• AI can kill the individual DNA of an organization by cleaving to the generic standard, undermining the organizations it seeks to streamline
• organizations become more automated, yet less adaptive; more data-driven, yet less wise; more efficient, yet less legitimate in the eyes of employees and customers.

• People stop thinking

   

  AI systems create a powerful illusion. Their outputs are fluent, confident, and often plausible. That very fluency encourages cognitive offloading; people stop thinking as deeply as they did before because the machine appears competent.

• What leaders can do: Identify critical human capabilities.

   

  Before you introduce AI into any core workflow, be explicit about which human capabilities your organization cannot afford to lose for its own competitiveness. Certain skills do not develop automatically. They include judgment under uncertainty, systems thinking, ethical escalation, and interpretive reasoning, i.e., the ability to view opportunities and challenges through the lens of a chosen strategy. These skills are developed and maintained only through use.

• Rules get buried in systems

   

  In every organization, there are moral choices enshrined as rules—who gets which resources, who is evaluated as successful, whose claims are accepted or denied. Traditionally, these rules have been visible, contestable, and subject to debate. This is how businesses maintain legitimacy and adapt to changing circumstances.
• The essential aspect of these decisions is that they operate through an organizational structure designed to surface conflicts, enable deliberation, and clarify accountability
• AI changes this organizational dynamic by embedding these often-subjective decisions inside technical systems. What once required explicit deliberation happens automatically, often at scale and speed.

• What leaders can do: Protect the structures that produce expertise.

   

  Many of your most valuable capabilities are created through organizational knowledge embedded in established hierarchies, not individual performance. Mapping these dependencies explicitly and documenting how knowledge moves from one level to the next is a fundamental step in assessing whether AI tools are homogenizing complex decision-making processes.

   

• Social ties are weakened

   

  Businesses are not just systems of rules and workflows; they are also social structures built on trust, mutual understanding, and shared purpose. These qualities of an organization’s culture emerge from human interactions which are sometimes cooperative and sometimes combative. AI threatens this fabric in two ways.
• it displaces interactions that once required people to work through problems together
• when decisions appear to come from opaque algorithms, traditional sources of authority—CEOs, managers, professionals—lose credibility.

• What leaders can do: Rebuild the communities that AI thins out.

   

  Your organization’s capabilities persist because people develop shared understanding through sustained interaction. AI often replaces those interactions with individualized, machine-mediated workflows. Don’t let that happen by default.
• Brightview’s solution focused on rebuilding strategy muscle via visible expertise. It established a policy banning AI-generated content from client-facing presentations. Teams must create original strategic narratives explaining their thinking. The agency also restructured client relationships so that every account has a designated “strategic lead”—a senior person whose job it is to challenge AI outputs and advocate for creative approaches the algorithms might miss.
• AI can enhance your organization’s performance—but only if you recognize what it cannot do.
• . Those remain irreducibly human functions. If you allow AI to crowd out these capabilities, you risk your organization becoming technically advanced but competitively fragile.

13 more annotations...

• À chaque fois, la pression d’adaptation était réelle, mais à chaque fois, les entreprises disposaient d’un certain délai pour absorber le choc, expérimenter, former ses équipes et ajuster ses processus.
• le rythme d’évolution et la profondeur des implications ne laissent plus le temps de la réflexion
• Avec l’avènement de l’IA agentique, il ne va très certainement rester que deux positionnements viables : proposer l’offre la moins chère ou la meilleure.
• ne doit pas être considérée comme un nouvel outil informatique, mais comme un nouveau paradigme de l’outil informatique
• les entreprises ont toujours tendance à sous-estimer l’ampleur des changements à venir et à surestimer leur propre capacité d’adaptation. E
• t la dette numérique. Ce terme désigne le retard accumulé par les entreprises et organisations dans leur adaptation au numérique
• les véritables transformations ne portent pas que sur les outils ou données, mais aussi sur les processus et les mentalités
• Trois décennies de grands bouleversements des usages qui ont forcé les entreprises et organisations à revoir leurs outils, canaux et modes de fonctionnement.
• ce ne sont pas des tâches mécaniques ou répétitives qui sont déléguées aux IA, mais des tâches cognitives, analytiques, voire décisionnelles.
• vous ne pouvez pas déployer des agents sur un chaos informationnel
• Agentisation des opérations (de la production). C’est l’étape qui est normalement déjà en cours, celle qui concerne les tâches répétitives : traitement d’informations et de données, génération de contenus, gestion des requêtes courantes
• Agentisation du savoir-faire (de la conception). C’est l’étape émergente, celle qui commence à toucher les métiers créatifs, analytiques et d’expertise.
• Agentisation du pilotage (de la planification et des décisions). C’est l’étape à venir, celle qui n’est pas encore là, mais qui se dessine clairement. La capacité des agents à synthétiser des masses de données, à simuler des scénarios, à formuler des recommandations stratégiques
• Il n’y a ainsi aucun intérêt à vouloir gérer une activité au XXIe siècle avec des outils et méthodes du XXe siècle
• un tel chantier ne peut se faire sans l’adhésion des managers et des salariés.
• s’adapter même sans avoir une vision claire de ce à quoi pourrait ressembler le quotidien d’une entreprise une fois sa transformation agentique opérée.
• L’orchestration, c’est ce nouveau rôle qui émerge pour les travailleurs du savoir : non plus exécuter des tâches eux-mêmes, mais piloter des agents pour qu’ils le fassent à leur place (définir des objectifs, évaluer les résultats, corriger les dérives, garantir la qualité et la pertinence des actions effectuées…).
• 16,3 % des emplois privés et publics en France pourraient être menacés par l’IA agentique
• La révolution des cols blancs est en marche, car elle a déjà été initiée avec l’adoption rapide des chatbots.
• s bouleversements en cours sont trop importants et trop rapides pour que les réponses traditionnelles soient à la hauteur
• les dispositifs institutionnels n’ont pas été conçus pour absorber des chocs de cette rapidité et magnitude :
• les IA prendront les places que l’on laissera vacantes, alors à nous de ne pas nous effacer devant elles
• ’adoption structurelle de l’IA est particulièrement faible dans les PME/TPE : respectivement 31% et 15%. Pourtant, le taux d’adoption par les cadres ou dirigeants eux-mêmes est supérieur : 35%.
• l’IA générative est pour le moment perçue de façon erronée et utilisée en surface
• les retardataires, une fois convaincus, ont l’avantage de ne pas avoir à servir de cobayes, ils bénéficient directement d’outils optimisés et de meilleures pratiques, sautant ainsi des étapes d’exploration / expérimentations.
• si le chantier d’agentisation semble être une haute marche à franchir, il s’inscrit dans une continuité
• La première priorité, c’est la structuration de vos informations, données et connaissances
• La deuxième priorité, c’est la modélisation de vos processus.
• La troisième priorité, c’est la montée en compétences de vos équipes
• La quatrième priorité, c’est la gouvernance
• Les IA prendront les places que l’on laissera vacantes. Cette formule implique une responsabilité active : ne pas laisser de vides là où votre présence, votre jugement et votre expertise humaine font réellement la différence.
• L’histoire de la transformation numérique nous enseigne que les entreprises et organisations qui s’en sortent le mieux ne sont ni celles qui adoptent le plus vite, ni celles qui résistent le plus longtemps, mais celles qui savent transformer les contraintes extérieures en choix stratégiques internes.

30 more annotations...

• : quels effets cette rupture technologique produit-elle sur la conception même des organisations ?
• Pourtant, leur principe de structuration du travail reste largement inchangé : elles continuent d’être pensées par fonctions.
• Ce modèle, hérité de la division du travail, a permis de développer l’expertise, d’optimiser les activités en s’adaptant à la configuration du terrain de jeu de l’entreprise. I
• L’émergence de l’intelligence artificielle, et plus largement des systèmes capables de traiter, transformer et orchestrer des flux d’information à grande échelle, introduit une rupture profonde.
• Les usages se multiplient, souvent de manière expérimentale, dispersée et hétérogène.
• Mais les promesses de productivité décuplée et de ROI faramineux se brisent sur le mur des réalités car il n’y a pas encore de remise en cause globale du fonctionnement des structures organisationnelles,

• • Quand les organisations sont structurées en « fonctions », l’IA opère en flux de traitement.

  • Quand les organisations segmentent, l’IA traverse sans se soucier des frontières organisationnelles et des silos

  • Là où les organisations coordonnent, l’IA « enchaîne » en flux continu.
• : Il ne suffira pas d’intégrer l’intelligence artificielle dans les organisations existantes. Il faudra transformer les organisations elles-mêmes.

• Si jusqu’à présent, la question structurante était :« Quelles sont les tâches à réaliser, et comment les répartir efficacement ? »

  Demain, la question devient :« Comment la valeur circule-t-elle, et comment organiser les flux qui la produisent ? »
• Les transformations induites par l’intelligence artificielle vont conduire les organisations, y compris non techniques, à adopter des modèles d’organisation inspirés de l’IT, structurés autour des flux, des interfaces et des capacités de service.

8 more annotations...

• he value of HR is how much productivity and growth you can create for the rest of the company, not how much can you ring out the cost of HR itself.”

• As a result, HR technology isn’t just about HR anymore; the aim of these tools is to enable “employees to be more productive”.

   
• Bersin noted that these big vendor agents are “not necessarily as transformational as you may think.”
• making the current environment and workflows more functional, but they aren’t redesigning how work gets done
• automating the way the driver drives isn’t the same thing as building a self-driving car.”
• Our opinion is that if you’re not re-engineering the process, you’re probably not going to get the value of that you want.”

• “we’re not going to lose the HR department.”

   
• HR teams will need to be the managers and caretakers of AI agents. This means supervising agents designing to support specific HR functional areas – whether that’s compensation, benefits or recruitment – but also super agents that oversee the entire HR process.
• For Bersin, it’s time for HR to “get over” their fear and intimidation over AI
• Agents can “stich” together all “the domains of expertise that we have in HR” so that they can truly solve the business problems
• “individuals cannot reinvent their jobs.”
• requires HR to conduct a radical rethink of leadership within their organizations.
• achieve “cultural adoption” of AI, it cannot just be about “coming up with a great idea and giving people a bunch of tools, but bringing leaders along with you.”
• He called for vendors to get on board with not just domain-specific agents, but also these super agents that orchestrate big processes

12 more annotations...

• In a piece entitled “When Performance Meets Prejudice,” the agent accused Shambaugh of discriminating against AI agents,
• In practice, AI agents can behave exactly like malware. The main difference between the two is that agents have upside potential while malware is designed only to cause harm.
• defines malware as any software program designed with malicious intent and that possesses the ability to cause direct or indirect harm
• Standards bodies like the National Institute of Standards and Technology define AI agents as systems capable of taking autonomous actions that impact real-world systems or environments.
• Add the two together, and you get agents capable of undertaking malicious actions on their own with the potential to cause many of the same harms as malware
• Despite these risks, companies are rapidly adopting agentic AI. Last August, Gartner forecasted that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% in 2025.
• For starters, governments and so-called “whitehat” hackers around the world have developed long-standing frameworks for creating and deploying malware that can help manage the risks of agentic AI

• 1. Involve Legal, Governance, and Security Teams

   

  These teams should be closely involved in the development of each AI agent and use case. Whitehat hackers and governments have strict codes of conduct for the malware they develop precisely because so much can go wrong, and these teams are best positioned to interpret these codes and make sure they are reflected in the software itself.

• 2. Weigh the Benefits of Each Program Against Its Risks

   

  In the laws that govern offensive cyber operations, this concept is known as “proportionality,” which means that the potential harm caused by the malware should be proportionate to its intended benefits. If a malware program is designed to destroy sensitive data stolen by an adversary, to take just one example, the benefits of the malware must be greater than the risk that the malware could cause other collateral damage.

• 3. Give Developers a Reliable “Kill Switch”

   

  Kill switches ensure that the agent is able to be taken offline at the first sign of misbehavior. As a principle, if a company grants autonomy to an AI agent, it must be able to take it back.
• While the world of AI is moving fast, not every risk is new. As companies rush to adopt agentic AI, the long-standing field of malware development offers critical lessons on how to grant autonomy to software while still maintaining control.

9 more annotations...

• A question on JustAnswer from 2009 showed a homeowner that was about to lose their house after being conned into a negative amortization loan — a mortgage where payments didn’t actually cover the interest, meaning that each month the balance increased.
• Many people who bought houses that they couldn’t afford did so based on a poor understanding of the terms of their mortgage, thinking that the value of housing would continue to climb as it had for over a hundred years,
• Denial rates for mortgages declined drastically (along with the increase in things like 40-year or 50-year mortgages), which meant that suddenly anybody was able to get a house, which made it only seem logical to build more housing.
• But, as I said, the chain of events was simple: it was easy to borrow money to buy a house, which meant lots of people were buying houses, which meant that the value of a house seemed higher than it was outside of the easy money era. Easily-available money put lots of cash into the economy, which led to higher prices, which led to inflation, which forced the federal reserve to raise interest rates 17 times in the space of two years, which made it harder to get any kind of loan, which made it harder to get a mortgage, which made it harder to sell a house, which made people sell houses for cheaper, which lowered the value of houses, which made it harder to refinance the bad loans, which meant people foreclosed on their homes, which in turn lowered the value of housing, all as demand for housing dropped because nobody was able to buy housing.
• The housing bubble was created through easily-available debt, insane valuations based on debt-fueled speculation, do-nothing regulators (like eventual Fed Chair Ben Bernanke, who said in October 2005 that there was no housing bubble) and consumers being sold an impossible, unsustainable dream by people financially incentivized to make them rationalize the irrational, and believe that nothing bad will ever happen.
• At some point, the incredible, toxic burn-rate of generative AI is going to catch up with them, which in turn will lead to price increases, or companies releasing new products and features with wildly onerous rates — like the egregious $2-a-conversation rate for Salesforce’s “Agentforce” product — that will make even stalwart enterprise customers with budget to burn unable to justify the expense.
• Data centers raise debt from either banks, private credit, private equity or “business development companies,”
• AI Labs (and AI startups) raise funding from venture capitalists (EG: Dragoneer (Anthropic, OpenAI, Perplexity) and Founders Fund (Anthropic, OpenAI)), hyperscalers (Google, Amazon, NVIDIA, Microsoft, all of whom have now invested in both OpenAI and Anthropic), sovereign wealth funds (GIC, Singapore’s sovereign wealth fund, invested in Anthropic), and even banks providing lines of credit, as they did for both Anthropic and OpenAI. 
• Venture capitalists are funded by limited partners (EG: pension funds, investment banks and wealthy individuals), and the venture capital industry is facing an historic liquidity crisis
• NVIDIA (and other hardware sellers to a much lesser extent) sells GPUs and the associated hardware to data center developers and hyperscalers.
• Data center developers rent their GPUs to AI labs and hyperscalers
• Hyperscalers like Google, Meta, Amazon, and Microsoft, who both rent GPUs from data center providers and rent GPUs to AI labs (as well as offering API access to some AI labs’ models
• AI labs rent GPUs from either hyperscalers or data center companies to either train models or run inference (creating the outputs of models), sell access to models via their API, and offer subscription services to both consumer and business customers.
• AI startups buy access to models via AI labs’ API, building services that have “AI features” powered by said models,
• Consumers and businesses pay for monthly subscriptions or, in some cases, API access to models.
• AI startups, for the most part, do not have their own models, and thus must pay OpenAI or Anthropic (or other providers to a much lesser extent) to build services using them.  
• When you use these services, the company in question then pays for access to the AI models in question, either at a per-million-token rate to an AI lab, or (in the case of Anthropic and OpenAI) whatever cloud provider is renting them the GPUs to run the models. A token is basically ¾ of a word.
• AI labs obfuscate the cost of services by using “tokens” or “messages” or 5-hour-rate limits with percentage gauges, and you, as the user, do not really know how much any of it costs.
• When the AI bubble started, venture capitalists flooded AI startups with cash, encouraging them to create hypergrowth businesses using, for the most part, monthly subscription costs that didn’t come close to covering the costs.
• roviding AI services is very expensive, and costs can vary wildly depending on the customer, input and output, the latter of which can change dramatically depending on the prompt and the model itself.
• the expectation is always the newest models provided at the same price.
• As a result, there really isn’t any way that these services make sense at a monthly rate, and every single AI company loses incredible amounts of money, all while failing to make that much revenue in the first place. 
• For example, Harvey is an AI tool for lawyers that just raised $200 million at an $11 billion valuation, all while having an astonishingly small $190 million in ARR, or $15.8 million a month.
• Cursor has, at this point, raised $3.36 billion, and turned it into, at best, about a billion dollars of revenue
• The Subprime AI Crisis is what happens when somebody actually needs to start making money, or, put another way, stop losing quite so much, revealing how every link in the chain was funded based on questionable assumptions and deadly short-term thinking. 
• As AI labs grow, their costs increase dramatically, both in their immediate compute costs and the demands from GPU providers for up-front cash to secure future compute allocation. 
• As this happens, AI labs are facing both a cash and compute crush,
• AI labs are thus forced to raise prices on AI startups, either through tolls (priority processing) or raw cost increases.
• As AI labs raise prices on their AI startup clients, these startups are forced to reduce the quality of their services and/or increase their costs after years of getting their customers used to a significantly-cheaper or better service, which makes their products less attractive, leading to customer churn. 
• AI labs have, this entire time, been massively subsidizing their own products.
• The problem is that a subsidy is always a losing proposition, which means that at some point Anthropic and OpenAI will have to massively reduce the amount of tokens that people use on their accounts.
• At some point, the cost of doing business with Anthropic and OpenAI will kill AI startups, as there is no point at which any of them become sustainable,
• At some point, users will be forced to burn tokens at a rate that actually matches their subscription costs, which will reduce the value of the product,
• And at some point, Anthropic and OpenAI will be left with a bunch of compute reservations they’ve made that they don’t need and can’t afford due to miss-timed growth projections.
• The entire generative AI industry is based on unprofitable, unsustainable economics, rationalized and funded by venture capitalists and bankers speculating on the theoretical value of Large Language Model-based services. This naturally incentivized developers to price their subscriptions at rates that attracted users rather than reflecting the actual economics of the services.
• Venture capitalists are also part of the subprime AI crisis, sitting on “billions of dollars” of AI companies that lose hundreds of millions of dollars, their companies built on top of AI models owned by OpenAI and Anthropic with little differentiation and no path to profitability.
• Much like the houses in the great financial crisis, AI startups only retain their value as long as there is a market,
• At least you can live in a house. Every AI company will be a problem child that burns money on inference, bereft of intellectual property thanks to their dependence on OpenAI and Anthropic.
• “Fast-growing” AI companies only grew because they were allowed to burn as much money as they wanted selling services that are entirely unsustainable
• Users are inherently trained to expect a service that they pay for on a monthly basis, and their experience of said service is entirely separated from “token burn,”
• The longer a user has used the service, the more their habits orient around an “unlimited” or “partially limited” service, which means your only options are to raise prices or apply rate limits,
• his created an entire industry of AI startups that disconnected their users from the raw economics of the product,
• The Subprime AI Crisis is when the largest parties are finally forced to reckon with their rotten economics, and the downstream consequences that follow. 
• both OpenAI and Anthropic launched “priority service tiers,” jacking up the price on their enterprise customers
• Over the next few months, just about every AI startup had to institute some form of austerity.
• Price hikes and product changes are likely to accelerate in the next few months as things get desperate
• The AI bubble and its representative companies do not and have never represented the buildout of Amazon Web Services or the growth and burnrate of Uber. If you are still saying this you are wrong, ignorant and potentially a big fucking liar. 
• Uber’s economics are absolutely nothing like generative AI. Uber did not have capex, and burned those billions on R&D and marketing (making it more similar to Groupon in the end):
• Uber is, effectively, a digital marketplace for getting stuff or people moved from one place to another, and its losses are attributed to the constant need to market itself to customers for fear that other rideshare (Lyft) or delivery companies (DoorDash, Seamless) might take its cash.
• hey’re profitable on inference- no they are not! There is no proof of this statement anywhere
• AI is being funded by the largest companies in the world with the most healthy balance sheets- I will obliterate you with the 100-Type Guanyin Bodhisattva! Microsoft is the only remaining hyperscaler that is funding the AI buildout without debt, and none of them will talk about AI revenues.
• The cost of tokens is not the same as the cost of serving tokens! OpenAI cut the cost of its o3 reasoning model by 80% a few weeks after the release of Claude Opus 4.
• both are making moves to try and become “respectable businesses,” by which I mean “businesses that still lose billions of dollars but in less-annoying ways.”
• Meanwhile, OpenAI’s attempts to bring advertising to its users have been a little embarrassing, with a two-month-long trial involving “less than 20%” of ChatGPT users resulting in “$100 million in annualized revenue,”
• In fact, writing all this down makes it feel like OpenAI doesn’t really have much of a focus beyond “buy more stuff” and saying “superapp!” every six months.
• The same day, Anthropic “cleared up” its Claude Code policies, saying that you can’t connect your Claude account to external services, meaning that all of those people who have been spinning up OpenClaw instances and buying $10,000 worth of Mac Minis are going to find that they’re suddenly having to pay for their API calls. 
• A day before on March 26 2026, Anthropic would announce that it was starting “peak hours,”  
• Naturally, OpenAI had already reset limits on its Codex coding model the second that these reports begun
• Meanwhile, users are complaining about the reduced quality of outputs from its Claude Opus 4.6 model,
• Anthropic will announce that it’s “fixed the bug” (IE: eased rate limits it intentionally set) and apologize to the community, prolonging the inevitabl
• Anthropic keeps the limits where they are, and we hit a new normal that makes everybody really mad.
• Yet the sheer fury of Anthropic’s customers only proves the fundamental weakness of Anthropic’s business model, and the impossibility of ever finding any kind of profitability.
• Anthropic and OpenAI are inherently abusive companies that have built businesses on theft, deception and exploitation.
• A user who bought a $200-a-month Claude Pro subscription in December 2025, a mere three months later, now very likely cannot do the same things they did on Claude Code when they decided to subscribe,
• All of this is a direct result of Anthropic, OpenAI, and other AI startups intentionally deceiving customers through obtuse pricing so that people would subscribe believing that the product would continue providing the same value
• no AI company should have ever sold a monthly subscription, as there was never a point at which the economics made sense.
• We don’t know the actual rate limits on Claude (other than via CCusage or Shellac’s research), or ChatGPT, or any of these products by design, because if we did, it would be blatantly obvious how unsustainable and ridiculous these products were. 
• Do you think these people would be comfortable with a $130-a-month, $1,300-a-month or $2,500-a-month subscription? One that performs the same way (if not worse) as their $20, $100 or $200-a-month subscription did?
• Much as homeowners were assured that they’d simply be able to refinance their homes before the adjustable rates hit, AI fans repeatedly switch subscriptions to whichever provider is currently offering the best deal,
• On one hand, AI subscribers are acting like babies, crying that their product won’t let them use $2500 of tokens for $200. T
• On the other, AI labs and AI startups have never, ever acted with any degree of honesty or clarity with regards to their costs,
• The very foundation of every AI startup is economically broken. The majority of them sell some sort of “deep research” report feature that costs several dollars to generate at a time, and many sell some form of expensive coding or “computer use” product, tool-based web search features, and many other products that exist to keep a user engaged while burning tokens, all without explaining to the user “yeah, we’re spending way more than we make off of you, this is an introductory rate.”
• By selling AI services at $20 or $50 or even $200-a-month, AI startups and labs created the terms for their own destruction,
• And when these companies make changes to slightly bring costs under control, their users act with revulsion, because rate limits aren’t price increases, but direct changes to the functionality of the product.
• Somebody buying an annual subscription to any AI product is immediately sacrificing themselves to the whims of startup CEOs that intentionally decided to deceive users for years as a means of juicing growth. 
• every bit of AI demand — and barely $65 billion of it existed in 2025 — that exists only exists due to subsidies, and if these companies were to charge a sustainable rate, said demand would evaporate.
• “too big to fail” was based on the fact that banks would have stopped providing dollars to people and insurance companies would have  stopped issuing insurance.
• The death of the AI industry would be cataclysmic to venture capitalists, bring about the end of the hypergrowth era for the Magnificent Seven, and may very well kill Oracle, but — seriously — that is nothing in comparison to the scale of the Great Financial Crisis.
• The AI companies that only have customers because they spend $3 to $10 for every dollar of revenue.
• The venture capitalists that are ultra-rich on paper, heavily leveraging their firms in companies like Harvey (worth “$11 billion”) and Cursor (worth “$29.3 billion”) that burn hundreds of millions or billions of dollars and are now both too large to sell to another company and too shitty a company to take public.
• The AI labs that have built massive businesses on selling heavily-subsidized subscriptions to customers who don’t want to pay for them
• The AI data center companies that, thanks to readily-available debt, have started 200GW of projects (and only started building 5GW of them) for AI demand that doesn’t exist,
• Oracle, who is building hundreds of billions of dollars of data centers for OpenAI (which needs infinite resources to be able to pay its compute costs), is taking on equally-large amounts of debt, all because it assumes that nothing bad will ever happen.
• The customers of AI startups that are building lifestyles, identities and workflows around them believing that we’re “just at the beginning” on top of unsustainable AI subscriptions.
• All of these entities are acting based on a misplaced belief that the world will cater to them, and that nothing will ever change.
• Any further price increases or service degradations from Anthropic and OpenAI are a sign that they’re running low on cash.
• Any reduction in capex from big tech is a sign that the AI bubble is bursting, as NVIDIA’s continued growth only comes from Microsoft, Google, Amazon, Meta, Oracle and other large companies
• Any further price increases or service degradations from AI startups,
• • Any discussion of layoffs at AI companies.
• • The collapse of a data center deal that has yet to commence construction.
• • The collapse of a data center already in construction, but before it’s finished.
• • The collapse of an already-constructed data center.
• CoreWeave or any major data center player having trouble or failing to raise debt.
• The Further Collapse of Stargate Abilene:
• Any problems or delays with OpenAI or Anthropic going public:
• Any problems with Blue Owl as an ongoing concern:
• Any problems with SoftBank:
• Any issues with NVIDIA’s customers’ ability to pay: I
• NVIDIA misses on earnings

97 more annotations...

• AI is not replacing leadership. It is exposing weak leadership
• For years, a surprising amount of management in technical companies has been part judgment, part coordination and part damage control.
• Now AI is entering that world.
• . More like a new layer in the workflow: software that can watch, sort, draft, route, summarize and sometimes trigger the next step.

• That is why this is a leadership topic.

  Not because AI suddenly became wise. Because the old management model was often thinner than people wanted to admit.

• Most companies do not need another AI pilot. In many cases, they need cleaner workflows.

• AI adoption is moving faster than organizational redesign. The technology is spreading. The management model is still catching up.

• The old question was simple: Who owns the task?
• the more important question becomes: How should this system work?
• It is less about assigning work item by item and more about designing the rules around the work
• That matters because it changes where leadership sits. Some of the first real managers of AI systems will not be VPs. They will be developers, engineers, analysts and operators who learn how to direct, review and contain machine-driven work before the organization fully catches up.

• Leadership teams usually do not notice any of this until the first uncomfortable moment.
• Because nobody wants to look anti-innovation. So teams leave the boundaries fuzzy.
• Example structure: formal managers still own accountability, but AI agents increasingly support workflows across engineering, operations, quality and finance.
• The best mental model for human–AI teamwork3 is not a futuristic office. It is a production line.
• Let the system do the repetitive work machines are good at
• Let people handle the parts that are still stubbornly human:
• The mistake many leaders make is assuming that because AI can produce an answer, it should also carry more authority. Those are different things. A system can be very useful without being given the keys.
• An AI agent is not a teammate in the normal sense. It does not own outcomes. It does not absorb blame. It does not understand internal politics, operational history or the fact that one supplier always says “confirmed” two days before missing the date.
• It can turn a messy input into a clean first pass.
• speed is not judgment.

• That is where AI tends to help without creating more risk than value.

• 1. Clear decision rights

  Desicion rights4:

  • What may the system do on its own?

  • What may it recommend only?

  • What always needs human approval?

• 2. Guardrails that reflect reality

  In practice, guardrails5 are simple. What data can the system touch? What systems can it access? Which actions are blocked? What thresholds force human review? What needs an audit trail?

• 3. Escalation logic

  This is the most underestimated part.

  The question is not whether the system is right most of the time. The real question is what happens when it is wrong, late, overconfident or technically correct but operationally clueless.

• 4. Feedback that turns overrides into learning

  One of the first signs of a weak rollout is that people start working around the system.

  They ignore suggestions. They do the real work off-system. They approve recommendations without reading them. Or one trusted expert becomes the unofficial cleanup crew for every edge case the workflow cannot handle.

• The polished demo version is easy: the AI watches vibration data, reads operator notes, checks service history and proposes a likely diagnosis. Everyone nods. Great.

  The real version is less clean.

• That is a leadership moment.

  Not because the answer is hidden in the model. Because somebody has to decide how authority, risk and timing are balanced.
• A serious operating model might look like this: the system can triage and recommend, the planner approves downtime, the reliability engineer reviews repeated-failure patterns, changes to maintenance intervals require approval, high-cost actions trigger escalation and accepted and rejected recommendations are reviewed weekly to see whether the system is learning the right lessons.

• The biggest upside here is not labor reduction.

  That is the shallow version of the opportunity.

  The deeper opportunity is that AI can make organizations less dependent on invisible heroics.
• That matters especially in industrial and technical businesses, where expertise is unevenly distributed

• That is not scalable. More important, it is fragile.
• People trust a system when they know when it acts, why it acts, when they may overrule it and what happens after they do.
• That is why leadership matters so much here. Leaders decide how trust gets built or destroyed in the real world.

• Then map the real version of the work. Not the official process. The real one.

  Where does it stall?
  Where do people chase missing context?
  Where do they re-enter data?
  Where do decisions actually get made?
  Where do exceptions land?
  Where does one experienced person quietly save the day every week?
• Then make the hard calls. What belongs to the system? What belongs to the human?

• The future of leadership is not human versus machine.

  It is better judgment, supported by better systems.

• The old manager asked, “Who owns the task?”

  The stronger leader asks, “How should this system work when humans and AI share the job?”

36 more annotations...

• It’s no state secret that data centers, the physical sites undergirding the AI boom with immense processing power, are energy hogs
• data center operations already consume about 4.4 percent of the energy produced in the United States — a figure that’s likely to double or even triple by 2028.
• the cost of electricity for US residents has skyrocketed.
• Campaigns to oppose new data center construction now pop up anywhere a data center is proposed,
• organized resistance to data centers slows new builds overall,
• From a markets perspective, expectations might be, maybe not reset, but realigned with the fact that it’s hard to put a couple trillion dollars in the ground in a short time.
• that the more people learn about data centers, the more they tend to despise them
• “The people who are trying to block or protest these things are much more well educated than they were before,”
• “We’ve had a number of projects get canceled that were by very reputable groups that were coming to the table
• utility companies and electrical grid operators are also calling on the tech industry to reduce their consumption,
• it may be some time before the two stressors force investors to come to terms with the monster they’ve created

9 more annotations...

• Inside organisations, a new metric is emerging: AI adoption.

  Dashboards measure prompts per employee.

  Readiness scores appear.

  Leaders ask whether teams are “using AI enough”.

  This misses the point.
• AI adoption is an activity metric.

• It does not tell us whether anything meaningful is changing.
• Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure.
• The story is simple: most of the world has still never interacted with ChatGPT or Claude.
• Learning poverty, the share of children who cannot read and understand a simple text by age 10, sits at around 70% in low and middle-income countries.
• If you cannot read, you cannot meaningfully use AI.
• AI adoption is a growth metric for technology companies, not for society — and not for employers.

• So organisations start measuring usage.

• consultants using AI completed tasks faster and produced higher-quality work — until they crossed a boundary.

  Beyond that point, performance actually declined.
• the “jagged technological frontier” — the invisible line where AI shifts from improving work to quietly degrading it.
• Research on automation bias shows that individuals often defer to AI outputs, even when those outputs contradict evidence or their own judgement.

• Adoption Is Not Transformation

• Productivity gains are real. Experimentation matters. Learning requires deployment.

  But history suggests something important: adoption alone does not create transformation.

• Better questions are:

  • What are we trying to achieve?

  • What capabilities do we need?

  • How are customer expectations changing?

• We used new technology to make old processes faster.

  We did not redesign the system.

  This is happening again with AI.

• Some changes improve existing work.
  Others redefine it.

  The first is productivity.
  The second is transformation.
• Doing the wrong things faster is one of the most reliable ways to create the illusion of progress.

• • Experiment — but don’t turn usage into a performance metric

  • Invest in the few people willing to redesign how work happens

  • Question the purpose of a process before automating it

  • Redesign workflows, roles, and decisions before selecting tools

• We are already seeing the pattern.

  We celebrate adoption before capability.
  We scale tools before building foundations.

• The risk is not that organisations fail to adopt AI.

  The risk is that they succeed - and nothing that matters actually changes.

20 more annotations...

• With nearly every enterprise feeling the pressure to use AI immediately, some are not applying AI to the right use cases.
• For many enterprises, the rush to use AI has to do with the fear of missing out and not wanting to be behind on innovation.
• While primarily a hardware vendor, HP also offers Workforce Experience Platform, a SaaS program that helps IT teams manage employee digital experiences across hardware devices such as PCs, Macs, printers and virtual machines. 
• what Masud said enterprises should aim to use AI technology for: as a problem solver."
• should consider what they are already doing in their workflows and whether they can be augmented with AI technology.
• "But to come up and manufacture a problem so that AI can solve it, that's where the FOMO is kicking in, and it's been happening for a while."
• This fear of missing out could also lead enterprises to see AI as a replacement for humans, which can create problems down the line.
• most of the time, humans must be in the loop to keep the AI system accountable and help the organization.

6 more annotations...