• global employee engagement fell from 23% to 21% last year—matching the decline seen during COVID lockdowns
• Now layer onto that: repeated restructuring cycles, investor pressure for margin discipline and transformation programs increasingly tied to automation and AI-driven efficiency.
• Most change frameworks still focus on sequencing, stakeholder mapping and communication cadence—necessary, often well-executed, but no longer sufficient for the intensity of today’s environment.
• Change does not usually fail because the structure is flawed. It falters because human response is mismanaged—particularly at senior levels.
• In today’s environment, decisions around workforce redesign feel harder to reverse. The stakes are visible and personal
• When leaders are under sustained pressure, decisions move faster than people can absorb them,

• They held town halls, they allowed questions and they provided support resources. But when I speak privately with affected employees, the narrative shifts. They describe discovering their future through a spreadsheet. Being told their role exists “for now.” Receiving reassurance that does not match reality.

   
• It is the byproduct of organizations optimized for speed, control and legal defensibility.
• High performers become more cautious. Managers wait for the next wave before committing energy. Survivors of restructuring often sit in roles that have expanded but not been redesigned, carrying additional workload while wondering if they are next

• 1. Make pressure visible in decision-making

   

  Senior leaders operate under intense pressure—markets, boards, public scrutiny. Under that pressure, timelines compress and debate narrows. If pressure distorts judgment, governance must account for it.
• Instead of asking only, “Are we aligned?” HR can ask: What will this timeline mean for manager capacity? What will this mean for retention six months from now?

• 2. Protect organizational absorption

   

  Markets may require pace. But pace does not remove the need for sequencing discipline if change is to embed rather than rebound.
• HR’s role here is to sequence discipline. Ensuring conversations precede system updates. Ensuring managers are prepared before teams are informed
• When roles change but work does not, survivors absorb the strain.

• 3. Strengthen managerial steadiness

   

  Gallup’s data makes this plain: Managers drive engagement, yet manager wellbeing is declining most sharply. Much of current investment equips managers to deliver outcomes—metrics, systems, transformation tools. Far less equips them to hold difficult conversations without over-reassuring, distancing or shutting down.
• Many HR and change functions are understandably measured on program delivery. But the steadiness of managers under pressure is what determines whether change embeds or fractures.

• A more honest question

   

  Perhaps the issue is not that change management is failing. It is that the emotional conditions surrounding change have intensified, and our frameworks have not kept pace. We have perfected the process. We have not strengthened the human capability required to lead repeated change under pressure.
• If we continue to treat change as structural choreography while ignoring the psychological load borne by managers and employees alike, engagement will continue to slide.
• The question for today’s HR and executive leaders is not whether change is necessary. It is whether we are willing to examine how it is experienced—

17 more annotations...

• I was fascinated, for example, by the economic reality that so many knowledge work organizations systematically undervalued focus, and was convinced that this provided a massive opportunity for those willing to correct for this mistake.
• “The problems I focused on in Deep Work, and in my writing since, have been getting steadily worse. In 2016 my main concern was helping people find enough free time for deep work. Today I think we’re rapidly losing the ability to think deeply at all, regardless of how much space we can find in our schedules for these efforts.”
• Distractions in the workplace intensified over the past decade with the addition of instant messaging tools like Slack and low-friction digital meeting programs like Zoom
• Outside of work, social media, which was generally still admired when Deep Work came out, has morphed into an addictive TikTok-ified slurry of optimized brain rot
• Meanwhile, new AI tools offer quick-fix short-cuts to whatever intellectually engaging work activities remain.
• we launch a revolution in defense of thinking.
• • Stop consuming social media (which is, if we are being honest, digital junk food and something adults largely need to eliminate from a healthy content diet).
      
  • Keep your phone plugged in and charging when at home instead of on your person.
      
  • Push Congress to follow Australia’s lead and ban social media for kids.
      
  • Build work cultures in which phones and laptops stay out of meetings, and find collaboration strategies that don’t require constant messaging.
      
  • Stop vague demands to “use AI” and instead carefully integrate these tools where they actually make us smarter, not just busier.
• I’m done ceding my brain — the core of all that makes me who I am — to the financial interests of a small number of technology billionaires or the shortsighted conveniences of hyperactive communication styles,”
• “It’s time to move past fretting about our slide into the cognitive shallows and decide to actually do something about it.”

7 more annotations...

• Instead of adoption we need adaptation, and conventional ‘change management’ is not going to get us there.
• But for now at least, the Capability Absorption Gap in most organisations outside of software engineering means that enterprise AI usage is developing far slower than the technology.

• Despite near-universal belief in AI’s potential, most organizations are struggling to translate adoption into real business value. Executives are facing growing pressure and challenges around AI strategy, productivity expectations, security and governance, and shifting power dynamics.

  The 2026 survey findings reveal 79% of organizations face challenges in adopting AI — a double-digit increase from 2025 — with 54% of C-suite executives admitting that adopting AI is tearing their company apart. This is despite the fact that 59% of companies are investing over $1 million annually in AI technology.
• If we continue down the incremental change management route and target marginal productivity gains from applying AI to existing ways of working, then we could end up with slightly leaner versions of last generation organisational systems, rather than better organisations overall.

• Financial Times to declare recently that the jobpocalypse narrative has been over-done:

  “Can AI do this task?” is a useful starting point for thinking about how it might impact employment, but it is an ambiguous signal that forms only one part of a large and complex picture. Considering the other factors that can shape job growth, directly or indirectly, helps to explain why thus far those occupations that are most exposed to AI are as likely to have grown as to have shrunk.
• One reason for the gap is a focus on targeting tool use and basic adoption rather than cultivating AI readiness in areas such as knowledge engineering.
• If you write a manifesto about AI transformation and never once use the word “knowledge”… you might be missing what AI actually runs on.
• I suspect there are many ways in which AI can help people collate, organise and share their knowledge that are more protective of the human element than just extraction and codification.

• We are working a lot on the challenge of cultivating AI literacy for large firms in a way that brings together all layers of a large organisation and all knowledge levels in a common narrative of world-building.
• If our ambition is not just to adopt new tools to speed up the old manual process structures, but rather to adapt the organisation to what AI makes possible, then we need people to expand their literacy and their thinking in various ways, and not just learn how to operate a chatbot.

• For everyone else, AI is just a demanding slop monster. It’s a threat. I’m not saying regular people don’t use Excel or Airtable to plan their weddings or have fun throwing PowerPoint parties, or even that AI won’t be useful to regular people over time. I think a lot of people enjoy data and tracking different parts of their lives. I’m wearing a Whoop band as I write this. I’m just saying these things aren’t everything. Not everything about our lives can be measured and automated and optimized, and it shouldn’t be.

• With AI, we are building something like self-guided machines. Whether these systems liberate or merely displace is not settled. But the possibility of leisure at scale is real enough to become a serious question.

  If AI can compress parts of instruction, it may deepen learning where it is used and clear ground for formation where it gives time back. But only if it preserves productive struggle rather than bypassing it.

  The alternative is already visible: autocomplete for life. Not just help with expression, but the slow outsourcing of judgment itself. That is Bildung’s antithesis.
• If organisations respond to AI by forcing people to become more legible to machines, they will fail, socially, culturally, and ultimately economically.
• But if they invest in human capability - judgment, learning, knowledge-sharing, and strategic thinking - they have a chance to build something far more powerful: organisations that are not just more efficient, but more adaptive, more coherent, and more human.
• It may be quicker and more effective to build net new capabilities and functions, rather than deploying AI in support of old ways of working, as factories initially tried with their electrification around the turn of the 20th Century.
• ‘Change’ is a very messy, human challenge; but there might be ways in which agentic AI can help us involve everybody and guide a distributed approach, helping track progress, and making the organisation more legible in real time.

15 more annotations...

• he business world is grappling with an AI paradox: Adoption of generative and agentic AI is growing, investment is accelerating, but sustained impact on performance is elusive.
• 94 percent of respondents report not seeing “significant” value from those investments.²“The state of AI in 2025: Agents, innovation, and transformation,” McKinsey, November 5, 2025
• This disconnect suggests a misunderstanding of how AI creates and redistributes economic value.
• Productivity improvement is unlikely to expand profit pools or provide companies with a durable advantage because competition tends to erode productivity gains, benefiting customers more than the companies that implement them.
• Real value from AI will come from reshaping offerings, business models, and market structures in ways that expand or reallocate profit pools. What matters most is not how AI will change your industry but how it will change the economics of competition.
• The questions business leaders should ask are: Through which mechanisms and at what pace will AI create, expand, or shift profit pools? And what does that mean for our strategy?

• Productivity gains: Rapidly becoming table stakes

   <!-- -->
• AI’s most immediate impact has been on productivity. By automating tasks and streamlining workflows
• Larger gains emerge when AI is embedded across entire processes.
• At the leading edge of productivity improvement, AI is beginning to orchestrate execution dynamically and at scale.
• While productivity improvements are valuable, they rarely increase profit pools—they typically affect the floor of industry performance, not the ceiling.
• To capitalize on this dynamic, companies need a head start. Early movers can scale faster, lock in lower cost positions, and make it harder for competitors to catch up once the benefits begin to compound.
• Differentiation: Fueling growth
• Product, service, and business-model innovation is where AI moves beyond improving how work is done to opening the door to new growth.
• Organizations capturing outsized value from AI-enabled innovation tend to build a few reinforcing advantages that compound and provide the foundation for new competitive moats.
• Foundational models (including large language models) are powerful enablers. However, lasting differentiation will not stem from access to these technologies—which will be increasingly commoditized—but from the way organizations use them to redesign offerings, revamp business economics, and reconfigure ecosystems.
• The first stage improves customer experience within existing industry structures
• The second stage expands the product frontier. Platforms built around some ride-hailing services have extended the model into adjacent segments
• More broadly, AI enables offerings that were previously infeasible.
• The third and most consequential stage—and the biggest catalyst of profit pool expansion—is business model reinvention.
• Advantage accrues disproportionately to organizations that move early, learn faster than peers, and build capabilities that compound even as AI-assisted practices become industry norms.
• Transaction cost reduction: Driving deep systemic shifts

• The last wave of change is the most impactful—and the least intuitive. AI promises to not only change products or operations but also market structures by radically reducing transaction costs.

   
• Many industries are built on friction.
• AI (particularly agentic systems) undermines these foundations by making information transparent, coordination automated, and decision-making near instantaneous. As a result, value shifts away from friction management and concentrates instead around new technological or strategic control points such as ownership of the customer interface or data, privileged access to demand, or the ability to orchestrate complex ecosystems.
• As AI becomes embedded in customer-facing interactions and eliminates friction, it has the potential to reshape markets in three fundamental ways
• it may change how customers discover and choose providers.
• Second, the inclusion of AI in purchase decisions sharply reduces transaction and switching costs. By automating comparison, onboarding, and migration, AI agents weaken advantages that had stemmed from customer inertia.
• Third, AI reconfigures intermediation as customer interfaces move from websites, call centers, and brokers toward conversational agents and platforms
• Given these three shifts, incumbents may need to rethink where and how they compete as friction steadily diminishes across value chains.
• Each of these pathways is less about achieving a technological advantage than securing a position that enables continued profitability as market structure shifts.
• As transaction costs fall, profit pools are not simply expanded but reallocated, with value shifting toward new control points in discovery, coordination, and customer relationships

• Competing in an economy reshaped by AI: Implications for leaders

   
• Assess the impact of AI on profit pools. Leaders should develop a detailed view of how AI will reshape industry economics—where value is created, lost, or shifted.
• Build or strengthen AI-powered competitive moats. As AI adoption spreads, differentiation will stem from how companies build and combine hard-to-replicate competitive moats.
• Turn speed into a structural advantage. The ability to test, learn, and scale faster than competitors is a primary source of advantage in an AI-shaped economy.
• Rewire the business. Capturing the AI opportunity requires a fundamental rewiring of business units and functions around a scalable AI backbone.
• In time, AI will affect every industry, but it will not create value in the same way everywhere—or for everyone.
• s: AI is not a productivity revolution—it’s a competitive reset.
• AI will not create value evenly. In many cases, it will simply redistribute it.

38 more annotations...

• Microsoft will now charge users based on the actual cost of the models they’re using, which it calls “...an important step toward a sustainable, reliable Copilot business and experience for all users
• ranslation: "we cannot continue to subsidize GitHub Copilot users, or Amy Hood will start hitting people with a baseball bat." 
• , it’s not that Microsoft was subsidizing nearly two million people’s compute, it’s that AI has become so strong, powerful and complex that it’s basically a different product!
• Microsoft was allowing users to burn more than their subscription costs in tokens every single month for three years.
• And that day has finally arrived, because every single AI service you use subsidized compute, and every single service is losing money as a result:
• AI startups and hyperscalers assumed that they’d be able to get enough people through the door with subsidized, loss-making products to get them hooked on services badly enough that they’d refuse to change once businesses jacked up the prices.
• Both assumptions were wrong, because the monthly subscription model does not make sense for any service connected to a Large Language Model.
• There has never been — and never will be — an economically-feasible way to offer services powered by LLMs without charging the actual token burn of each user, and in the process of deceiving said users, these companies have created products with illusory benefits and questionable return on investment.
• On an economic basis, a monthly subscription only makes sense with relatively static costs.
• Conversely, an AI subscriber’s costs can vary wildly. One user might only use ChatGPT for the occasional search, while another might feed in reams of documents, or try and refactor a codebase, or try and use it to put together a PowerPoint presentation.
• Anthropic, OpenAI, and other AI companies could grow their user bases, as the majority of AI users perceive its real or imagined benefits entirely through the lens of being able to burn anywhere from $8 to $13.50 for every dollar of their subscription in tokens. 
• This intentional act of deception had one goal: to make sure that the majority of people were never exposed to the true costs of generative AI.
• nobody writing about AI in the mainstream media actually understands how much these services cost,
• LLMs are inherently unpredictable. 
• These are far more forgivable if you’ve not paying on a per-token basis, because in the mind of a subscriber, that’s just another turn or two with a chatbot rather than something that’s incurring a real cost.
• I think it’s inevitable that the majority of AI subscriptions move to token-based billing
• The real thing to look out for — a true pale horse — will be a major AI lab like Anthropic or OpenAI moving all of its subscribers to token-based billing. Once that happens, you’ll know it’s closing time. 
• Uber’s CTO said at a conference that it had spent its entire AI budget for 2026 in the space of a few months,
• This is the direct result of training every single AI user to use these services as much as humanely possible while obfuscating how much they really cost.
• Sure, sure, you’re gonna say that engineers are “shipping code faster” or some such bullshit, I get it, but how much faster, and how much money are you making or saving as a result? If you’re spending 10% of your headcount on AI tokens, are you seeing that extra expense reconciled in some other way?
• For the most part, everybody you’ve read gooning over the many possibilities of generative AI has experienced it without having to pay the true costs.
• In reality, that 10-person, $1250-a-month Teams subscription likely burns anywhere from $5000 to $10,000 a month in API calls, if not more.
• Can somebody who pays $200 a month foreseeably afford $350, $400, or $500?
• If we assume an average of 21 working days in a month, that puts the average cost of a Claude Code user at around $273 a month, or $3,276 a year. At $30 a working day, that works out to $630 a month, or $7560 a year. 

• Let’s play with that $30 number a little more. 

  • For a ten person dev team, that’s $75,600 a year, and we’re only counting working days.
  • If you raise a mere three months to an average of $50 a working day, that raises to $88,200 
  • If you add a single month where you go over $100, you’re spending $102,900 a year.
  • If you spend $300 a day, you’re now spending $756,000 on tokens for ten people.

• Right now, I think most companies fall into three camps:

  • Enterprise deployments in massive organizations like Spotify or Uber with AI-pilled CEOs that allow budgets to run wild.
    • I’d also say this is the case in large, well-funded startups.
  • Smaller startups that use the subsidized “Teams” subscription.
  • Individual users paying a monthly fee to access Claude or other AI subscriptions. 
• I also think that startups burning 10% or more of their headcount in AI tokens will have a tough time convincing investors in six months that doing so is necessary.
• Per Jerome Darling of TD Cowen, it costs around $30 million in critical IT (GPUs and the associated hardware) and $14 million per megawatt of data center capacity.
• Of the 114GW of data centers supposedly being built by the end of 2028, only 15.2GW is under construction in any way, shape, or form.
• It’s also unclear whether the customer base for AI compute exists outside of OpenAI and Anthropic, whose demand accounts for 50% of AI data centers under construction,
• In any case, it’s also unclear what kind of ongoing rates these data centers charge.
• To be clear, we must differentiate between the spot cost — which is the cost of randomly spinning up GPUs on somebody else’s servers — and contracted compute, the latter of which makes up the majority of data center capex.
• Now, I should be clear that most data center companies you know of don’t actually build them, instead leaving that job to companies like Applied Digital,  who are also known as “colocation partners.” For example, CoreWeave pays a colocation fee to Applied Digital to use its North Dakota data centers. CoreWeave is responsible for all the GPUs and other tech inside the data center.
• Wrong! You’ve got $3.126 billion in IT gear to depreciate, which works out to around $521 million a year over the six years you’re depreciating it. That’s $837.4 million a year, leaving you with around $168.6 million in yearly profit, or around a 16.7% gross margin…
• …if you have 100% tenancy at all times! You see, data centers can take a month or two to get those GPUs installed and a customer onboarded, all while making you exactly zero dollars in revenue and losing you a great deal more, as you’re stuck paying your colocation, electricity and opex costs the whole time, albeit at a much lower rate (I’ve modeled for 10% electricity and 15% colocation/opex costs), meaning you’re losing about $3.27 million a day.

• And I must be clear that this is if you have 100% utilization and a tenant that pays you on time, every time.
• Stargate Abilene, an eight-building, 1.2GW data center campus with around 824MW of critical IT, was first announced in July 2024. As of April 27 2026, only two buildings are operational and generating revenue, and the third barely has any IT gear in it. I estimate the total cost of Stargate Abilene to be around $52.8 billion.
• Based on my calculations and reporting, I estimate Abilene’s rough gross margin is around 37.47% once it’s fully operational: 
• should also be clear that Oracle is mortgaging its entire fucking future on projects like Stargate Abilene, incurring billions of dollars of costs up front for a business that will take years to turn a profit even if OpenAI makes every single payment in a timely manner.
• Four months later on April 22, 2026, Oracle tweeted that “...in Abilene, 200MW is already operational, and delivery of the eight-building campus remains on schedule.” It is unclear if that’s 200MW of critical IT capacity or the total available power at the Abilene campus, and in any case, this is only enough power for two buildings, which means that Oracle is most decidedly not “on schedule.” 
• This is a huge issue. OpenAI can only pay for compute that actually exists, and only 206MW of critical IT is actually generating revenue, with the third at least a month (if not a quarter) away from doing so. 
• 
  To be clear, OpenAI’s projections have it making $673 billion over the next four years, and burning $218 billion to get there. It is an incredibly unprofitable business, and even if it wasn’t, it would have to make so much more money than it currently does to pay Oracle on an ongoing basis.
• OpenAI’s numbers come directly from The Information’s reported leaks of OpenAI’s projected burn rate and revenues, which have the company making $673 billion in revenue through the end of 2030 and burning $852 billion to get there:

• In other words, in two years OpenAI projects it will make more revenue than TSMC, in three years almost as much annual revenue as Meta, and by the end of 2030, as much annual revenue as Microsoft ($300 billion or so on a trailing 12 month basis). 

  And if OpenAI can’t pay for that compute, Oracle dies, because it’s taken on around $115 billion in debt just to build Stargate’s data centers, and needs another $150 billion to finish them:

• I really need to be clear that Oracle has no other path to making this revenue without OpenAI, and these projects are entirely financed and paid for using the projected cashflow of the data centers themselves.
• OpenAI recently missed its own targets for new users and revenue, stumbles that have raised concern among some company leaders about whether it will be able to support its massive spending on data centers.
• Chief Financial Officer Sarah Friar has told other company leaders that she is worried the company might not be able to pay for future computing contracts if revenue doesn’t grow fast enough, according to people familiar with the matter. 
• Now, I should add that Google and Amazon are far-savvier and less-desperate than Oracle, meaning that they can take the hit if Anthropic ends up running out of money. The “up to” part of that deal gives them some much-needed wiggle room that Oracle simply does not have. 
• Nevertheless, for Anthropic to actually meet its commitments, it will have to agree to spend anywhere from $25 billion to $100 billion a year on compute by the end of 2030. 

• If we assume that the 15.2GW of data center capacity under construction (due by the end of 2028) has a PUE of around 1.35, that leaves us with roughly 11.2GW of critical IT. At $14 million per megawatt, that works out to around $156.8 billion in annual revenue in GPU rental revenue required to actually make these data centers worth it.

  When you calculate for the 114GW of capacity theoretically coming online by the end of 2028, that number climbs to $1.18 trillion in annual revenue.
• To give you some context, CoreWeave — the largest neocloud with Meta, OpenAI, Google (for OpenAI), Microsoft (for OpenAI), Anthropic, and NVIDIA as customers — made around $5.1 billion in revenue, and projected it would make $12 billion to $13 billion in 2026. 
• This means that a large part — if not the majority of — AI compute revenue is dependent on a continual flow of venture capital and debt, both of which are only made possible by investors that still believe that generative AI will be the biggest, most hugest thing in the world.
• And I’ve been asking this question for fucking years. Every time we get a new update on Anthropic or OpenAI, we hear they’ve lost billions more dollars than expected, that margins are decaying, that costs are skyrocketing, that everything is more expensive despite promises that the literal opposite would happen.
• Even Cursor, a company that briefly (before its pseudo-acquisition by Musk’s SpaceX) claimed it had positive gross margins, actually had negative 23% gross margins as of January, or negative 31% if you include the cost of non-paying users, which you fucking should if you actually care about your accounting.
• The economics are built for perfection, with zero fault tolerance. They must have consistent, 100% utilization and tenancy, or they end up burning millions of dollars and fail to chip away at the years-long wall of depreciation created by the tech industry’s most expensive mistake.
• And that’s before you consider that the majority of the customers are unprofitable, unsustainable startups.
• I realize it seems a little much, but I genuinely believe that subscription-based AI services were an act of deception tantamount to fraud, as they misrepresented the core unit economics and thus the possibilities of a Large Language Model
• I need to be very clear that the product that Anthropic offers — by virtue of recent rate limit changes — is substantially different (and far worse) than the one that you read about everywhere.
• Members of the media, I say this with abundant respect: Anthropic is mistreating its customers, and it is doing so because it believes it can get away with it. T
• AI was rushed and pushed in our faces as quickly as humanly possible in the least-efficient yet most-accessible form it could be presented, even if said form was never going to result in anything resembling a sustainable business.
• and, crucially, use it in a subscription-based form that would make people experience it without ever asking how much it costs to provide.
• Think of it like this: as a reporter, an investor, an executive, or a regular LinkedIn Lounge Lizard, you might read here and there that it’s $5 per million input tokens and $25 per million output tokens, but you’ve never really experienced how fast or slow one loses that money, because it’s important to do so to truly understand this product.
• OpenAI’s CFO has now said multiple times that she doesn’t believe OpenAI is ready for IPO, and has material concerns about its growth and continued ability to meet its obligations.
• OpenAI’s CFO, the literal person who would know best, is saying that she is worried that OpenAI cannot pay its fucking compute contracts if revenue doesn’t grow.
• What the fuck does that mean? Excuse me? This company has allegedly raised $122 billion and is allegedly worth $852 billion god damn dollars and expects to burn $852 billion dollars by the end of 2030.
• For the amount of oxygen that both Anthropic and OpenAI consume, both of these companies should be fucking flawless, both as products and businesses.
• Generative AI is an insult. It is unreliable, its economics don’t make sense, its outcomes don’t justify its existence, and the perpetrators of its con are boring, oafish and avaricious men disconnected from society and anybody who would ever disagree with them.
• LLMs are constantly humming with the sense that they’re about to do something new, even though they’re mathematically restricted to repeating other actions.
• To be an “AI expert” requires you to actively ignore the worst economics of any industry in history, to constantly explain away obvious, glaring issues with products, and to actively convince others to do the same.
• For Larry Ellison’s sake, at least. Ellison has already pledged 346 million shares of his Oracle stock — or around $61.5 billion — “to secure certain personal indebtedness, including various lines of credit,” meaning “many big, beautiful loans against his Oracle shares.”
• If OpenAI can’t raise $852 billion in revenue and funding by the end of 2030, it won’t be able to pay for Stargate.
• That’ll kill the value of Oracle’s stock, leading to a series of margin calls, leading to Ellison having to sell shares, leading to further margin calls.
• What I’m saying is that Ellison’s future rides on Sam Altman’s ability to raise funding and make revenue to the tune of $852 billion in the space of 4 years.

71 more annotations...

• There’s always some uncertainty, but when you’re confident about your ability to see 30 years ahead, you build a skyscraper and a railway. When you can see months ahead, you pitch a tent and buy a bicycle.
• Think of how uncertainty about basic economic realities keeps multiplying. Dario Amodei or Sam Altman forecast the arrival of super-intelligent AI systems in the next few years. The potentially looming “jobpocalypse” and the very-much-in-the-present “SaaSpocalypse” have revealed how vulnerable our current economic situation may be to AI-powered disruption. The same goes for the large gyrations in public market valuations seemingly caused by new feature releases from the major AI companies
• Will AI eliminate the need for almost all software engineers, or will it counterintuitively create more of these jobs? How capable will robots be in a couple years, and what will that mean for the job market? Is the traditional customer service center actually toast?
• Every question like this reveals the limits of our visibility into the short-term future, and this extreme uncertainty challenges the criteria we use to commit to forward-looking investments.
• Today, the decision to invest in human capital is harder to justify. Who will attend medical school if we can’t articulate what a “doctor” even is in 2035? I don’t mean this as a cheap provocation. I mean it literally.
• The argument turns on a distinction economists rarely apply to human capital: the difference between risk and uncertainty
• Risk is quantifiable; one can assign probabilities that allow us to price a bet. Uncertainty describes environments where the probability distribution itself is unknown, which is the precise consequence of invisibility.
• For the companies that depend on skilled workers, the AI fog poses an unprecedented workforce planning problem. Leaders must decide how many people to hire, in which roles and at what salaries, against a backdrop in which the shelf life of any given skill set is unknowable.

• Simply put, expectations about corporate hiring plans and individual human capital investments are two sides of one coin.

   
• The haze of the future also throws into question how companies are valued, and therefore if and how they raise capital and communicate their worth to investors and other constituents.
• If AI fog brings into question the long-term viability of a company’s core product or service, then we must also question whether it will earn its terminal value. In turn, if we doubt the terminal value, the valuation collapses. The company is simply worth less.
• But if AI substitutes for existing software functionality, commoditizes software creation, shifts the revenue model from high per-seat subscription fees to on-the-margin token consumption, or fundamentally disrupts the enterprise software stack, what becomes of these businesses
• Can the market still justify valuation multiples of 50 times free cash flow or higher like we currently see at companies such as Samsara, Cloudflare, and others?
• The enveloping fog also affects internal capital markets. Nearly every large organization runs a stage-gate resource allocation process. A team proposes a new initiative that requires a capital outlay. Leaders evaluate the proposal against an ROIC hurdle.
• Should leadership greenlight the investment? As the fog rolls in, they must ask whether, within two to three years, a new generation of AI-controlled robots learn tolerances from a few sample runs rather than from months of setup will allow competitors (or even their own customers) to bring this production in-house at a fraction of the capital cost?
• We have entered a period in which conviction in the prospects of many long-term investments has declined.
• That stability made it rational to invest in long-lived infrastructure projects, in specialized career tracks, in sprawling supply chains, and in corporate structures designed for evolutionary—but not revolutionary—change. We have had the luxury of investing as though tomorrow would be a slightly updated version of today.
• First, it is time to shift from multi-year capital commitments to staged investment with explicit decision points.
• Second, rather than locking in organizational structures optimized for a stable future, think about how to organize for the arrival of agentic AI. If the future “operating model” of the firm is uncertain, as is perhaps foreshadowed by companies like Block, which laid off roughly 40% of its staff as part of what CEO Jack Dorsey claimed were AI-related changes, and if doubt surrounds the product portfolio and unit economics, then leaders can’t hard-wire the company for a single forecast.
• Third, leaders need better sensing systems. We are witnessing a period of constant, unprecedented change in the scale and scope of enterprise-grade AI capabilities.
• Leaders need to put in place a small team whose full-time job is to monitor frontier AI capabilities and translate technical progress into managerial implications for the firm, including what to adopt, when to adopt it, how to implement it, and what it all means for the company.
• Markets and firms will surely adapt to the fog. Investors will recalibrate. AI-native companies will be founded. Existing firms will restructure, although many may fail. None of this forecloses a generally optimistic outcom
• But the fog means that explosive growth in some sectors will coexist with wrenching dislocation in others, and it is premature to know which ones yet.
• In short, we are entering a paradigm shift in how we place bets on the future. The long-duration investments that define modern economic life—mortgages, law degrees, high-CapEx product development—emerged in an era in which the future was legible enough to justify committing to it.

• The fog is here. It’s no longer a question whether it will reshape how we invest, hire, build, and plan. The truly pressing question is: Will we redesign our institutions for opacity before investment of all sorts seizes up?

23 more annotations...

• At that time companies had on-premise client/server systems and the only alternatives were SaaS, hosted solutions run by others.
• Workday introduced a new architecture: an object-oriented database, an integrated security and business rules engine, and a novel user interface that excited companies in every industry.
• The company grew rapidly, focusing on the message “Power of One” – a single system that could handle all HR and Financial needs,
• During these years of growth, the company maintained its strong employee-first culture
• Let’s start with the main problem: what is Workday’s role in a world of newly developed, easy to build AI Agents?

• Workday is transforming the system of record into a platform for agents.

   
• If we unlock the data, security, and business rules in a company, agents can be designed and built with scale, security, and speed. Workday, as a trusted system of record, provides the company rules, policies, security model, and compliance that enables agents to run at scale. These “rails” exist in Workday today and recreating them outside of Workday is expensive, slow, and risky

• First: AI complements enterprise software. It does not replace it.

   

• Second: Workday’s “rails” are the core of enterprise AI.

   

• Third: governance and agent management can be productized by Workday.

   

• Fourth: the new Workday experience is unified and liberated through Sana.

   

• Fifth: Workday’s new commercial model aligns to outcomes, not seats, and therefore can grow rapidly over time.

   
• A customer could, in principle, extract their Workday data into a data lake, connect a large language model to it with tools, and try to rebuild the same agent capabilities outside of Workday.
   Workday’s answer: this approach produces a shadow ERP that costs millions to build, misses the unified object graph, misses the configuration system, misses the compliance machinery, and remains fragile.
• First, you can now extend Workday with the new UX and Sana. Register agents in the Agent System of Record, and it’s easy to build new apps
• Second, and even more important, Workday introduced the “Deployment Agent,” a dynamic system testing, configuration, and consultative deployment system that lets customers deploy changes more quickly.

• 1. The founder returns, a new level of energy

   

  As with Steve Jobs at Apple and Howard Schulz at Starbucks, sometimes a pioneering company loses its way. Aneel, as a leader who deeply understands technology and the Workday market, has the energy and passion to define “the next chapter.”

• 2. Workday can lead the charge for agentic HR and Finance

   

  Rather than focus on hundreds of mini-agents within Workday as-is, the company is now betting on big transformation agents for the future.

   

  By acquiring Paradox and Sana, two market leaders in AI (recruiting, agents, and learning), the company has a management team with deep experience in Agentic applications.

• 3. Sana and Paradox change workday: new leadership

   

  Workday is leveraging new leaders to make this happen.

• 4. Workday can try to define Enterprise AI Infrastructure

   

  My fourth observation is that Workday has an opportunity to define how enterprise AI architectures should evolve. This is a messy topic that seems to be baffling every company I meet.

   

  The questions we have to ask are complex: how many agents do we build and what Superagents vs. Subagents do we design.

• 5. Workday understands the context and semantic layer challenge

   

  “Everyone is ignoring the big boring problem of bad context.” — Joel Hellermark

   

  (This is music to my ears.)

   

  There’s an important lesson everyone should learn about AI: context is what creates value, and if we can’t trust the content, we can’t trust the agent.

18 more annotations...

• Peter Drucker warned in 2002 that large organizations must learn to innovate or they will not survive. Two decades later, the challenge has intensified, and at its core it remains a problem of clarity, not capacity.
• . It begins with an honest view of reality. That is precisely where most organizations fail, not because they lack information, but because they prevent existing knowledge from reaching the place where decisions matter.
• The task of management, he argued, is not to manage people but to make strengths productive.
• After a failed product launch, a missed market shift, or a costly mistake, I hear the same sentence in debriefs: “Actually, we already knew that.”
• This is not a failure of intelligence. It is a structural failure, and it is far more common than most leadership teams admit.

• Observation. Organizations tend to see only what their existing categories allow them to see. If customer dissatisfaction does not appear on a KPI dashboard, it often does not exist in management discussions, even when frontline staff encounter it daily.

      

  Description. What is observed must be translated into language, and that translation reduces complexity. Nuance disappears. Uncertainty is flattened into apparent certainty. Complex situations become colour-coded updates: green, yellow, red.

      

  Dialogue. Even in organizations that claim to value feedback, conversations are shaped by hierarchy and social norms. People quickly learn what is safe to say. Inconvenient truths are softened, delayed, or left unsaid.

      

  Intention. Strategic priorities and cultural assumptions act as invisible filters. Information that confirms the current direction flows easily. Information that challenges it meets resistance; nothing explicit, but through countless small decisions about relevance.

      

  Structures and processes. Formal workflows determine what information has a path upward. If it fits the agenda, it gets airtime. If it does not, it disappears.
• Improving how organizations hear themselves is not about adding more tools or increasing communication volume. It requires addressing the filters directly.
• Make observation explicit.
• Slow down description.
• Protect dissent in dialogue.
• Audit intentions.
• Open alternative pathways.
• Most organizations are already rich in ingenuity. It exists across functions, levels, and generations. What is scarce is clarity: the conditions that allow what people know to reach the decisions where it matters.

11 more annotations...

• The companies that survive the next decade will not be those with the best algorithms or the most data. They will be those that have the courage to abandon how decisions get made.
• Consensus management was, in its time, a rational response to complexity.
• The first is speed. Decisions pass through gaggles of lawyers, marketing teams, PR, investor relations, risk managers, and compliance, each of whom is incentivized to mitigate personal and organizational risk.
• The second is that it distorts information. In our work across dozens of corporate turnarounds, the most dangerous threat to a company was often the systematic filtering of reality as information travels upward through layers of management.
• AI pushes us into rougher waters and turns both weaknesses into critical liabilities. On the first, because it rapidly accelerates the speed at which companies can—and must—operate

• The Architecture of Speed

   

  Leaders need to redesign their organizations around a different decision-making architecture, and to invest in advanced AI-enabled information systems.

• The “Autonomous Scrum”

   

  Most organizations give small teams permission to recommend. High-velocity organizations give them permission to act. This, of course, requires rebalancing priorities in addition to changing the decision-making architecture.
• The risks of this architecture producing a suboptimal decision pale against the paralysis of endless deliberation. AI agents equip the Autonomous Scrum with an army of savants, deep subject matter expertise, and almost instantaneous feedback on what actions are producing positive results.
• We’ve had success with a different approach: the OVIS framework. One person Owns the decision. Two or three people Veto or Influence it. Everyone else Supports the outcome.

• What Boards Must Demand

   

  While the CEO and her management team must execute these changes, it is unlikely to happen without a Board of true believers—devoid of the discomfort of radical mandates
• Boards can’t continue to operate as they have in the pre-AI era. If they’re working off filtered, committee-approved reports from the C-suite they are not exercising governance—they are perpetuating the distortion field.
• The executives who thrive in this environment will share a particular disposition. They will be comfortable— even excited—making consequential decisions on incomplete information. They will trust signals over instinct, speed over process, and small teams over consensus. They will demand a world of continuous feedback loops and recognize that the greatest sin is not making the wrong call but making no call at all.
• The Great AI Reorganization will not be won by companies with the most compute or the deepest pockets. It will be won by those with the clarity to see what the new environment demands, and the courage to build it.

11 more annotations...

• The only way out of that trap is a decision. A real one. Not a committee consensus, not a hedged bet, not a roadmap that tries to please every stakeholder in the room. A genuine, uncomfortable, this-is-who-we-are-and-this-is-what-we’re-building decision.
• But there is something specific that the statistical model underneath every LLM is not designed to do. It cannot generate exceptions. That’s not a criticism. It’s just math.

• It finds the most statistically significant answer to your question based on everything it was trained on. Everything that already exists. Everything that has already been built, written, described, and shipped by every company that came before you.

  Including all of those copied playbooks.
• The model learned from an industry that was already converging. It was trained on the teardowns, the frameworks, and the “what’s working right now” content that spread because something worked once, and then everyone grabbed it.
• That happened before anyone was prompting a model. It happened because the moment something worked, everyone noticed. Someone wrote a teardown. Someone posted the conversion numbers. Someone gave a conference talk about it. And then seventeen other teams rebuilt the same thing because it felt like the safe bet.
• When you ask the model what you should build, it reaches into that pile and hands you the consensus.
• The convergence doesn’t slow down with AI. It just stops requiring any effort at all.
• When your product decisions are coming out of a model, nobody has to own them. Nobody planted the flag. Nobody said “this is the bet we’re making and here’s why I believe in it.” The model said so. And the model can’t be held accountable.

• The LLM doesn’t know your customers the way you know your customers.

  • It doesn’t know the conversation you had with a client last Tuesday where they told you something that completely reframed how you understood their problem.

  • It doesn’t know the weird strength your team has that nobody else in your market has figured out yet.

  • It doesn’t know the contrarian bet you’ve been sitting on for a year because you haven’t had the nerve to make it.
• Saying out loud: we are going to be the best in the world at this specific thing for this specific kind of customer, and we are going to say no to everything that pulls us away from that.

• Humans generate exceptions. Always have. That’s not going away.

  ---
• When was the last time your product roadmap surprised someone? Not overwhelmed them. Not confused them. Surprised them in the way that makes a customer say “I didn’t know I needed that until you built it.”
• Because the market doesn’t need another version of what already exists.

11 more annotations...