of a founder is not a pitch competition or podcast interview. It is confusing customer conversations, difficult product decisions, hiring mistakes, cash constraints, and long periods where nobody tells you how to continue, except yourself.

Some people thrive in that environment. Others do not. Both are valid preferences.

Entrepreneurial thinking will be crucial everywhere

This is why I think we need to separate two ideas that are often thrown together: becoming a founder and becoming more entrepreneurial.

AI will not make everyone a founder, and it should not. Many people do not want that life. Many people would be unhappy in it. Most people are excellent inside organizations and would create far more value there than by forcing themselves into a startup because the current culture says founders are cool.

But AI may require many more people to become entrepreneurial. That is different.

“Entrepreneurial” does not mean quitting your job to announce a stealth startup. It means developing agency and independent thinking. It means becoming less dependent on someone else giving you permission and defining your next step. It means learning how to identify problems, build solutions, talk to customers, create value.

That can happen inside a company. Or it can happen inside a university, in a side project, an independent consulting business. The distinction is important because many people hear “be entrepreneurial” and think “take existential risk.” But the most useful entrepreneurial behavior often starts with small, reversible risks.

AI allows you to take more risks, in any context

The question is not whether AI makes entrepreneurship easier. It obviously does. The better question is what kind of safety is still available. The old safety came from institutions: the name on your CV, the credential, the promotion path, and the assumption that a system would continue to reward you if you performed well. That form of safety is not disappearing, but it is becoming far less stable.

The new safety may come from something more portable: judgment, adaptability, reputation, customer understanding, and the ability to create value directly with small teams and powerful tools. What are you learning that is effective inside and outside your current institution? What judgment are you building? What visible proof of value are you creating?

AI lowers the cost of small experiments dramatically and allows more people to create something from scratch, inside a company or outside. The only bottleneck will be agency.

I do not think the future of careers is that everyone becomes a founder. The future is that more people will need to behave a little more like entrepreneurs, even when they are not founders. They will need to create solutions independently, keep learning (continuously and rapidly) after their formal training, build visible proof instead of relying only on institutional credentials. And they have to become more comfortable with uncertainty, not because uncertainty is fun, but because certainty is becoming rarer.

When I started a company in 1995, the strange thing was choosing entrepreneurship at all. Today, the strange thing may be believing that any institution can protect you from change. The safest career might be the one that helps you deal with uncertainty, and that’s what entrepreneurial thinking enables you to do.

Originally published at https://innospective.net on May 14, 2026.

Most discussions about AI and work still start from the wrong place. We ask which jobs disappear, which functions get automated, which layers get cut, and whether the entry-level analyst, junior developer, consultant, marketer, or lawyer still has a future. Those are legitimate questions, but they are too defensive. They describe the transition from the perspective of the old organization trying to protect its shape.

The more interesting question is what kind of company becomes possible when intelligence is no longer the scarce input; not a normal company with some AI tools added to its workflows, and not a smaller version of today’s SaaS company with a slightly leaner payroll.

A genuinely AI-first company will probably not look like a traditional corporate pyramid at all. It may look more like a compact institution built around a small number of high-context people, many AI agents, a continuously updating memory system, and a culture that turns talent into leverage much faster than the old career ladder ever could.

This is where the discussion gets more interesting and more optimistic. AI may weaken parts of the traditional apprenticeship model, but it can also create a better one.

The best AI-first companies will not simply remove junior work. They will move ambitious young people closer to real problems, real context, and real responsibility earlier. They will make a new kind of person possible: someone who is not just a specialist or manager, but a high-agency builder who can frame problems, orchestrate machines, verify outputs, understand customers, and compound judgment at a speed that used to be impossible.

The old company was an inefficient ladder

For most of the last 100+ years, knowledge-work companies were built around a fairly stable hierarchy of learning and decision authority. Juniors did structured work: spreadsheets, memos, legal drafts, lead lists, market maps, first-pass code, and all the slightly boring artifacts that keep organizations moving. Middle managers reviewed, coordinated, prioritized, and turned this raw work into something usable through higher context. Senior people defined problems, made judgment calls, managed relationships, allocated resources, and carried accountability.

This system was never as rational and efficient as it looked from the outside. It created bureaucracy, politics, status games, fake urgency, and a huge amount of performative work. But it had one crucial advantage: it trained people.

The career ladder was not just a compensation structure. It was a learning architecture. You became good by touching the raw material of the business over and over again, usually under pressure, usually with blunt feedback, and always by doing work that felt below your eventual ambition. But it was the path to becoming an effective professional.

AI puts pressure on exactly this architecture because many of the old training tasks were also production tasks. If a senior person can now produce a decent market map, legal memo, product prototype, financial model, customer analysis, or investor deck in an afternoon with AI, the old staffing logic starts to erode. Why assign five people to intermediate artifacts if one strong operator with the right tools can produce the same (or maybe better) output in a fraction of the time? That sounds threatening, but the truth is that the old ladder was never the ideal form. It was just the best we could do when intelligence, information processing, and coordination were expensive.

The AI-first company is not a tool stack, but a new shape

The tempting but lazy definition of an AI-first company is “a company where everyone uses AI.” That is not enough. A company does not become remote-first because employees occasionally take Zoom calls from home. It becomes remote-first when documentation, meetings, hiring, trust, and decision-making are redesigned around distributed work (and as we have learned during and after the pandemic, that’s a very tough switch to make).

The same will be true for AI. The AI-first company is not defined by tool adoption, but by organizational shape and structure.

That AI shape will have a few unusual properties. It will be smaller than traditional companies at the same output level, but not merely because it has cut headcount. It will be more fluid and porous across functions because AI lowers the cost of crossing domains.

We’re already starting to see this in the best startups: Product managers who produce features that go into production. Software engineers analyzing customer conversations and deducing product requirements from them. Salespeople running their own market research. Founders not waiting for the data team to provide analytics, but just asking an agent in real time.

The old boundaries between departments will likely not disappear entirely, but they will matter less because the unit of work shifts from “my function” to “the outcome we are trying to produce.”

The real moat will not be the visible product, because that is already much easier to copy. It will be the institution underneath: how the company attracts exceptional people, gives them context, distributes authority, turns customer needs into product, and makes every cycle of work improve the next one.

You can already see some of that in very successful companies: Google and NVIDIA for example have always had unusual cultures and management structures, long before modern AI. Most companies that tried to imitate them failed because there are so many dependencies. That’s what an organizational moat looks like. And AI is an opportunity to build something like this for new companies.

The company memory becomes the operating system

One of the most important differences between old companies and AI-first companies will be the treatment of memory. In almost all organizations today, memory is broken. Some of it lives in Slack or Teams, some in decks and spreadsheets, some in CRM notes, much in people’s heads, and some in rituals that are never written down. This is annoying for humans, but it is a complete blocker for agents.

AI systems become dramatically more useful when the company’s context is explicit, structured, searchable, and connected to decisions. That’s now a fairly obvious insight, and that’s why hundreds of startups and many incumbents are chasing the goal of becoming the context layer for the next generation of companies. But context will not be a software product that you just install. It’s a deep organizational habit.

The AI-first company will treat memory as essential infrastructure. Every customer call, product decision, failed experiment, sales objection, support escalation, hiring lesson, and strategic debate becomes part of a living context graph. Not in a bureaucratic “please document everything” way, but because the work itself produces traces that humans and agents can reuse. The company learns not only what happened, but why it happened, what alternatives were considered, what assumptions were wrong, and what changed after a reality check.

Young people should optimize for access to context and reality, not credentials

All of this matters for young people because it changes apprenticeship. In the old model, you learned by being lucky enough to sit near the right people, overhear the right conversations, and slowly infer how judgment worked.

In the AI-first company, much more of that judgment can become visible. A 23-year-old joining the company should be able to inspect how important decisions were made, how customers actually behave, how product bets evolved, where the company was wrong, and what standards good work needs to meet. That is a more powerful learning environment than being assigned tiny fragments of work with little context.

The obvious advice to young people is to learn AI tools. That is true, but too shallow. Tool fluency will be table stakes. The deeper skill is learning how to own outcomes in an environment where machines can produce most intermediate artifacts. The valuable person is not the one who can generate the longest report or the prettiest deck. It is the one who can ask the right questions, define goals, assemble the right context, use AI to explore possible solutions, iterate on the result, and push it into reality.

This means young people should deliberately seek environments where they are close to customers, decisions, and consequences. Purely internal work that never touches reality will become weaker training. Work near the economically relevant source of truth will become more valuable.

In an AI-first company, a 23-year-old operator might listen to ten customer calls in the morning, ask an internal agent to cluster objections, prototype a workflow with an engineer by lunch, test it with a customer in the afternoon, and write the decision memo that updates the company memory by evening.

If you understand why a customer is upset, why a workflow breaks, why a sales process stalls, or why a product decision changes behavior, you are building judgment. If you are only formatting the output based on your manager’s instructions, you are not getting enough signal. You’re in the wrong place.

The positive version is that ambitious young people may get access to serious work much earlier. In the old model, you often waited years before anyone trusted you with real decisions. In the AI-first era, the bottleneck is not your ability to produce artifacts, but your ability to handle context and accountability. A young person with taste, persistence, technical curiosity, and good judgment can operate with much more leverage than before. The new apprenticeship will be less linear, less credentialed, and probably less comfortable, but it may be much faster.

Founders need to build a culture that makes new paths possible

For founders, the important lesson is not “hire fewer juniors.” That is the unimaginative interpretation. The real question is what kind of person can only become themselves inside your company, no matter if they’re just out of school or have decades of experience.

Great companies have always done this. They create a playing field where the best people can grow quickly and follow their ambitions. The best people do not only choose compensation, category, or job title. They choose the place where they can truly move things, and AI puts this on steroids.

But this is also where founders (and leaders of established organizations) need to be honest. An emotional promise has to become a structural truth.

If you tell people that ownership matters, they need real decision rights. If you say that customer proximity is your moat, customer-facing work cannot be low status (witness the growing importance of Forward Deployed Engineers in the best AI companies). If you say that speed matters, decisions can’t be held up by committees. If you say that talent is essential, average performers can’t set the pace and quality bar. If you say that young people can grow faster here, they need scope, feedback, tools, visibility, and compensation that implement the promise.

This will be one of the biggest cultural differences between mediocre companies and great ones. Mediocre companies will use AI to extract more output from people while keeping the old hierarchy intact. Great companies will use AI to redesign authority. They will give smaller teams more scope, make customer reality more visible, push responsibility downward, and build review loops strong enough that young people can take on more without being set up to fail. That kind of company will feel intense, and sometimes quite chaotic, if you’re used to the old ways of doing things.

The future company may be part startup, part school, part AI OS

The more futuristic version is that the AI-first company becomes a new kind of institution altogether. It is part company, because it sells products or services and must survive economically. It is part school, because the main human work is the acceleration of judgment. It is part operating system, because much of the work is executed by agents. It is part missionary belief system, because exceptional people still need a reason to sacrifice comfort and optionality for a particular higher-level goal.

In reality, great companies have always been containers for ambition. They tell people what kind of work matters, what kind of person is admired, and what kind of future is worth building. AI makes this even more important, because everything visible becomes more copyable. The enduring question becomes: what kind of institution concentrates the right people around the right problem in a way others cannot reproduce?

This also means that “company culture” will stop being a soft afterthought. Culture will become essential infrastructure, maybe the only permanent moat left.

Experience will matter less in some ways and more in others

There is a reason young founders may have a real advantage in this transition. Nobody has decades of experience building AI-first companies. Much of the traditional startup playbook is useful, but some of it is actively dangerous. Static pricing, large functional departments, rigid career ladders, and product roadmaps designed around easily observable user behavior might be a thing of the past.

At the same time, experience is not dead. People problems remain. Trust still matters. Distribution might be more important than ever. Focus, and the deliberate decision of what to focus on, is even more crucial. Customers still need to believe that a company will solve their problem and be around tomorrow.

What changes is the half-life of specific operating knowledge. The best experienced people will be those who can separate durable principles from expired playbooks. The best young people will be those who can move fast without mistaking novelty for wisdom.

That combination may define the strongest AI-first teams: young talent with native fluency in the new tools, paired with experienced people who understand markets, customers, trust, and consequences. But the relationship should not be old-style supervision. It should be more like high-speed joint exploration. The younger person may be closer to the frontier of capability, while the experienced person may be better at knowing what actually matters. The company that combines those two loops well will learn faster than both traditional incumbents and naive AI-native startups.

The optimistic version

The negative version of AI and work is easy to imagine: fewer jobs, thinner middle layers, more pressure, more inequality, more people feeling replaceable. Some of that may happen, especially in companies that treat AI mainly as a cost-cutting tool. But that is not the only possible outcome.

The positive version is much more interesting. AI can make companies less bureaucratic, less trapped by functional silos, and less dependent on slow apprenticeship rituals. It can let young people take on meaningful work earlier. It can make judgment more visible, context more accessible, and execution less constrained by headcount. It can turn a small team into something that has the operational impact of a much larger organization, without inheriting all of the old organization’s coordination overhead.

But this future has to be consciously designed. Young people should not use AI to take shortcuts in their learning process; they should use it to get closer to harder problems, more quickly. Founders should not use AI only to shrink teams; they should use it to build companies where people can grow faster. Companies should not remove the middle of the ladder without replacing the learning function it served.

The company of the future is not a company without people. It is a company where people are less trapped by the old machinery of work. It is a company where the best young employees are not waiting politely for permission to touch real problems, and where leaders understand that culture, context, learning, and a fluid organizational structure with decision power at the edges are essential for success.

The old company made people fit into jobs. The AI-first company will make new kinds of careers possible.

Originally published at https://innospective.net on May 9, 2026.

For the last two decades, the default answer to every business problem was: buy another specialized cloud tool. CRM for sales, helpdesk for support, HRIS for people, BI for reporting. Something for procurement, planning, workflow automation, contract management, and then some integration layer like Zapier to pull the pieces together (hopefully).

Sure, this was a huge improvement over clumsy, expensive, on-prem installed enterprise software. SaaS made software much easier to buy and deploy.

But it also created a strange effect: companies outsourced more and more of their internal structure into dozens or hundreds of point solutions. You adapted to somebody else’s idea of a process, piece by isolated piece.

Every function got its own dashboard, and each of those created its own version of reality. Every workflow generated more handoffs, integrations, meetings, and eventually a team to manage the system. Whole categories, like “the modern data stack”, were created around this complexity. SaaS reduced deployment friction, but made coordination much more complicated and processes more fragmented.

AI is now pushing us in the opposite direction, almost back to a past where software was much more custom and deeply integrated.

If software becomes easier to build and operate, the advantage will shift away from buying the best point solution for every sub-task. It will shift to owning the full workflow again, but not in the old monolithic ERP sense with its over-customized SAP instances.

This future could look more like a modular full-stack company, built from highly customized components.

A lot of SaaS exists because doing the work manually was too labor-intensive and building custom software was too expensive. AI changes that equation. It lowers the cost of automated service delivery, the cost of software creation and the cost of integration.

In many situations, the question of the future is not going to be: should we use a software product with a standalone UI and a bit of AI on top of it? It is: should we use a traditional SaaS product in the first place? Or should we rather use AI to build a system that pulls intelligence and functionality from the best sources, then bundles it in a way that makes the most sense for our company?

This is not the same as “everybody is going to vibe-code their own CRM”. It is more like assembling a solution from the best Lego pieces of functionality, combining them with deep business context, coordinated through highly specific agents.

Of course we’re already seeing this trend in the market. Salesforce just announced a new strategy that lets users access its functionality in a “headless” way, i.e. directly from AI agents through MCP, APIs and CLIs. The CRM is not a full app anymore, but a library of capabilities that can be remixed.

OpenAI and Google just announced new agent platforms, providing powerful capabilities to do exactly this remixing. And OpenAI also started new implementation partnerships with large systems integrators, because somebody has to build these custom solutions. Recombining the whole stack is powerful, but far from trivial.

That opens the door to various forms of re-bundling. As Netscape co-founder Jim Barksdale once said: “In business, there are two ways to make money. You can bundle, or you can unbundle.”

Historically, fortunes were made in tech when one of these unbundling-rebundling cycles happened.

Frequently, the owners of the old bundled solutions were the losers (just ask IBM, AOL or Nokia). This might be why financial markets are punishing SaaS stocks so massively: They are selling a product that is not shaped like the future.

The winners were the providers of the best components (like Intel and Microsoft in the PC cycle) or the ones that provided the best integrated bundles (like Apple with its rich ecosystem).

We are already starting to see some of these patterns: The big four providers of AI platforms (OpenAI, Anthropic, Microsoft, Google) are trying their best to build an integrated stack that allows third parties to plug in their own components. They want to own the bundle, but they need other providers to bring their value to the platform.

Other companies are bundling on a different layer: That of a complete service that provides a defined outcome. This initially looks more like consulting, but it can turn into a complex hybrid business. Palantir is of course the best example for this, and we are seeing similar patterns in legal, finance, industrial solutions and even robotics.

Yet other companies provide the modular building blocks, both in terms of infrastructure (such as AI agent optimization) as well as business functionality, for example rich data APIs for specific problems. It’s quite hard to avoid commoditization at this layer, but if you can find an attractive niche, it can be lucrative.

All of this is uncomfortable for entrepreneurs and investors alike because it breaks familiar categories. Gross margins may be messier. The roadmap and pricing model may be more complicated. The early product may look suspiciously like consulting, or it might look like a narrow feature. You might end up playing in somebody else’s walled garden, with all the downsides of such dependencies.

But there is no way around it: The way software solves problems is changing fundamentally. The stack is being rearranged completely. And there are opportunities everywhere in this new environment.

Maybe SaaS was not the final form of software. Maybe it was just the unbundling-bundling phase before software could become a deeply integrated part of the work again.

Originally published at https://innospective.net on April 25, 2026.

Here we go again: Another fight for platform dominance. This time it’s of course about winning in the AI race.

Anthropic’s launch of Claude Managed Agents this week is the clearest signal yet that the AI market has entered a new phase. This is not just another model release. It is Anthropic moving from selling intelligence to selling execution infrastructure: secure sandboxing, long-running sessions, orchestration, and governance, all tied together with Anthropic’s existing products such as Claude Code.

That is platform behavior, not model-vendor behavior.

The IT industry has always been shaped by large platforms that competed fiercely. Wintel vs. Mac, iOS vs. Android, Playstation vs. Xbox vs. Nintendo Switch, AWS vs. GCP vs. Azure, the list goes on.

Silicon Valley is obsessed with platforms, for a simple reason: The most valuable companies in the world are all platform companies. Winning a platform race can be extremely lucrative, but it’s also not cheap. For example, there’s a reason why Mark Zuckerberg spent dozens of billions hoping to win the VR platform race, so far without much to show for it.

By far the most important platform race is the one in AI, and in particular the race for control of the enterprise AI stack.

From model war to platform war

For the last couple of years, the AI market mostly looked like a model war: Better reasoning, better coding, better multimodality. Better benchmarks. But that was always an incomplete picture.

Platform wars begin when the leading players stop shipping isolated components and start assembling the full stack above them. That is what is happening now. Anthropic’s move into managed runtime is one clear proof point. OpenAI’s Frontier vision is another: it explicitly frames the future as shared context, onboarding, feedback loops, permissions, and AI coworkers operating across real business systems.

The stack is now visible. At the bottom sit the foundation models: LLMs, code, video, image generation. On top of that comes the agent runtime: hosted execution, sandboxes, tool use, sessions, browser or computer access, and long-running jobs.

Above that is context: enterprise search, connectors, memory, permissions-aware retrieval, and company knowledge. Then comes the builder/orchestration layer: workflow builders, reusable skills, agent coordination, evaluation, and deployment.

On top of that sit the application surfaces: coding agents, productivity agents, and finally the general-purpose chat interface that acts as the front door. And sitting across all of it is the layer most people still underestimate: governance — identity, permissions, tracing, audit, security, and control.

Right now there are four serious players competing for the entire stack: OpenAI, Anthropic, Google, and Microsoft. Of course many others offer bits and pieces in the stack, sometimes even leading in a particular area.

But if history is any guide, the really important race is that for dominance of the entire platform stack. And the big four are the only serious contenders for this.

What the big four are offering

Anthropic is probably the company that made the biggest transition in the last few months. Until recently, Anthropic was a company with very strong frontier models, a premium chatbot, and an increasingly strong coding product. With the invention of the MCP integration protocol Anthropic for the first time clearly showed platform ambitions.

Then came massively extended versions of Claude Code as a universal agent platform, Claude Cowork as a productivity tool, and now an agent execution environment. Anthropic still has gaping holes in its lineup, such as lacking models for image and video generation, but it has focused remarkably well on what enterprises really value.

OpenAI on the other hand may still have the clearest articulation of what the end state looks like. Frontier is not pitched as a collection of tools, but as an enterprise platform for building, deploying, and managing AI agents with shared context and explicit permissions and boundaries.

While it has fallen behind Anthropic recently in the coding and agent areas, OpenAI is filling in the stack quickly. Codex is now a serious multi-agent surface for parallel, long-running work, expanding beyond coding. ChatGPT agent combines several operating modes and connectors. OpenAI’s runtime story may be less clearly implemented than Anthropic’s, but strategically it is obvious where this is going. OpenAI is trying to become the operating system for AI coworkers as well, and with a broader model capability footprint than Anthropic.

Of all four players, Google may have the broadest technical surface. On the cloud side, Vertex already offers a real agent runtime environment and many other enterprise components. On top of that, Google is building a serious context and distribution layer, benefitting of course from its strong Google Workspace product. It has an agent and workflow builder and a range of capable (but not yet very popular) coding tools as well. Gemini is of course a strong chatbot product, benefitting from the quality of its models, including the very popular image and video generation tools.

Google’s weakness has historically been a product strategy that can feel quite fragmented. It’s the same in AI: The perceived quality of Google AI products ranges from mindblowing to “what were they thinking when they released that?” And deep integration between the pieces seems to be a challenge as well.

Finally, Microsoft may have the strongest enterprise endgame, even if it is not always obvious. Its advantage is not the model layer, where the company combines some specialized models of its own with the big models from OpenAI and more recently Anthropic. The real strength is the fact that Microsoft already owns so much of the environment where work happens in enterprises: Microsoft Office/365, Teams, GitHub, and so on. Copilot is certainly not perceived in the market as the strongest chatbot product, but it is still ubiquitous in many companies.

Microsoft is also building the runtime and builder layers aggressively with Foundry, Copilot Studio and various Azure products. However, navigating the complexity of Microsoft’s offerings is challenging.

What they are fighting over

The surface-level interpretation is that all of these vendors want a bigger share of wallet from B2B customers. That’s true, but incomplete. What they are really fighting over is being the default environment in which AI work happens.

Being the default platform for work can be very lucrative. Just ask Bill Gates.

In prior platform wars, the winner usually owned the point of highest control and value creation. In the PC era it was the standard stack of CPU and OS, which then allowed Microsoft to dominate productivity apps as well. In mobile it was the operating system and app store. In cloud it was the control plane and service breadth.

We don’t know yet what this “choke point” will be in AI. For a while, it seemed like model capabilities would decide the race, but with several large players now playing on very similar levels, this doesn’t seem to be true anymore. For B2B environments, it might be a combination of runtime, context, and governance, closely integrated with the best models.

OpenAI and Anthropic already train their latest models in close connection with their agent harnesses. This is a clear sign for vertical integration that is not just motivated by business considerations, but also by technical performance.

The winning platform will be the one that makes it easiest to deploy agents that are trusted, connected, and embedded in real workflows. This is also why I think governance deserves to be treated as a first-class platform layer. All four players talk about this, with Microsoft probably having the most experience. Over time, this may become the least glamorous but most important part of the platform war.

What about other players?

Of course it’s not just the big four mentioned here that want to be the owner of a platform that runs your agents. Various other players are in the race as well, but they seem clearly weaker by comparison because they don’t own enough of the stack.

AWS has built strong infrastructure, but has no meaningful footprint in models and UIs. Salesforce is pushing its agent infrastructure as an extension of its core product line-up, but that’s not terribly appealing to non-Salesforce customers. Things look similar for SAP, Oracle and other incumbents.

Then there are of course a bunch of smaller, specialized players that have a best-of-breed product in a particular layer of the stack.

That’s a tough spot to be in. Just look at how Cursor went from everybody’s favorite coding tool to something people barely mention anymore, in a matter of months. Competing against a reasonably capable tool provided by a platform owner is difficult.

Will interfaces be open or closed?

There is one interesting difference from earlier platform wars: The underlying plumbing and integration points in the AI agent stack may become standardized and open even while the platforms compete on a UX, governance and model level.

Historically, platforms tried to control their inner workings closely and define precisely where 3rd parties are allowed to play. Apple has to approve your app for the App Store. Microsoft tells you which exact APIs your Windows application can use. The cloud providers try to lure you into using their bundled services instead of rolling your own from open source components. It’s all about control and customer lock-in.

In AI, we have seen an early proliferation of open standards. MCP allows the connection of external tools and data, and all major players are supporting it. Google introduced its Agent2Agent protocol as an open standard. And it is quite easy already to transfer skills between the platforms. I often develop skills in Codex (most capable coding) and then run them in Claude or OpenClaw. This is surprisingly easy.

But will it stay that way? I doubt it. The big winners in previous platform wars were those that used open standards to win market share, but at the same time captured value with their own proprietary solutions. Sometimes a fully open stack ended up winning (such as LAMP for websites), but that’s rare.

What this means for startups

For startups, the easy era of “just build a nice wrapper on top of the frontier models” or “provide an infrastructure tool that enables cool stuff behind the scenes” is clearly ending. The big platforms have very broad ambitions, and they have to have them, considering all the money that is at stake.

Many startups currently position themselves as what is generally called “platform complementors”: They fill a hole in the line-up of the large platforms where the big players don’t have a solid offering yet: Workflow engines, agent orchestrators, LLM security, memory platforms, and so on. These are all good ideas, and there is a clear market need. That’s a reason why some of these companies can grow very quickly.

But historically, it’s a dangerous position to be in. If you’re under 35 years old or so, you probably have never heard names like Borland, PowerBuilder, Lotus Development, Ashton Tate, or Sybase. These were all very successful software companies in the 80s that filled holes in the line-up of Microsoft and IBM. They largely disappeared when the platform owners encroached on their territory.

Of course you can build successful companies even in a world of large platforms. For example, Snowflake and Databricks became arguably the most important players in the data warehouse space, despite the big cloud providers having competing offerings. If you truly have a best-of-breed product that is wide enough not just to be a feature on somebody else’s roadmap, there is plenty of space.

Another successful strategy is to be the cross-platform vendor on a relevant layer. Oracle is the most famous example: It became the most important database vendor because it was able to run on all large computing platforms.

It’s likely that things will play out in similar ways in AI. The durable opportunities are either above the platforms or between them: vertical agents, proprietary workflow UX, domain-specific context, governance and compliance, specialized orchestration, and multi-platform abstractions. The dangerous place to be is the middle: thin applications or infrastructure features that depend on one platform, while owning none of the broader context themselves.

That is what platform wars do. They compress the value of undifferentiated layers and increase the value of control points.

Startups should study the history of platforms. How you position your company in this new phase of the market will determine whether you succeed or fail.

-

Book tip: “ The Business of Platforms “ by Michael Cusumano, Annabelle Gawer and David B. Yoffie

Originally published at https://innospective.net on April 11, 2026.

We live in the age of the endless video stream — TikTok, Instagram Reels, YouTube Shorts, videos on X. This now dominant form of media consumption has influenced many aspects of society, and it’s changing radically how startups work.

A recent article by Derek Thompson has an apt description for what’s happening to our media diet. Everything is turning into television — or rather its modern form: a continuous stream of short, emotional, interchangeable clips, algorithmically arranged into feeds. The individual piece matters less; what matters is that the feed never stops. The grammar of this world is immediacy, brevity, spectacle and emotion. It doesn’t invite you to think in paragraphs and coherent arguments. It trains you to think in clips and emotionally impactful hooks.

Netflix and other media outlets have already changed their content strategy to accommodate this kind of media diet. Screenwriters have been instructed to write simpler plots that work for viewers with short attention spans who probably are on their phones while simultaneously watching a show.

Founders turn into creators

What happens when this logic spills over from media into the startup world, and especially into AI? More and more, it feels like we are building, funding and evaluating startups as if they were shows, or even just clips in an endless feed of superficial messages.

Founders have always needed to tell a story. The difference is how deep and nuanced this story is. Ten or fifteen years ago, you pitched investors in a room, maybe spoke at a conference, wrote a long blog post. Today, in some sectors, your entire company exists in feeds: customers meet you through short demo videos, talent discovers you on TikTok and YouTube, journalists get their first impression from a launch trailer, investors might scroll past your product long before they see a deck.

The founders who win attention are those who behave like creators. They know how to compress a complex idea into 30 seconds of feed magic. They know how to craft launch videos, how to pick a fight for attention, how to deliver a simple, compelling narrative (“Look, we went from 0 to $50M ARR in just two weeks!”). The line between “startup founder” and “creator of a TikTok channel” is getting blurry.

Demo theatre and benchmark fluff

You can see this very clearly in the current AI wave. There is a whole category of products that are, if we’re honest, mostly “demo engines”: an LLM with a thin wrapper around it, designed first and foremost to look impressive on video. The core asset is not the depth of the product, but the founder’s ability to generate spectacular clips that win on Product Hunt and X. The success metric is not retention or gross margin; it’s how often the demo gets reposted.

The same dynamic shows up in how we judge technology. Television cares about spectacles and scores. AI now has its own versions: dramatic launch videos and cherry-picked benchmark charts. A new model or product appears, accompanied by glossy footage of mind-blowing use cases and a few carefully chosen numbers that place it on top of some leaderboard. Some companies deliberately optimize their product to make it to the top of a chart. What rarely survives the journey into the feed are the boring questions: what are the assumptions behind these evals, how robust is this in real-world conditions, what does inference actually cost at scale, who owns the data, where are the safety rails?

In other words: we get demo theatre and benchmark theatre. The point is not to show how the system behaves in the real world. The point is to deliver a memorable scene with a clear score.

You can feel the pressure this creates on product teams. If your distribution is video-first and your growth depends on staying in other people’s feeds, your roadmap starts to tilt towards features that look good on camera. Meanwhile, the unglamorous work that actually creates defensibility — reliability, proprietary data, robust tooling, deep integration into existing workflows — is hard to film and impossible to explain in 30 seconds. It’s the first thing that gets pushed down the backlog. Short term, the spectacle wins. Long term, the debt piles up.

Thinking, fast and faster

All of this sits on top of something deeper: what this media environment does to our thinking in tech, how it may distort how we reason as founders and investors.

Thompson connects feed logic to declining literacy and eroding deep thinking capabilities, arguing that when everything becomes a stream of short clips, we lose the ability to sit with long-form arguments. Neil Postman warned that a TV-driven culture would start to treat politics, science and news as performance rather than discourse. That sounds uncomfortably close to where parts of the AI conversation are today: continuous drama about “AGI is near” vs. “everything is a toy”, daily outrage cycles (even driven by online spats among billionaires), tribal fights between doomers and accelerationists, breathless coverage of every new model drop.

If you live in that stream all day, your own thinking becomes episodic. It jumps from one extreme to the other in response to whatever is trending, instead of building a stable mental model of where value might actually accrue in five or ten years. That doesn’t just affect journalists or Twitter personalities; it affects founders and investors as well. It’s much easier to react to the latest spectacular demo than to spend an afternoon reading a dense research paper and tracing through what it really implies for your product or your portfolio.

All of this is amplified in AI because the underlying technology really is spectacular. When an LLM or a video model does something surprising, it’s genuinely fascinating — and extremely easy to clip. Sora clips, synthetic influencers, AI agents independently completing little tasks: they are made for television. The risk is that we start to mistake “AI that makes for a good clip” for “AI that makes good businesses.”

TikTok time vs VC time

One place where this is very visible and, frankly, quite worrying is venture capital. VC is structurally a long-duration game: you write a seed cheque in year zero and, if things go well, see real cash returns somewhere around year eight to twelve. Strong exits have become slower and rarer, not faster. On paper, we can mark things up along the way (and LPs reward that), but actual DPI still arrives on decade-long timescales.

Now overlay that with a market psychology trained by TikTok. Attention cycles compress into weeks, sometimes days. The half-life of excitement around a company is often one launch video or one funding announcement. Even large AI rounds feel like content objects in the feed more than milestones on a long operational journey. Everyone moves on to the next “hot” model, agent, or device before the previous one has even had a chance to prove whether it works outside controlled conditions.

That creates a fundamental tension: how do you run a ten-year strategy in a market that mentally resets every ten days? One obvious answer is that VCs start behaving like momentum traders. They surf narrative waves, invest into companies that are good “marks” for the next round, and they hope to offload risk to a later-stage investor or a strategic buyer before anyone looks too closely at retention, unit economics, or real moats.

Celebrity-driven investing

Maybe as a consequence, fundraising feels increasingly detached from tangible technical substance or revenue traction. Celebrity is becoming the product. If a luminary like Mira Murati or Ilya Sutskever starts a company, capital in the billions lines up before the product exists, because the “IP” is the person and their narrative.

You can see similar effects with younger, far less accomplished founders who have managed to get their 15 seconds of fame. That’s rational to some extent — these people can attract talent, capital and distribution — but in a feed-logic market the pendulum often swings too far. Founder-audience fit starts to overshadow founder-market fit.

Sure, many of these companies generate real revenue, often growing at unprecedented rates. In today’s AI market, hype helps winning paying customers because everybody wants to try out the latest cool toy. But once you look behind the curtain at unit economics and churn rates, the picture gets a lot darker.

We will see what the returns on these hype-driven investments are going to look like in the long run. If history is any guide, there’s not a lot of reason for optimism, apart from — as usual — a few outliers.

Every wave and hype cycle has its stars that deliver outsized returns, but only very, very few make it big. Sometimes fame correlates with financial success, but much more frequently it doesn’t. Yes, we all know Elon Musk, Mark Zuckerberg and Jeff Bezos. But do you know Kim Polese, Louis Borders or Joe Kraus? They were big media stars in the dot-com boom, running some of the sexiest startups. But their companies failed or delivered mediocre outcomes, and they declined into obscurity again. And their investors for the most part lost a lot of money.

This echoes patterns that we have seen historically in the creator industry. For every Mister Beast or Charli D’Amelio, there are countless once-famous creators that have lost their audience (remember Michelle Phan? Smosh? Ryan Higa?). Fame is very ephemeral, particularly in the age of the short-form video feed. Startups that rely on the sugar high that comes from feed fame might very quickly find themselves on the wrong side of the next hype wave.

Finding balance in the age of spectacle

So what do we do with this? I don’t think we can or should roll back the clock. Video is a powerful way to communicate complex ideas (I should know; my last own startup focused on video for a reason). Founders with strong public personas can be a real asset to their companies. Attention is not fake, it’s a real input into distribution and hiring.

But if everything is quietly becoming a TikTok feed, we need to be much more explicit about when we are optimizing for the show and when we’re building the substance behind it. We need to notice when we’re funding narrative arcs rather than unit economics. We need to catch ourselves when we’re judging companies by how they look in the feed instead of how they perform in the dark, when nobody is watching.

In ten years, cap tables won’t remember who had the best launch trailer. They’ll remember who quietly shipped boring infrastructure that never broke.

Good TV and good companies are not mutually exclusive. But they are not the same thing.

Originally published at https://innospective.net on November 16, 2025.

The AI industry currently produces new startups in something that is probably best described as a Cambrian explosion. We have seen this frantic pace of company creation before in other technology waves, such as the dot-com boom of the late 90s and the SaaS boom of the early 2010s.

Nobody knows what the AI startups of the future will look like exactly and what will them make successful. What will their business model be? Which technical architecture decisions will be crucial? How will they go to market?

Is there something we can learn from previous technology waves to help us think about success recipes? Can we maybe even see early glimpses of what a new playbook for success might look like?

The Software-as-a-Service wave might be a good place to look for historical echoes. It’s hard to remember now, but software was built and sold very differently before SaaS became mainstream ca. 2010.

What we did and didn’t know about SaaS in 2007

I had the pleasure of writing my master’s thesis at MIT about the then emerging SaaS market in 2007. My advisor was Michael Cusumano, a professor who had already done research about the software industry for decades (and has continued to do so). 2007 was a moment quite comparable to where we stand in AI now: It was clear that there is going to be a big wave, first successful companies were growing very quickly, but there were many things still to be figured out. My old thesis probably describes quite well what the consensus in the industry was back then, and in particular how many open questions there were.

For example, here are some things that we now take for granted that were unclear in 2007:

• Where do I host my SaaS applications? AWS had just started out, Azure and GCP didn’t exist. The assumption was that SaaS startups needed to own and manage their own servers and that doing so efficiently might be a strategic advantage.
• How do you create a UI efficiently? Modern standards like React didn’t exist, and it wasn’t even clear that the dynamic HTML UIs of today were the right way forward. Some people believed in hosted backends with locally installed clients.
• How do you integrate SaaS with existing systems? Doing so in 2007 was clumsy and manual. The elegant API methods of today didn’t exist yet.
• How do you find customers? In 2007 software sales were still dominated by field salespeople in suits who flew to clients to take them to expensive dinners. Product-led growth, free trials and self-serve onboarding were only being tested in tiny niches.
• How do you price your software? Seat-based pricing already existed, but many traditional software products were priced on a per-CPU basis and other now obsolete schemes. Usage-based pricing — now increasingly popular in SaaS — was not practical for locally installed software. On the other hand, many people in 2007 believed that there would be such a thing as advertising-financed software, but that hasn’t happened outside of gaming.
• How often do you release a new product version? For locally installed enterprise software, doing this even quarterly was ambitious. Now SaaS companies often push new updates several times a day — unimaginable in 2007.
• How do you measure your business, and which KPIs matter? A vague concept of LTV and CAC existed, but things like net revenue retention, rule of 40 or magic number had yet to be invented. And it wasn’t clear what a reasonable gross margin for a SaaS company should be (I guessed 85%, but that was too optimistic. Median margins for public SaaS are currently 76%).

This shows that while general assumptions (SaaS is going to be big and will disrupt many incumbents) were obvious, many of the crucial details had yet to be figured out. I’m certain it’s the same with our current AI wave.

Market expansion out of nowhere

Even with all its optimism, the tech industry sometimes falls into the same biases when it comes to predicting market growth.

In the case of SaaS, there was (as so often) a bit too much optimism about the short term. It turned out that enterprises needed a bit more than just a few years to wrap their heads around this new model. But the longer-term analyst predictions were too pessimistic. Many thought that SaaS would top out below 50% share of the total software market, and of course we’re now way beyond that.

When I wrote my thesis in early 2007, I found a grand total of 108 companies that offered some kind of SaaS product. Almost all of them provided horizontal solutions such as CRM or project management.

Now there are (according to most estimates) about 30'000 SaaS companies in the market. And presumably most of them offer very focused vertical solutions for certain use cases in certain industries.

This huge market expansion in vertical SaaS was something that was still hard to see in 2007. For traditional installed software, selling to small verticals was much too expensive. You can’t send field salespeople to customers with small budgets. SaaS with product-led growth, lean tech stacks and quick feature iteration cycles was ideal to crack these smaller market segments.

Great disruption or great times for incumbents?

A frequent discussion in every tech wave is what it will mean for the existing incumbents in the industry. Historically, predictions of total annihilation (“AI will kill the entire SaaS industry!”) tend to be very wrong. Incumbents are not idiots, and they see tech waves as much as everybody else. The question is how quickly they can and want to adapt while new players are emerging. Sometimes they emerge as the big winners, sometimes they get disrupted. The truth is often in the middle.

In the case of SaaS, the transition was certainly not easy for incumbents such as Microsoft, Adobe, SAP or Oracle. Interestingly, even as early as 2007 they all experimented with some kind of SaaS product. But who wants to give up a comfortable business with 95% gross margins for an experimental new approach?

By now this is of course history. A subscription-based business model is now the standard, and hosting software in the cloud is too, apart from some conservative corners of enterprise IT. The big incumbents have all transitioned to the new world, and in some cases spectacularly well.

However, the story was less glamorous in the segment behind the biggest names. Many mid-market vendors in the ERP and CRM spaces saw their market share shrink due to competition from SaaS. The lesson is that big incumbents with ample resources are in a better position than smaller players to deal with a disruption.

What are the lessons for AI startups?

The takeaway from the history of SaaS is that many macro trends are pretty obvious early on, but the details of implementation are not. Who could have predicted that the majority of SaaS companies would host their products on the IT infrastructure of an online book store? No doubt, we’re going to see similarly crazy-sounding things in the future of AI. Market expansion is not obvious either and often goes beyond what people are able to imagine in the early days.

There are a few things we probably can say about the future of AI startups:

1. AI products will look different from SaaS, but we don’t know how. Right now most AI products still are structured like SaaS. If history is any guide, this will be a very temporary pattern. For example, it doesn’t make much sense for a powerful autonomous AI agent to have a login page, dashboard and settings page as the primary units of interaction. It’s hard to predict where this is going. The “dominant design” (to use James Utterback’s term) typically takes years to emerge. But it’s likely that we will interact with AI products in many more granular ways.
2. Software (if that’s still what we will call it) will be able to solve many small problems. Very much like SaaS enabled niche solutions for even small verticals, AI will make it possible to provide solutions for tiny problems. Of course we’re already experiencing this when we use today’s chatbots. The countless small problems we solve on ChatGPT are not something that would be worth creating a software product for. But where is the boundary between one-off solutions (like what we would have done in a spreadsheet in past times) and repeating problems that would warrant creating something that feels more like a product? Hard to tell, but the boundary will shift towards much more granularity.
3. Moats will change towards speed, vertical expertise and depth of integration. In traditional pre-SaaS enterprise software, moats for software companies were much about customization, long deployment cycles, sales relationships and expensive product creation. This changed with SaaS to agility, network effects, ease of use, and flexible pricing. We’ll see probably more of this with AI, but potentially taken to the extreme.
   Let’s say I need an agent that solves a particular workflow problem for me. Why would I want to evaluate software products? Why wait until tomorrow? Can somebody just solve the problem for me in an intelligent way, right now, based on a vague description? The AI companies that will be able to do that will build on deep experience with a certain class of use case in certain verticals paired with infrastructure that is optimized for speed. This will be hard to copy.
4. Many of today’s infrastructure concerns will go away. In 2007 startups worried about which server type and database to buy, today they worry about foundation model pricing. This kind of broadly used infrastructure often gets commoditized quite quickly, and we’re already seeing signs of it in AI foundation models. Infrastructure worries will move to a higher level. It’s a good idea to plan for this.
5. There will be new forms of distribution. Most AI startups still follow the good old SaaS GTM playbook (PLG, inside sales, etc.), but it’s very likely that the future will find other ways to take products to the market. Maybe agents will be bought and sold on marketplaces, with buyer agents automatically finding the best solution. In the enterprise, we’re currently seeing a return to a consulting-heavy model with Forward Deployed Engineers. Is this due to market immaturity, or will it be the way of the future? Experimentation with fresh routes to the market could pay off hugely for innovators.
6. Pricing will evolve to something new. The standard model now is something like $20 per seat per month for many AI products. That’s of course just a leftover from SaaS pricing, not reflecting true product value at all. Most likely we’ll see some kind of seat+usage hybrid pricing for a while, but future schemes could look very different. Value-based pricing for example has been quite elusive for SaaS, but maybe AI will enable it at scale.
7. Dealing with regulation will be table stakes. Early SaaS companies didn’t have to worry about regulation and certifications. Later we got GDPR, SOC2 and many other great inventions of the regulatory-industrial complex. It’s likely that regulatory pressure will further increase with AI. Dealing with this responsibly, but also efficiently, will be a crucial skill for AI startups.
8. The dance with big platforms will continue to be both an enabler and mortal threat. Most software companies have to build on somebody else’s platform. That’s just efficient for technical enablement and distribution. In the good old days it was the Windows OS or an Oracle database. In SaaS it’s hyperscalers, API providers and marketing channels such as Google. In mobile it’s Apple and Google. In AI it’s foundation model providers.
   What these have in common: They are elephants that don’t particularly care about all the small players that build on their platform. If the elephant feels like releasing a new feature that will kill off dozens of its small customers, it will do it. If the elephant decides to change rules that will critically affect the business of thousands of its customers, it will do it. Balancing these threats with the great opportunities that platforms provide has always been a crucial skill. There are startup concepts that are not very smart (like filling a temporary hole in a big platform’s functionality) and there are others that have historically worked well (like being the cross-platform glue for people who want to use multiple platforms).
   It’s the same in AI. Of course countless startups are one OpenAI or Google announcement away from obsolescence. Others are not. Studying these platform dynamics is one of the most useful things that founders can do.

AI is at a moment that is quite similar to the early days of the SaaS boom. Everyone sees a tidal shift, but the money is in the details that haven’t solidified yet. Just as SaaS rewired many dimensions of the software industry(hosting, UI, integrations, distribution, pricing, release cadence, plus a surprise boom in vertical niches), AI will change many dimensions. But we should be just as confident that our current assumptions about the “how”-the infrastructure we rely on, the way we price our products, the channels we use to find customers-will look profoundly naive in retrospect. Founders who move fastest on these learning loops will be the winners.

(Thanks to GPT-5 Pro and Gemini 2.5 Pro for research and inputs, DeepL Write for copy editing and GPT-5 Thinking for the visualization. I’m sure thanking AIs for their contribution will pay off some day…)

Originally published at https://innospective.net on August 9, 2025.

Unpopular opinion: While everyone talks about AI’s ability to efficiently produce cognitive deliverables — such as writing code, creating content, and analyzing data — a massive opportunity is hiding in plain sight: coordination.

Some studies suggest that up to 50%-60% of employees’ time in large corporations is spent coordinating rather than creating value. Tackling this problem could be AI’s most transformative impact on the economy, fundamentally reshaping how organizations operate.

There are three important dimensions of coordination inefficiencies in which AI might have an impact:

1. The Context Window Problem

Humans, like AI models, have limited processing capacity — relatively short context windows, if we want to apply this AI term on humans. We can process only a certain amount of information and can only understand facts and connections up to a certain degree of complexity.

Organizations compensate by breaking down complexity into individual goals — essentially creating personal “reward functions” for each employee. Much like a reinforcement-learning AI model is being trained to optimize against a certain goal, human workers get goals that are specific enough to handle for one person. The same applies to teams, departments, and divisions on a higher level.

What’s the problem with that? People optimize their own metrics without seeing the bigger picture. These personal reward functions don’t necessarily add up to an optimal organization-wide reward function.

Procurement departments are an example. Anyone who’s sold to enterprise clients knows this pain. While procurement teams serve vital cost-saving and quality control functions, they often operate in silos. They’ll spend months negotiating a 10% discount on a strategically important software product or service, while the delay could cost the company millions in missed opportunities. They’re optimizing their reward function (cost savings) without considering the organization-wide impact.

In the near future, AI agents might be able to collaborate toward a unified organizational objective. They could process vastly more context and adhere to more complex priority structures than any individual human. Companies have long used computer models to optimize complex goal hierarchies, but integrating goal optimization with actual work execution and dynamically adjusting priorities based on current reality seems much more promising.

2. The Social Factor

Humans are inherently social beings. We seek approval, status, and connection — goals that don’t always align with corporate objectives. Office politics, relationship management, and status optimization consume significant mental bandwidth and decision-making capacity in many organizations. I have personally had the privilege of working in very socially inspiring teams with great cultures, but have also experienced quite the opposite. And even the most pleasant work environment suffers from social friction at times.

Of course social dynamics between humans can lead to magical moments of innovation or culture building that AI can’t replace. But it’s probably fair to say that this upside comes often at a considerable price. Consider how many meetings exist primarily to manage social dynamics rather than make decisions.

AI doesn’t need to be liked, doesn’t compete for promotions, and doesn’t get distracted by interpersonal dynamics. This neutrality allows for purely objective optimization toward defined goals.

The interesting thing to watch will be how autonomous AI agents work alongside human co-workers. Will social dynamics in these “cyborg” teams be very different from human-only teams?

3. The Principal-Agent Problem

Every employee is, understandably, optimizing for their entire life, not just their current job role. Career advancement, work-life balance, skill development — these personal priorities naturally influence professional decisions, potentially creating misalignment with company objectives.

For example, software engineers sometimes choose technologies or architectures that enhance their personal marketability rather than serve the company’s needs. Using the latest trendy framework might make an engineer more hirable elsewhere, even if a boring, proven technology would better serve the current project. It is often very difficult for engineering managers to balance these needs.

AI agents don’t have careers to manage or personal lives to balance. They can dedicate 100% of their processing power to organizational outcomes.

This will certainly be a source of conflict. We’re already hearing senior software engineers complain that AI coding tools don’t produce “beautiful” code with all the fancy frameworks and abstractions engineers tend to be proud of. But what if AI is simply more efficient, ignoring secondary aspects that matter only to humans?

What will an AI-Human-Hybrid Company Look Like?

The companies of the future — maybe the near future — will likely be structured as dynamic networks of AI agents coordinated by strategic human leaders who define the overarching organizational reward function.

AI might become the organization’s operational backbone, autonomously handling routine tasks, data analysis, real-time decision-making, most sales and marketing tasks, and internal communication. Human leaders focus exclusively on strategic vision, innovation, critical decisions, and building meaningful stakeholder relationships.

The efficiency gains would be staggering. No more endless email chains, misaligned incentives, or coordination overhead consuming the majority of productive time. The coordination revolution may prove even more transformative than AI’s impressive cognitive capabilities.

How far away is this future? It’s hard to say. AI is amazingly good at certain cognitive tasks, but the challenge is still to feed it with all necessary context, to give it the right guardrails and to make sure it sets the right priorities. The necessary technical capabilities won’t be ready overnight, but they are progressing every week.

Most of all, the adoption of AI as a coordination layer will be very transformative for established organizations. Fundamental transformations take time and tend to be messy.

I therefore think that we will see this kind of AI-first organization initially in relatively new companies — startups and scale-ups — that don’t have to deal with a lot of legacy. First signs are already here. And it’s possible that these AI-first companies are so much more efficient than incumbents that they are going to have a relatively easy time winning in their markets.

The Social Impact

To state the obvious: Yes, this will have a massive societal impact. We are already seeing that many companies are becoming much more efficient (i.e. are employing fewer humans) thanks to AI, and we’ve probably only scratched the surface. AI has not even really started to solve the coordination problems described in this post, and the impact on employment could be massive. Who needs middle managers if AI can do their jobs much better?

And consider the impact on human identity and self-image: It’s one thing to use AI as a coding buddy or copywriting assistant. But what happens to human identity when AI starts making decisions? What if it gives you orders in your daily work because it’s better at determining what’s good for the company? What if AI becomes your manager? There is little doubt that there will be job functions where humans are going to be superior to AI for quite a while, but if AI is better at coordination and management, we could see inverted roles.

If you’re the CEO of an AI-first company, how do you know that the AI is presenting correct and unmanipulated data for your own decisions? How do you supervise AI agents that operate at incredible speed?

As so often, the emergence of a fundamental innovation raises more questions than it answers.

(Thanks to Claude Code, o3-pro and GPT-4.5 for their suggestions, as well as DeepL Write for copy editing and proofreading. And yes, em-dashes are all mine.)

Originally published at https://innospective.net on August 2, 2025.

What’s the future of productivity: Pre-programmed, AI-enabled workflows or fully autonomous, general-purpose AI agents?

There are currently two schools of thought in the AI industry when it comes to the near future of AI agents: One perspective thinks that AI works best when it’s combined with more traditional code and ways to access data. Think of it as “ AI agents on rails “: the AI can make certain decisions, but the degrees of freedom are strongly limited by workflow code that guides the AI. This enables better consistency and deeper integration with traditional IT mechanisms, such as existing rich APIs to connect legacy systems.

Many providers of workflow tools have started to embrace this concept, and with a bit of “vibe coding” in code generator tools such as Cursor it’s quite easy to roll your own, fully custom AI workflow. In a way, the underlying AI models are really a tool used by traditional code.

The other concept thinks that AI agents will be entirely autonomous and much more flexible. In other words, the AI is going to be the dominant force and interacts with its environment. The agent gets prompted by the user and then figures out independently what exactly to do. Agents come up with a plan, execute each step in the way they see fit and can use tools to access other systems.

For this last integration part, the new MCP (Model Context Protocol) framework is currently evolving as the new standard, bridging the AI agent world and traditional APIs. Importantly, the user or creator of the agent doesn’t have to know what a particular API can do. The agent figures that out independently.

Confusingly, both approaches are referred to as “AI agents” in the current market, but they are quite different. So which approach works best for which use cases?

Fully autonomous agents have made a lot of progress just in the last few weeks. New tools such as OpenAI’s o3 model or the Manus AI agent can plan and execute remarkably complex tasks. Several vendors (e.g. OpenAI, Google, Grok, You.com, Perplexity, Manus) now offer “Deep Research” capabilities that conduct extensive research jobs independently, googling and assembling information for you. And experimental browser use agents can even use the web for you to solve problems.

But can these agents run more complex workflows that go deeper and have to be consistent? As a test, I tried to implement several of my own vibe-coded AI workflows with general-purpose agents. The results were quite mixed.

One of my workflow agents that can research and draft a PowerPoint presentation about any topic is now topped in most cases by the flexibility of o3 and Manus AI. Both agents can create a presentation end-to-end with sometimes surprisingly good results. These broader tasks are probably best covered by general-purpose AI. Particularly when the input data can take on many forms (such as web search results), the flexibility of a general-purpose agent is a huge plus. It can not only decide what to best do with the data, but also much more dynamically find a way forward to get the best results.

But in other use cases the pre-programmed workflows still prevailed:

• Finding relevant industry events based on precise criteria
• Summarizing e-mail newsletters
• Conducting an in-depth competitor analysis
• Running semantic search on a large table

In all these cases the general-purpose tools showed some promise, but ultimately failed.

A big problem was the scope and precision of processing. For example, Claude with an MCP connector to Gmail was able to summarize newsletters in my inbox very nicely, but it stopped every time after newsletter no. 10 or so, while the pre-programmed workflow can process any number of inputs. Somehow the agent seems to have a limit in the length of its processing run. Similarly, the competitor analysis always stopped at a small number of relevant companies (often less than a third of what the workflow AI was finding), and the agents struggled to find correct numerical data, such as fundraising information. In some cases, the agents ended up hallucinating wildly.

A huge concern was the lack of reliability. A semantic search over a table of company profile data with OpenAI’s o3 and Google Gemini 2.5 delivered plausible and very well summarized results, but unfortunately they were different each time I ran the query, despite using the exact same prompt and data. This is not likely to instill much trust in users.

Similarly, accessing external systems with MCP is still quite a bit of a mess. Apart from a complicated setup process and the usual questions around security (do you really want to have a fully autonomous AI access your CRM with read/write privileges?) the results tended to be quite inconsistent. This approach is clearly promising, but will need to mature quite a bit.

Bottom line: Both the pre-programmed as well as the general-purpose approach have their justification in today’s market. For broad, fairly unpredictable tasks where exact precision is not crucial, general-purpose agents such as o3 or Manus can already be tremendously useful. But for anything that needs more reliability and depth, a healthy dose of good old deterministic code will do the job.

To use an analogy: Pre-programmed workflows are a bit like SaaS. They bring structure and deep capabilities, but also limitations. General-purpose agents are like a spreadsheet: Wide open to do pretty much anything, but a lot can go wrong if the user doesn’t know what they’re doing (and is not able or willing to check results). Obviously both approaches can and will co-exist, but you have to choose wisely to get the right tools for each job.

However, as always with AI, the really interesting question is not where we are today, but where we are going. Judging from the recent progress of general-purpose AI agents I wouldn’t bet against this category. The improvements in terms of planning scope, precision of instruction following and also reduction in hallucinations has already been tremendous. MCP and other tool integration approaches are still flimsy, but it’s clear that the industry needs these things and is investing massively, which sooner or later will lead to mature platforms. I wouldn’t be surprised if general-purpose agents would take over many of today’s structured workflows in the next couple of years.

Originally published at https://innospective.net on April 26, 2025.