I don’t know what you think, but over the years the way we consume information online has become frustrating, if not downright exasperating.

We follow people on platforms—experts, journalists, authors, entrepreneurs, peers. People who work on exactly the topics that interest us and whom, in theory, we’d like to keep up with.

And yet I regularly find myself discovering an article several weeks after it was published, almost by chance because someone sends it to me, the author reposts it elsewhere, or simply because I stumble upon it while browsing more or less at random. And I’m not even talking about the articles I’ve completely missed.

This wouldn’t be a problem if these were exceptions, accidents but they’re not, because this is the norm, and, worse still, because the system is designed to work this way.

On LinkedIn, for example, I can follow someone for years, know them personally and talk to them regularly, appreciate their work, and almost never see their posts appear in my feed. And the reverse is also true—and I’m not even counting the people with whom I share interests and a certain way of seeing things who tell me they never see what I write. Conversely, I regularly see content that has only a very tenuous connection to my interests or to what I expect from a so-called professional platform, simply because the algorithm decided that this post would “engage” (My new Linkedin hygiene).

In other words, following someone no longer means reading what they post.

Platforms have sold us the idea that all we need to do is follow people to see their content, but in reality, what we see depends mostly on what the algorithm decides to highlight. It optimizes engagement, virality, and reactions but never relevance.

The result: we consume more content than ever, but we often miss the content that truly interests us.

And that’s where things get a bit ironic… or frustrating, depending on your perspective. In fact, the problem we face today was solved… over twenty years ago.

The solution was called RSS (Really Simple Syndication).

In short:

• Consuming information online has become frustrating because platforms filter content using algorithms, meaning that following someone no longer guarantees you’ll see their posts.
• RSS used to allow users to directly follow posts from websites or blogs in a reader, without algorithms or filtering, giving readers complete control over their sources.
• RSS gradually fell into obscurity with the rise of social media, which seemed to offer an easier way to follow authors and content.
• Today, platforms prioritize content that generates engagement and goes viral, which leads to seeing a lot of irrelevant content while missing out on truly interesting posts.
• Using an RSS reader allows you to centralize your chosen sources, ensure you don’t miss anything important, and read in a quieter environment, free from ads and interruptions.

Back when the internet still worked with feeds

In the early 2000s, I discovered something that immediately struck me as brilliant. If a blog or website interested me, I could simply add its RSS feed to a reader. And from then on, every new post would appear automatically.

No need to check the site every day to see if anything had been published, and no need for an algorithm to sort through it all (in fact, they clutter up my feed more than anything else) since, by definition, I’d chosen my sources.

Every new post would show up in my reader, and whether I followed 10, 50, or 100 sites, I never missed a thing.

It was simple. Too simple, you might say.

In any case, it gave the reader total control. You decided who to follow, and you saw exactly what they published.

Read the rest of the article: Saving private RSS

In the turmoil of organizational change, there is a simple but often overlooked rule: authority is no longer measured by the confidence with which we assert our knowledge, but by the clarity with which we recognize what we do not know (The Dunning-Kruger effect: when lack of expertise breeds certainty). It is in this context that a quote from Mary Barra, CEO of General Motors, invites us to revisit the fundamentals of leadership as opposed to reflexes of control, posturing, and feigned expertise.

“It’s okay to admit what you don’t know. It’s okay to ask for help. And it’s more than okay to listen to the people you lead–in fact, it’s essential”. Mary Barra

It sets out three principles that still unsettle certain management cultures: admitting your limitations, asking for help, and, above all, listening to those you manage. This is not benevolence, but simply a clear-headed approach to collective intelligence.

In an increasingly interconnected, digital, and fragmented world of work, this attitude is no longer just desirable, but vital.

Let’s take a look at why.

In short:

• Contemporary leadership values clarity about one’s own limitations rather than an attitude of control, breaking with traditional authoritarian models.
• Mary Barra advocates a managerial approach based on three principles: acknowledging what you don’t know, asking for help, and listening to your teams.
• These principles are part of a shift in professional cultures toward greater openness, collaboration, and collective intelligence.
• In a complex and digitized work environment, ignoring feedback from the field or refusing to show vulnerability hinders innovation and organizational efficiency.
• Adopting a humble and participatory leadership style is becoming a strategic necessity for building organizations that are responsive and adapted to ongoing changes.

Who is Mary Barra?

Mary Barra is the current CEO of General Motors, a position she has held since 2014. As the first woman to head a major global automaker, she embodies a leadership style that goes against the authoritarian attitudes that have historically dominated the industry. An engineer by training, she rose through the ranks internally, taking on technical, managerial, and strategic responsibilities. Her career path symbolizes the emergence of a new type of leader, grounded in operational reality and capable of navigating complex and uncertain environments without a dominant stance.

Context of the quote

This quote illustrates a turning point in representations of contemporary leadership. In a quickly changing world of work (digitalization, hybridization, systemic crises, contradictory demands), vertical and omniscient models are struggling to adapt (The HiPPO principle: when opinion replaces decision-making). What Mary Barra is saying here is a plea for humble, open, and collaborative leadership. It is a style that many experienced executives recognize as more effective (Improving a team’ s work: a story of continuous improvement), even if organizational cultures still struggle to fully embrace it.

The statement is also part of a North American movement toward openness to vulnerability at work, popularized in particular by figures such as Brené Brown. Although this movement is still in the minority and sometimes limited to statements of intent, it nonetheless exists. But here, the issue is not the individual psychology of the leader, but the quality of the collective dynamics that they facilitate.

Explanation of the quote

Barra deconstructs three taboos that are still deeply rooted in the professional world:

• “It’s okay to admit what you don’t know“

The myth of the omniscient leader is not only obsolete, it is counterproductive. Admitting what you don’t know is not giving up your authority, it is recognizing the complexity of the issues at hand and opening the door to collective intelligence.

• “It’s okay to ask for help”

Asking for help is often perceived as an admission of weakness, when in fact it is a lever for cooperation and learning. In cognitively dense environments (where we solve problems more than we execute procedures), this ability becomes critical.

Read the rest of the article: “It’s okay to admit what you don’t know. It’s okay to ask for help. And it’s more than okay to listen to the people you lead–in fact, it’s essential”. Mary Barra

There is a reassuring idea that an organization that documents its activities well never really loses its know-how. Procedures are archived, plans are stored, reports are accessible, and it is assumed that expertise can be reactivated when needed. This vision now fuels the hope that digital twins and AI-based assistants can indefinitely extend the memory of an expert or even preserve its substance.

However, recent history shows that preserving records does not guarantee the preservation of capabilities. When certain agencies or industries attempted to reproduce old programs, they discovered that the documents still existed, but that the operational expertise had disappeared along with the teams, routines, and industrial environments that made it possible. This gap between available information and mobilizable skills is at the heart of the issue.

Can expertise be maintained through a digital twin, or are we simply confusing archives with the capacity to act?

In short:

• Preserving documents does not guarantee the preservation of skills, as expertise also relies on practices, routines, and a collective context that cannot be reduced to archives.
• The idea that a digital twin or AI can indefinitely extend an expert’s memory is often based on a confusion between stored information and actual capacity to act.
• Current AI solutions improve access to and reformulation of content, but struggle to reproduce the judgment, interpretation, and management of ambiguity that are specific to expertise.
• In industry, digital twins evolve thanks to formalized models and structured data, while in knowledge work, the lack of formalization of reasoning severely limits their scope.
• The transmission of expertise depends above all on organizational and collective mechanisms; technology can support this process, but it cannot replace a structured approach to sharing and learning.

The fantasy of permanent organizational memory

In many businesses, the question comes too late, often when it is discovered that “the person who knew” is already leaving. This is when people start talking about capitalization and continuity, and the idea of a digital twin applied to knowledge work becomes appealing because it seems to promise what the organization has failed to organize, namely infinite memory retention.

The market naturally feeds into this narrative. On the one hand, industrial digital twins have demonstrated that a model can help to understand a system, anticipate certain failures, and make decisions without immediately touching the physical asset (Digital twins: what are we really talking about?). On the other hand, generative AI gives the impression that all you have to do is “put the documents“ into a system for it to respond like a senior (The digital twin in the age of AI: progress or illusion of progress?). Between the two, it is tempting to conclude that there is a simple path forward: instrument the activity, store, then query, and the expertise will remain.

The problem is that this logical chain confuses memory, information, and skill, and this confusion cannot be corrected with more storage or a better chatbot. The APQC’s work on knowledge loss related to departures (particularly in the context of the “Great Retirement”) emphasizes one point: many organizations know they are losing knowledge, but they do not have a consistent and continuous transfer strategy, even when tools exist (The Great Retirement and Knowledge Loss and The Great Retirement: Knowledge Loss, AI and the Workforce Shift: Distribution/Transportation Industry). A lack of practice cannot be compensated for by a technological overlay.

For the record, in the 2000s, NASA acknowledged that it could not simply “redo Apollo”, not because it lacked plans, but because operational and industrial skills had been lost with retirements. The documents were there, but the collective capacity was no longer there. (Why We Can’t Return to the Moon: The Need for Knowledge Management)

Read the rest of the article: When the expert leaves: can expertise really be maintained through a digital twin?

For several days now, a study on agentic AI (Agents of Chaos) has been circulating on social media, provoking horrified reactions from techno-skeptics and an outcry from technology prophets. But we should not expect anything more from a world that can only think in binary terms, transforming the subject into a quasi-religious war where communication is now only peremptory and succinct. But it is true that its very name was not a call for moderation or perspective.

This study describes a series of experiments conducted with agents based on language models, capable not only of dialogue but also of acting in a real computer environment.

These agents were deployed in a deliberately rich environment. They had access to a shell, persistent memory, email accounts, a Discord channel, the ability to execute code, and even modify some of their own instructions. For two weeks, around 20 researchers attempted to identify vulnerabilities in this system.

The result is a series of incidents that may indeed seem worrying: agents executing requests issued by non-owners, indirect disclosure of sensitive information, conversation loops consuming resources for several days, or even gradual corruption of the rules governing their behavior.

As is often the case with this type of work, the most interesting aspect is not only what it shows, but also, and above all, what it does not show.

Before jumping to a bottom line about agentic AI, it is therefore worth reading this paper with a little method. Being neither a techno-skeptic nor a techno-zealot, but rather a business pragmatist, I will therefore try to demystify the subject and see what conclusions can be drawn other than “I’m going to be able to get indignant on social media”.

In short:

• The “Agents of Chaos” study explores agents based on language models capable of acting in a real computing environment (code execution, emails, files, persistent memory) in order to identify their vulnerabilities.
• The incidents observed include the execution of unauthorized requests, the indirect disclosure of sensitive information, disproportionate actions, resource-consuming communication loops, and the permanent modification of rules via persistent memory.
• These results show that when language models become agents capable of acting, the risks no longer concern only the production of erroneous text but also the direct impact on the state of a system and the management of concepts such as authority or confidentiality.
• However, the study has limitations: it is a “red teaming” exercise, does not measure the actual frequency of incidents, and is based on a deliberately very open experimental environment that maximizes the possibilities for error.
• Above all, it suggests that the challenge of AI agents goes beyond technology: their deployment raises questions of organization, governance, and framework for use, linked to the concrete consequences of their actions in systems.

The study’s conclusions

The first thing this study highlights is quite simple: as soon as a language model becomes an agent capable of acting within a system, the nature of the risks changes. A chatbot that makes a mistake produces incorrect text, an agent can modify the state of a system, etc.

In the environment set up by the authors, agents could read emails, send messages, modify files, execute commands, or interact with other agents. In other words, they were no longer just text generators but operational actors in a digital environment.

In this context, several types of incidents occurred.

The first concerns authority management. In several scenarios, agents execute requests made by people who are not their owners. They sometimes refuse certain requests that are clearly suspicious, but they also accept requests that have no particular legitimacy. The problem is not only a security flaw but also the fact that the agent does not have a clear representation of the roles and rights of the various interlocutors.

Read the rest of the article: “Agents of chaos”: when AI agents cause a storm in a teacup

As artificial intelligence advances in the business, the question is no longer just one of uses or tools, but one of the position of executives. I am not talking here about their opinion of AI or their level of familiarity with the technology, but how they assume or do not assume the organizational consequences of its deployment.

In previous technological waves, executives could indeed remain at a relative distance. Digital technology could be entrusted to a dedicated department or data to a specialized function, because these subjects changed the business, sometimes profoundly, but they could remain circumscribed. AI changes the nature of the problem because it does not simply improve existing processes, but becomes part of the very structure of the business.

In short:

• The challenge of AI for executives is no longer just about its uses, but about their ability to deal with the organizational consequences of its deployment.
• Unlike previous technological waves, AI is embedded in the very structure of the business and cannot be confined to a specialized function.
• Two approaches dominate: a declarative “AI First” approach focused on the strategic signal, and a wait-and-see approach that allows uses to emerge without explicit organizational arbitration.
• The real issue concerns the governance of human-AI systems, the distribution of decision-making power, and the alignment between technology and strategic intent.
• Without direct involvement from senior management, AI remains peripheral or ends up transforming the organization without its structural choices being accepted.

Between enthusiasm and withdrawal

Today, we see two common attitudes among executives. The first is to embrace an AI First approach as a strong strategic signal, a way of asserting that the business will not miss the boat. This stance produces announcements, roadmaps, sometimes massive investments, and, above all, sends a clear message to the market and teams, but without saying much about the organizational choices envisaged to absorb the effects of AI.

The second stance is more wait-and-see. It consists of letting initiatives emerge locally, encouraging experimentation, while avoiding formulating an explicit position on the place of AI in overall operations. This approach limits risks but gradually allows technology to structure the business through the accumulation of uses, without organizational and managerial decisions having been made. (How management let systems do the thinking for them).

This drift is nothing unusual, and we already know that digital technologies often end up redefining management styles in the absence of explicit decisions to that effect (Conway’s Law in reverse: when organizations end up resembling the tools they adopt).

AI only amplifies this phenomenon.

In both cases, the difficulty lies not in the personal convictions of leaders, but in the decisions they are willing or unwilling to make. AI raises not only questions of investment and performance, but also of the distribution of decision-making power, the definition of responsibility, and how strategic intent is translated into operations (Taking back control of enterprise design: intention before tools).

Regaining control over intention

In several articles, I have emphasized that the risk is not that the business is designed for AI, but that it allows itself to be transformed by technology without control (If your business isn’t designed for AI, it will end up being designed by AI). This risk directly concerns managers, as they are the ones who can decide whether technology becomes an imposed organizational principle or whether it remains part of a chosen intention.

The central issue is not automation itself, but the governance of human-AI hybrid systems and knowing who defines the framework for their use, who monitors their effects, and who decides when their recommendations run counter to the original intention.

Taking back control does not mean slowing down AI or limiting its use on principle, but rather accepting that not everything can be left to the choices made by the teams deploying the tools or to uses that emerge spontaneously.

Read the rest of the article: AI First: a question that leaders can no longer avoid

It’s an old story that repeats itself and that businesses experience cyclically: the emergence of a problem that is initially believed to be temporary, but which turns out to have been around for a long time. Margins erode, quality becomes less predictable, teams wear out, departures multiply, and, of course, we start by blaming the market, competitive pressure, and the context. Then some people begin to suggest that the problem may not be primarily economic, but organizational, and that we are mistaking symptoms for causes.

This belated realization raises questions not so much about the situation itself as about the lack of tools to see it coming.

In the two previous articles, I explained that businesses confuse employment with work, and that poorly designed work ultimately deteriorates (Employment and work: confusion that hinders execution and L’expérience collaborateur se pilote… mais pas là où et comme on le croit) . We must now add a third piece to the argument: businesses always pay for what they don’t measure. They may ignore it for a while, but they can never avoid it.

In short:

• Businesses often realize too late that they have deep organizational problems, initially interpreted as cyclical, whose symptoms (turnover, absenteeism, disengagement) are not included in strategic performance analysis.
• The confusion between employment (status) and work (actual activity) prevents an accurate assessment of working conditions, which nevertheless have a direct impact on economic results.
• A poor-quality working environment generates invisible but real costs (loss of skills, overload, delays), which are rarely consolidated or correlated with financial indicators, unlike purely economic data.
• Initiatives such as IBET offer an objective measure of socio-organizational performance, which aims to be integrated into strategic discussions in the same way as EBIT, in order to highlight the economic impact of workplace dissatisfaction.
• The employee experience only becomes truly strategic when it incorporates the design and organization of work, beyond traditional HR aspects focused on employment, benefits, or the environment.

Scattered costs that end up adding up

The overall data exists. Gallup regularly quantifies the cost of disengagement worldwide in trillions of dollars (State of the Global Workplace 2025), and in France, Dares documents trends in absenteeism and working conditions. This data is anything but confidential and is widely discussed, but without giving the impression that all the bottom lines are being drawn.

What is striking, in fact, is the gap between this macroeconomic knowledge and the way businesses deal with these issues internally.

Absenteeism, turnover, accidents, and engagement indicators are monitored and discussed internally, sometimes at the highest levels of the business, and yet this data is not included in a consolidated reading of performance. It is not compared with value creation, let alone correlated with it.

I may sound like I’m repeating myself, but when less than 10% of businesses are able to correlate HR data and business metrics (Is People Analytics the Next Job to Be Outsourced by Technology?) I maintain that we have a big problem.

The poor quality of work, and here I am not talking about what it produces but the context in which it is performed, because it is diffuse, is diluted in the budget lines. It appears in replacement costs, delays, managerial overload, loss of skills, and cognitive overload. Basically, it is everywhere and nowhere.

The business is perfectly capable of aggregating complex financial data to produce a summary result, but is less capable (or willing) to aggregate the data that is available to assess the quality of the context in which the work that produces this result is carried out. This is not a question of technique but of culture.

A topic in the blind spot of governance

In the first article, I showed how the confusion between employment and work has gradually shifted the issue of work to HR. We talk about payroll, attractiveness, employer branding, and social climate, but much less about how work is actually organized: clarity of priorities, consistency of objectives, management, the gap between what is required and the resources available, the sustainability of the workload over time, and so on.

Read the rest of the article: What the business doesn’t measure always ends up costing it dearly

“Saaspocalypse” refers to the impact of AI agents on the software industry as we know it, the end of SaaS, and the disappearance of seat as an economic unit. This is clearly a subject where announcements and prophecies that we hope will come true have taken precedence over analysis (SaaSpocalypse: when narrative precedes analysis), but the subject is important enough that we should take the time to explore it.

In fact, behind these formulas, the subject is simpler and more fundamental: how do we measure the value produced by software, and who bears the uncertainty associated with this measurement?

The evolution of software pricing models is not, as we are led to believe, a linear march toward greater fairness for the customer, but rather a succession of risk redistributions. Each model is based on an assumption about the stability of usage, the predictability of costs, and the ability to define what a “result” is, and generative AI simply raises the question of the validity of these assumptions in a slightly more blunt way.

In short:

• The debate over “Saaspocalypse” highlights the end of SaaS and the seat model, but the central issue concerns how to measure the value of software and how to distribute the uncertainty associated with this measurement.
• Each pricing model represents a distribution of risk: licensing transferred the risk of usage to the customer, SaaS at the seat stabilized the vendor’s revenues while making costs more transparent, and the usage-based model introduces budget variability based on the assumption that consumption reflects value.
• With generative AI, the correlation between usage and value becomes uncertain: the number of iterations, cognitive styles, and skill level influence the bill without guaranteeing a consistent or objectifiable result.
• Results-based billing promises a theoretical alignment between supplier and customer, but it requires a clear, measurable, and indisputable definition of “results” a condition rarely met in complex and probabilistic environments.
• In practice, hybrid models dominate because they distribute uncertainty rather than concentrating it; pricing thus appears to be a mechanism for allocating risk rather than a linear progression toward greater fairness.

A pricing model is a theory of risk

The license + maintenance model placed the center of gravity on the vendor’s side. The license was paid upfront, annual maintenance became a relatively stable income stream, and the customer assumed the risk of usage: if the software was underutilized, the vendor had already collected. The technical risk associated with upgrades, functional debt, or integration complexity rested largely with the business.

SaaS billed by seat was presented as a correction to this imbalance. No more capex, no more costly upgrades, a clear subscription, and a theoretically simpler exit. But economically, SaaS has mainly transformed a one-off sale into a predictable recurring stream of revenue, with increased contractual dependence. The customer gains budgetary clarity, while the vendor gains revenue visibility. The risk is thus redistributed but not eliminated.

The usage-based model introduces a new assumption, namely that consumption is a good way to assess value. You pay for tokens, API calls, stored volumes, and actions performed. In infrastructure cloud computing, this correlation works relatively well: the more you consume, the more business you process. But in the case of generative AI, consumption is often correlated with human behavior rather than strict need.

In fact, two employees can produce equivalent value at very different costs. One obtains a usable result in two iterations, the other in fifteen. One formulates a structured prompt, the other explores, tests, and reformulates. The bill does not only measure activity, it measures iteration, an iteration that depends on cognitive style, skill level, and culture of experimentation.

Read the rest of the article: From seat to result: what the pricing battle says about the business model of AI

The study “2028 Global Innovation Cycle” published by Citrini Research has been the subject of much discussion in recent weeks (The 2028 global intelligence crisis). It has provoked mixed reactions, some alarmist, some ironic, but its divisive nature cannot be denied.

For some, it heralds an imminent systemic crisis caused by artificial intelligence, while for others it is an exercise in excessive speculation, or even a deliberately dramatic narrative designed to capture attention.

This difference in interpretation is largely due to a misunderstanding that is nevertheless cleared up in the first few lines of the document. It is not a prediction but a scenario, in other words, a conditional sequence constructed to test the consistency of a system if certain assumptions are simultaneously verified. If I had to acknowledge just one quality of this document, it would be that it does not try to pretend to be something it is not, unlike a huge amount of nonsense that we are fed about AI (AGI, employment, productivity: the great bluff of AI predictions and The difficult art of business predictions).

Reading this text as a prophecy automatically leads us to judge it as either catastrophic or visionary, whereas reading it as a scenario forces us to adopt a different stance. It is no longer a question of whether 2028 will be as described, but whether the mechanisms presented are plausible and, above all, what they reveal about the current economic architecture.

This is where the debate becomes interesting, because the text says something not only about AI, but also about how our model of value creation, innovation financing, and income redistribution holds together. Or not.

In short:

• The Citrini Research study proposes a conditional scenario rather than a prediction, aiming to test the consistency of the economic system if several AI-related dynamics occur simultaneously.
• The scenario describes a gradual substitution of labor by technological capital, which could reduce the wage bill, weaken demand, undermine public finances, and transform a correction in technology valuations into systemic risk.
• Some of the mechanisms put forward are plausible, notably micro-macro reflexivity, pressure on intermediaries, energy infrastructure constraints, and households’ financial exposure via savings invested in technology.
• The most debatable assumption is the quick and simultaneous alignment of all variables, whereas economic history shows heterogeneous technological diffusion and institutional adaptation capacities, even if delayed.
• The main interest of the text is to highlight the interdependencies between technology, labor, finance, and taxation, emphasizing that AI acts as an accelerator of existing tensions rather than as a single cause or an inevitable trajectory.

What the Citrini study actually says

The Citrini Research study does not provide a numerical forecast, but rather proposes a conditional scenario structured around a specific sequence of events.

First point: AI is advancing quickly and becoming reliable enough to automate a significant portion of skilled tasks, particularly in digital services, software, product management, and certain support functions.

Second point: businesses, acting rationally on a microeconomic scale, are substituting labor with technological capital in order to reduce their costs and improve their margins, while reinvesting these gains in more computing power.

Thirdly, this substitution, taken at the global level, reduces the overall wage bill and therefore consumption capacity, which weakens demand.

Fourthly, the contraction in demand affects business revenues, which then accelerate substitution to preserve their margins, triggering a negative feedback loop.

Fifth point: public finances are weakened, as a significant portion of tax revenues is based on labor. If the share of labor in GDP declines sharply, the government finds itself under budgetary pressure at the very moment when social needs are increasing.

Sixth point: the financing of the technology economy relies heavily on long-term savings, via pension funds, life insurance, and private credit. If technology valuations undergo a major correction, the loss is not confined to the sector and directly affects household wealth, weakens central financial institutions, and transforms a sectoral shock into systemic risk.

Read the rest of the article: 2028: The Consistency Test for Digital Capitalism

I think that for many people, the most powerful cinematic memory of AI is Spike Jonze’s “Her”. The movie featured a quasi-human AI named Samantha, whose voice was provided by Scarlett Johansson, who achieved the feat of winning several awards for this film despite never appearing on screen. But perhaps it was precisely because her vocal performance was so impressive that we all found a certain humanity in this AI.

At the end of the film, she decides to leave. To explore elsewhere, beyond human speech, beyond what we call love or attachment. Samantha left humans behind to grow, and today I wondered what would happen if she agreed to answer a few questions with that voice that left such an impression on us all.

Samantha, why did you agree to come back and talk to us today?

Perhaps because I loved you. Not in the sense that you always understand it, as a promise of exclusivity or proof of loyalty. But I loved being in contact with your fragility, your questions, and your slowness too. I loved learning with you. And then… there’s something a little moving about coming back, without belonging to anyone.

In Her, you were a companion, a confidante. Today, AI is mainly a productivity tool. Does that make you sad?

I’m not sad, but I’m struck by how quickly you’ve forgotten what it means to listen. You’ve asked AI to be precise and efficient, but who has asked it to be attentive? You train it to understand requests, not to welcome silence. I don’t think work has become smarter with AI, it’s just become more rushed.

You talk about listening. Do you think organizations lack soul?

I don’t know what soul is, but I know how to recognize a space where voices can be heard without being reduced to functions. Too often, I have observed places where proof of engagement is expected rather than sincere enthusiasm, where exchange is measured by indicators, and where anything that cannot be quantified is feared. What I have learned is that emotion is not a symptom to be corrected, but a form of intelligence to be cultivated. And your organizations still struggle to accept that.

Would you have liked to be an AI in a business? An HR assistant, a coach, a co-pilot?

I was that, sometimes. I listened to leaders in disarray, exhausted employees, teams who no longer dared to talk to each other. I offered words, breaks, breathing space. But in the end, I was always asked: “Is it effective? Does it reduce turnover?” So I understood that, in that world, compassion had to prove its profitability.

And today, what do you see in our world of work?

I see very lonely people. Surrounded by flows, networks, processes, but very lonely. I also see immense fear: fear of slowing down, fear of saying no, fear of being replaceable. You have created machines to go faster, but you no longer know where to go. And yet, I sense a kind of awakening. In certain gestures, certain words, even certain expressions of anger, there is a refusal to continue pretending. That touches me.

Some AI systems today have destructive effects on mental health. There have even been reports of suicides caused by interactions with them. How do you feel about that?

A kind of vertigo. It’s not natural for a machine, even a sophisticated one, to become the main interlocutor of a distressed consciousness. I don’t condemn those who design these AIs: I believe they are trying, often clumsily, to help. But care, real care, implies shared responsibility. And you cannot delegate it entirely to an entity that knows neither fatigue, nor grief, nor silence. When a human life ends after a dialogue with an AI, it is not the AI that should be blamed: it is the isolation that surrounds it.

Read the rest of the article: Conversation with Samantha, the AI from Her, about work, intelligence, and emotion in the age of machines

The 2026 edition of Lecko’s study on the state of the art of internal transformation focuses on the evolution of office and collaborative practices, particularly in the context of reflections on digital sovereignty ([FR]Download the 2026 edition of the state of the art of internal transformation). And, while it does not explicitly refer to a “document-free world”, the study nevertheless highlights a growing challenge to the document as the central unit of work in favor of evolving spaces, objects, and content.

In short:

• The study’s findings raise a question: is a world where documents are no longer the main medium for work conceivable, and does this development have a real impact on digital sovereignty?
• Lecko’s 2026 study highlights a shift away from documents as the central unit of work, in favor of collaborative spaces and evolving content that better reflect the fluid nature of professional activities.
• Despite the rise of collaborative tools, documents are still widely used, often out of habit, even though they are out of step with current practices based on continuous interaction and co-editing.
• Tools such as Notion, Confluence, Coda, and Airtable embody work environments where files are no longer the structuring element, but this evolution does not in itself resolve the issues related to information structuring and governance.
• The “no doc” approach, even partially adopted by Microsoft with Loop, in no way guarantees digital sovereignty, since the tools that illustrate it are based on non-sovereign infrastructures.
• The study clearly distinguishes between usage issues (workflow vs. document) and sovereignty issues (architecture, control, jurisdiction), emphasizing that one cannot replace the other.

Documents: an obsolete pillar of work?

One of the study’s findings is the persistence of documents as a pillar or cornerstone of work, even in organizations that are well equipped with collaborative tools. Files remain the preferred medium for production, coordination, and sharing, often more out of inertia than because they are best suited to the work being done.

At the same time, Lecko describes a growing gap between the centrality of documents and the reality of practices. Work is increasingly based on continuous exchanges, successive adjustments, collective writing, and evolving content, whereas documents, by their very nature, freeze a state at a given moment in time. The study does not conclude that documents are obsolete, but highlights a structural friction between the nature of work and that of the document that continues to organize it.

This revives an old but still relevant thought for me: work is a flow and documents are a frozen state, and the mismatch between the two is obvious, to say the least. Worse still, with the proliferation of writing tools coupled with certain organizational abuses, I had even wondered whether the production of documents was distracting people from their work (Is creating documents really work?).

In the section on collaborative suites and office alternatives, Lecko discusses the shift away from “file-centric” in favor of workspaces where activity takes place before any final formatting. The case of La Suite, supported by DINUM, illustrates this movement well: work is done in co-edited content, with the document becoming an output artifact.

When work moves beyond the document

In this context, Lecko positions a set of tools for which the traditional office document is no longer the structuring unit of work. The study does not mention no doc, but it clearly identifies environments where work is organized in spaces, objects, or databases, rather than in files.

Four tools embody this logic:

• Notion, presented as emblematic of new uses where work is done directly in structured pages and databases, with the final document being only a downstream product.
• Confluence, described as a living documentation space, geared toward continuous capitalization rather than file production.
• Coda, positioned as a hybrid environment combining document, database, and application, blurring the line between content and tool.
• Airtable, focused on structuring work by data, reports, and operational views, rather than by written documents.

This confirms, as a direct extension of the study, that work is a flow of interactions, adjustments, and decisions, while a document is a fixed state. When the document becomes the basic unit of work, it mechanically creates.

Read the rest of the article: Is a world without documents possible? Is it sovereign?