• According to a 2025 EY survey, while 88% of employees report some AI use at work, only 5% say they’re using it in advanced ways that fundamentally transform how they work—meaning many organizations are likely missing out on up to 40% of potential productivity gains because of gaps in “human readiness.”
• At work, the people who make up a team need to have the right conditions to thrive as humans first before they can unlock innovation and productivity. Respect, compassion, clear communication and a sense of safety must be present to create the ideal conditions for success.
• In the context of AI implementation, taking time to understand human behavior and psychology provides valuable insight into how to successfully curate a human culture that enables technology innovation.
• What will separate the winners from the losers in this moment will be how well they unlock the human capabilities of their people: to adapt, grow and deliver value with AI.

• 1. Establish psychological safety and a culture of experimentation

   

  Foster environments where people feel safe to speak up, share ideas and admit mistakes

• 2. Provide structured learning, training, and support

   

  Most employees currently using AI focus on basic tasks, with only a small fraction using it in transformative ways.

• 3. Align AI tools with people’s work and purpose—don’t just drop tools in

   

  When rolling out AI tools, clearly communicate why: how they map to business goals, how they help reduce tedious tasks and how they free people to focus on more meaningful work.

• 4. Reward experimentation and learning, not just output

   

  Encourage risk-taking and celebrate small wins. Even failed experiments are worth highlighting if they lead to valuable lessons and progress.

• 5. Treat AI adoption as organizational change, not a one-time project

   

  Use metrics beyond tool usage (e.g. employee sentiment, collaboration, innovation frequency, learning uptake) to track how well people are adapting to AI-enhanced work.
• People and HR leaders can’t simply provide employees with tools. They also need to help employees understand how using tools connects to the business—linking AI usage not just to efficiency or speed, but to long-term strategy, value creation and even individuals’ career paths.
• only a minority have managed to build the capabilities necessary to generate tangible value
• that individuals’ strengths and motivations align with the AI-enabled tasks and that their work advances both personal growth and organizational value
• HR leaders who can show they maximize return from that cost center—by building readiness, fostering culture and enabling human-AI collaboration—will help shape the business’s direction for years to come.
• It’s about empowering people.

12 more annotations...

• ci ne signifie en rien la fin du Web3, mais simplement la fin des profits rapides reposant sur la mécompréhension de technologies complexes.
• ann LeCun s’e’est retrouvé isolé au sein d’une nouvelle organisation chez Meta où les priorités ont changé : il n’est plus question de faire progresser la recherche, mais de sortir des produits qui contribuent au chiffre d’affaires
• a recherche sur les modèles de langage s’essouffle. Comprenez par là que les gains de puissance ou de performance sur ce type de modèle se font à la marge.
• la marge de progression des LLMs est maintenant très faible.

• il existe de nombreux types de modèle génératif, et que certains offrent une belle marge de progression :

      
   
  • Les modèles multimodaux (des images et vidéos plus réalistes) ;
      
  • Les modèles de raisonnement (des réflexions plus poussées, mieux cadrées) ;
      
  • Les modèles d’action (toujours plus d’autonomie)…
• La dernière catégorie en vogue est celle des modèles de monde (« world models« ) qui offrent une appréhension plus précise de la réalité physique de notre monde. La finalité des modèles de monde est de générer des environnements réalistes qui s’appuient sur les règles de la physique
• l va ainsi pouvoir poursuivre ses travaux sur la fusion entre les IA et les robots pour pouvoir concevoir des entités 100% autonomes, capables de découvrir un environnement inconnu, de le comprendre et d’enrichir ses capacités que cette base :
• ’il y clairement un phénomène de surchauffe. À partir de ce constat, soit on tente de calmer les espoirs au risque de neutraliser la croissance potentielle, soit on les alimente pour justement stimuler la croissance
• ce qui stimule le marché n’est pas la recherche du progrès scientifique, mais de la capacité à faire rêver les médias et investisseurs pour perpétuer le cercle vertueux du capital risque
• Le prochain palier se fera avec la montée en puissance des IA neuro-symboliques qui reposent sur l’utilisation conjointe d’une base de connaissance et d’un moteur de règles.
• 2025 a assurément été l’année des LRMs (« Large Reasoning Models« ), ceux qui ont permis aux chatbots de passer d’un apport éditorial (mieux écrire) à un apport méthodologique (mieux raisonner = mieux travailler).
• ’évolution logique des modèles de raisonnement est d’en faire des modèles de travail (« work models ») qui qui auront une compréhension fine des différentes activités d’une entreprise (ex : compta, RH…).
• 2025 a également été marqué par les débuts des LAMs (« Large Action Models« ) avec des agents intelligents toujours plus autonomes et des échanges toujours mieux encadrés (standardisation)
• nous allons probablement voir émerger une nouvelle catégorie de super apps qui carburent à l’IA
• Sinon, LE gros chantier de 2026 sera la conception d’algorithmes d’apprentissage continu (contrairement au séquentiel) pour des modèles que l’on améliore en permanence :

• les grandes sociétés technologiques sont sur le pied de guerre :

      
   
  • Google savoure son retour sur le devant de la scène avec Gemini (How Google Got Its Groove Back and Edged Ahead of OpenAI) ;
      
  • Microsoft consolide son ancrage dans le monde de l’entreprise avec Copilot, et peut s’appuyer sur la clairvoyance de son Chief AI pour rassurer les patrons (Towards Humanist Superintelligence) ;
      
  • Meta dépense sans compter, mais le pivot depuis le métavers semble plaire aux investisseurs et actionnaires (Investors Are Still Bullish on AI Stocks) ;
      
  • Amazon est en embuscade avec Alexa+ (With new Alexa website, Amazon’s consumer AI vision finally comes together) et ses offres dans le cloud (AWS Updates Its Nova Models To Compete With Google, Anthropic and OpenAI et AWS puts Kiro and other AI agents to work on truly autonomous software development)
   
      
• Ainsi, l’éditeur de ChatGPT pourrait bien créer la surprise en 2026 avec le lancement prochain de modèles sectoriels, notamment dans le secteur financier pour lequel ils ont recrutés des anciens salariés des cabinets d’audit
• Sur le terrain des usages professionnels, Claude reste LA valeur sûre : le modèle le plus puissant et l’éditeur le plus responsable, ce qui plait beaucoup aux investisseurs institutionnels
• un élargissement de la fracture numérique entre ceux qui n’utilisent pas l’IA ou se servent de ChatGPT comme de Google (pour faire des recherches), et ceux qui investissent du temps et des efforts pour comprendre les innovations et explorer les nouveaux usages.

18 more annotations...

• s these tools become more capable, many organizations hope they will spark higher levels of creativity, enabling employees to generate more novel and impactful ideas.
• A recent Gallup survey found that only 26% of employees who use generative AI report improvements in their creativity.
• We find that generative AI can indeed boost employee creativity, but the gains are not universal. Specifically, employees with stronger metacognition—the ability to plan, evaluate, monitor, and refine their thinking—are more likely to experience creative gains from using generative AI,
• to unlock AI’s potential for boosting workplace creativity, organizations must go beyond simply rolling out new tools; they also need to invest in developing employees’ metacognition and promote the thoughtful, strategic use of AI so employees can translate AI outputs into more effective creative performance.
• Employees produce more creative ideas when they have sufficient cognitive job resources.
• These resources include two key elements: a) information and knowledge, and b) the opportunity to adjust work methods and tasks, such as switching between complex and simple tasks and taking mental breaks.
• First, by expanding knowledge: Although employees’ own knowledge is limited, generative AI can provide large amounts of information within seconds. This expands employees’ knowledge base and enables them to integrate insights across domains. Second, by freeing mental capacity: When generative AI handles tasks such as summarizing texts, managing data, and drafting content, it reduces employees’ cognitive overload, allowing them to redirect resources to complex problem-solving.
• However, access to AI tools alone does not guarantee that employees can acquire the cognitive resources needed for creativity
• For example, employees with strong metacognition usually think through the steps to perform a task, keep track of how effective their approach is, and adjust when they notice a lack of progress.
• By contrast, employees low in metacognition are more likely to accept AI’s first answer, rely on default outputs, and fail to check whether AI’s suggestions are accurate or relevant.
• To examine these ideas in real work settings, we conducted a field experiment with 250 employees at a technology consulting firm in China.
• Employees with stronger metacognition became more creative when they used AI—they generated ideas that were judged as more novel and more useful.
• Generative AI does not automatically make employees more creative. What matters is whether employees have the metacognition to use AI in a reflective way.

• 1. Help employees use AI to expand the cognitive job resources that fuel creativity.

   

  Generative AI can enhance employees’ creativity by expanding their access to information and knowledge and freeing up their mental capacity for creative problem-solving. Leaders should encourage employees to use AI to gather diverse information, explore multiple angles, and offload routine tasks to restore cognitive capacity.

• 2. Raise awareness that metacognition is the engine of AI-supported creativity.

   

  Leaders might assume that integrating generative AI into their workflows will automatically make all employees more creative. Yet our research shows that creative gains tend to occur among employees who can actively monitor their own thinking and then evaluate, question, and refine AI outputs.

• 3. Build metacognitive skills through targeted and scalable training.

   

  Leaders should consider employees’ metacognitive abilities when implementing AI and invest in developing these abilities through training. Notably, metacognitive skills can be strengthened through various methods. Companies can offer short training sessions that introduce metacognition and walk employees through real examples of AI errors, asking them to anticipate, detect, and correct those mistakes.

• 4. Design workflows that promote active, iterative engagement with AI.

   

  Leaders should design workflows that position AI as a thinking partner rather than a shortcut. Instead of encouraging employees to use AI for quick answers, leaders should establish processes that involve generating multiple perspectives, comparing and critiquing AI outputs, and refining ideas across several rounds.

• Caveats and Limitations

   

  When applying these insights, leaders should consider several limitations. First, our findings are based on a single organization in China. Although the underlying mechanisms are likely to generalize, people’s attitudes toward AI may vary across countries and industries.
• Second, other personal traits—for example, motivational traits such as a strong desire to learn or to pursue ambitious goals—may also influence how effectively employees engage with AI to enhance creativity.
• hird, our study examined short-term effects within a single week. The long-term consequences of sustained AI use remain open questions.
• In sum, our research shows that generative AI can meaningfully enhance creativity—but only for employees with strong metacognition. By pairing AI deployment with deliberate support for metacognitive thinking, organizations can unlock deeper insights, accelerate innovation, and ensure that employees drive the tool rather than letting the tool drive them.
• As generative AI becomes woven into global workflows, cultivating employees’ metacognition will be what separates organizations that are merely adopting AI from those that are truly unlocking its creative power.

20 more annotations...

• Productivity improves in pockets of the organization. Yet, somewhere between year one and year three, momentum fades.
• Transformation programs quietly stall, not because the technology failed, but because the organization never evolved how decisions were made,
• Most transformation discussions focus on what to implement. Fewer address how the organization itself must change to sustain that implementation.
• The operating model defines how work actually gets done. It includes decision rights, funding mechanisms, accountability structures, performance incentives, and governance rhythms.
• In many enterprises, digital initiatives are layered onto legacy operating models designed for stability, predictability, and functional silos.
• The result is friction. Teams are asked to move faster, but approvals remain slow. Leaders want agility, but funding cycles are rigid.
• Initial automation or AI initiatives often target contained problems. They improve a single workflow, optimize a specific function, or enhance a narrow customer interaction. These efforts succeed because they operate within existing structures and require minimal organizational change.
• Leadership interprets this success as proof that the transformation strategy is working, but scaling is different from piloting.
• Without explicit changes to how decisions are made and who owns outcomes, transformation stalls.
• In mature digital organizations, decision rights are clear.
• Technology teams build solutions, but business leaders retain final authority. Data teams generate insights, but operational leaders decide whether to act on them.
• AI models make recommendations, but no one is accountable for trusting or challenging the output.
• This ambiguity creates hesitation. Teams defer decisions upward. Leaders delay action. Innovation slows.
• Many enterprises still fund transformation initiatives as projects rather than capabilities. Budgets are approved annually, tied to fixed scopes and predefined deliverables
• Digital capabilities, however, require continuous investment.
• Leaders often underestimate how deeply funding mechanisms shape behavior. Teams optimize for project completion rather than long-term impact
• Cross-functional initiatives involve multiple stakeholders, each with partial responsibility but no clear ownership.
• When outcomes disappoint, accountability fragments. Issues are attributed to data quality, user adoption, integration complexity, or regulatory constraints.
• This diffusion of responsibility is one of the most damaging effects of a misaligned operating model.
• process optimization alone cannot overcome structural misalignment.
• True transformation requires aligning process excellence with operating model evolution.
• First, leaders must explicitly define decision ownership for digital capabilities.
• Second, funding models must shift from project-based investment to capability-based investment.
• Third, accountability mechanisms must be strengthened, not diluted, as complexity increases.
• Digital transformation is often described as a journey. In practice, it is a discipline.
• The difference is not ambition or investment. It is whether leadership is willing to redesign how the organization works, not just what tools it uses.
• The true test of digital leadership is not whether an organization can adopt new technology. It is whether leaders can evolve the operating model to support continuous change.
• Technology enables possibility. Operating models determine whether that possibility becomes performance
• Digital transformation does not break because leaders lack vision. It breaks because the organization remains designed for a world that no longer exists.

27 more annotations...

• Full utilization1 means trying to keep people busy close to 100% of the time.
• In many kinds of modern work, it quietly produces the opposite: long delays, missed commitments and a constant sense of urgency without real movement.
• Organizations don’t succeed or fail on effort alone.
  They succeed or fail on flow2.

• Somewhere between planning meetings and delivery dates, many organizations lose sight of something simple:

  Customers don’t see your internal effort.
  They experience waiting.
• A system can look full and still move slowly.
  Sometimes because it’s full.

• Most lead time isn’t spent on actual work.
  It’s spent waiting.

  Waiting for decisions.
  Waiting for handoffs.
  Waiting for capacity.
  Waiting for “the right moment.”
• You can perfect every step and still disappoint customers
  if most of the time is lost between them.

• There’s a simple relationship that shows up almost everywhere work moves through a system:

  Lead Time3 = Work in Progress4 ÷ Throughput5

• The more unfinished work you allow into the system, the longer it takes for anything to reach the customer.
• Little’s Law doesn’t tell you where time is lost.
  It tells you why it accumulates.
• Not because engineers are slow.
  But because the work has to move through builds, tests, iterations, suppliers and external constraints.

• When you look closely, those 18 months break down unevenly.

  Actual working time: roughly 6 months
  Designing hardware, writing firmware, bringing up boards, debugging, testing, fixing, integrating.

  Waiting and idle time: roughly 12 months
  Waiting for parts.
  Waiting for build slots.
  Waiting between EVT, DVT and PVT.
  Waiting for test results, approvals, lab availability, supplier lead times and manufacturing readiness.

• You reduce the time work spends sitting idle between steps:

  • fewer parallel initiatives competing for the same labs and decision-makers

  • clearer prioritization between hardware and firmware work

  • faster decisions between builds

  • earlier supplier and part choices to avoid blocked builds

• A project that’s “almost done.”
  A decision that just needs one more alignment.
  A release that slips by another week (again).

• In practice, throughput in knowledge-heavy work is often capped by things that don’t show up on a plan or org chart:

  • limited attention.

  • coordination overhead.

  • context switching.

  • dependence on other people’s availability

• Real work is uneven by nature:

  • some tasks are quick, others messy.

  • work arrives in waves.

  • priorities shift mid-stream.
• High utilization works best when work is standardized, predictable and heavily buffered.

• Here, waiting doesn’t look like a queue.
  It looks like stalled projects, delayed releases and work that is “almost done” for weeks.

  Applying the same utilization logic without the same buffers usually backfires.

• “Idle time is waste.”

  It sounds sensible.
  It’s also misleading.

  Slack8 isn’t about comfort or perks.
  It’s what allows a system to stay stable when reality refuses to cooperate.

• You don’t need formulas to recognize it.

  • everything is urgent

  • simple tasks wait for “the right moment”

  • planned work gets displaced by the latest escalation

  • projects start easily and finish painfully

  • people are busy, yet outcomes arrive late

• In practice:

  • limit what counts as active.

  • finish before starting more.

  • make queues visible.

  • protect slack deliberately.

  • pay attention to time, not just effort.

• • List what’s actively in motion

  • Notice what’s actually moving

  • Pause the rest (explicitly)

  • Watch one signal: time to market (or time to completion)

• Full utilization feels responsible.
  In many kinds of work, it’s a slow-motion failure mode.

• Ask something more uncomfortable:

  What are we starting that we don’t need to start yet?

22 more annotations...