• All those ambitions were vaporized, however, when the ship suddenly burst into flames as it attempted a landing in New Jersey. The horrific fireball was attributed to a critical flaw: the hundreds of thousands of pounds of hydrogen it was filled with were ignited by an unfortunate spark. 
• You’ve got a technology that’s very, very promising, but not as rigorously tested as you would like it to be, and the commercial pressure behind it is unbearable.”
• Perhaps AI could be responsible for a catastrophic spectacle, such as a deadly software update for self-driving cars, or a bad AI-driven decision collapsing a major company,
• Together, these can encourage a user’s negative thoughts and lead them down mental health spirals fraught with delusions and even full-blown breaks with reality. These episodes of so-called AI psychosis have resulted in stalking, suicide and murder.
• OpenAI alone admitted that ChatGPT that more than half a million people were having conversations that showed signs of psychosis every week.
• We need to understand that these are just glorified spreadsheets, they are tools and nothing more than that.”
• If AI has a place for us in the future, it should be as cold, impartial assistants — not cloying friends that pretend to have all the answers.
• “That’s not what we get. We get an overconfident AI that says: yes, here’s the answer,” he told The Guardian. “Maybe we need AIs to talk to us in the voice of the ‘Star Trek’ computer. You would never believe it was a human being.”

6 more annotations...

• Going by their customer relationship management (CRM) systems, their processes looked indistinguishable.

• they diverged in ways that materially affected revenue, risk, and speed.

   
• The first company serves regulated, mission-critical industries such as utilities, financial institutions, and public infrastructure. Over time, its sales organization has evolved a distinctive way of working.
• The second company serves digitally driven enterprises undergoing rapid transformation, competing on speed of rollout and time to value. Early momentum matters more than perfect certainty.
• In both cases, none of this behavior is captured in the CRM.
• Yet this context, ephemeral as it may be, makes up the fabric of an organization. These patterns of judgment, coordination, and trade-offs shape how work actually unfolds.
• In doing so, we have examined more than 200 distinct work patterns across more than 50 large enterprises, mostly Fortune 500 or European equivalents.
• Context is demonstrated execution: the workflows teams actually follow across systems, the signals they respond to, the order in which roles get involved, the exceptions that trigger action, and the judgment calls that repeat across real work. These patterns are visible only in execution, not in stated process.
• Strategy experts have long argued that sustainable advantage rests on resources that are valuable, rare, difficult to imitate, and non-substitutable. Organizational context meets all four criteria.
• Many companies use the same models and platforms, yet results vary dramatically. The reason is not model quality. It is context.
• AI systems are general by design. They cannot infer which signals matter, which trade-offs take precedence, or how risk and speed are balanced unless those patterns are made explicit.
• This also explains why many AI pilots fail to scale. Early experiments show promise in controlled environments, but when deployed across real workflows, impact plateaus.
• Historically, organizations embedded context by redesigning workflows and systems. Today, that same contextual knowledge can ground models and agents directly. The leverage shifts from redesigning systems to amplifying judgment.
• That distinction does not live in the CRM. It lives in the organization’s accumulated experience across many deals and roles.
• In finance, ERP systems record transactions but not the judgment behind credit decisions or capital allocation trade-offs. Across functions, systems of record capture outcomes. They rarely capture how execution unfolded.
• AI ROI is not determined by model capability alone. It is determined by how precisely intelligence is grounded in the organization’s operating context.
• To convert AI investment into measurable return, leaders must build a context layer that grounds intelligence in operational reality.

• 1. Instrument work to capture context

   

  Invest in technologies that map how work unfolds (the work graph) across systems and collaboration environments, creating a continuously updated view of decision sequences, coordination patterns, and trade-offs.

• 2. Make the context library a runtime layer for AI

   

  Every AI system should connect to the context library. That requires technology capable of inferring user intent and retrieving the precise slice of context relevant to the task.

• 3. Establish governance and trust

   

  Context engineering requires clear privacy, security, and access controls. Sensitive information must be managed or redacted.

• 4. Monitor impact and close the ROI loop

   

  Instrument outcomes alongside AI usage. Where forecasting accuracy improves, deal velocity increases, risk events decline, or cycle times compress, contextual grounding is working.
• In a model-driven enterprise, context becomes infrastructure: a continuously updated representation of how the organization works. Competitive advantage and AI ROI depend on how tightly these two layers are integrated.

20 more annotations...

• Si vous rajoutez à cela les innombrables prophéties de techno-disruptions aussi irrémédiables que subites, vous obtenez un raz-le-bol général, un rejet naturel tant le sujet est dense et omniprésent :
• Cette débauche publicitaire trahit l’empressement des éditeurs de services d’IA de nous faire adopter leurs outils en toute urgence
• . La dernière étude de Microsoft France est ainsi très révélatrice : L’usage de l’IA progresse dans les entreprises françaises, mais se heurte à un manque d’accompagnement.
• la dernière édition du Baromètre du numérique de l’Arcom qui confirme cette dissonance entre adoption et bonne utilisation : si la moitié des Français utilisent désormais l’IA (en progression de 47 % en un an), seul un quart s’en dit satisfait.
• Ceci ne signifie en rien que l’IA est inutile, mais que la majeure partie des utilisateurs n’en ont qu’un usage superficiel :
• les modèles génératifs de dernière génération nous offrent des capacités qui ne sont plus exploitées à travers des outils (que l’on utilise), mais des agents (qui agissent à notre place).
• un système de Plugins et de Skills qui semble devenir la norme et qui facilite considérablement le déploiement et l’utilisation des agents :
• l’approche agentique n’est plus une expérimentation, c’est un standard industriel en cours de déploiement.
• La situation est très préoccupante, car ce que l’on pensait arriver dans 3 à 5 ans, arrivera à coup sûr dans 3 à 5 mois, voire 3 à 5 semaines. Les cols blancs vont connaitre une révolution équivalente à ce qu’ont connu les développeurs, mais ils n’y sont pas préparés
• Mais ces changements sont invisibles pour ceux qui utilisent les versions gratuites des outils d’IA, celles qui sont fortement limitées en puissance et en fonctionnalités.
• D’énormes changements sont à venir pour des salariés qui sont bien souvent réfractaires au changement, contrairement aux populations informatiques qui sont habituées à s’adapter rapidement aux innovations technologiques (quasiment tous les ans).
• , l’adoption forcée de l’IA se fait au pas de charge : les entreprises se précipitent tels des poulets sans têtes qui courent par réflexe sans voir où ils vont
• . Une étude du MIT le confirme sans ambiguïté : l’IA générative ne diminue pas la charge de travail, elle l’augmente
• Nous retrouvons ici une parfaite illustration de l’effet rebond du paradoxe de Jevons : des outils plus efficaces engendrent plus d’usages, donc une plus grande utilisation.
• le travail serait très prochainement accaparé par des salariés séniors qui, assistés par des agents intelligents, pourraient fournir autant de travail que 3 à 4 ETP
• l’adoption de l’IA est une charge que l’on confie de façon implicite aux jeunes, en espérant qu’ils vont participer activement à l’évolution des méthodes de travail et former les plus vieux.
• : 94 % des étudiants français ont déjà utilisé l’IA, entre usage massif et dérives inquiétantes. Autant vous dire que miser sur la génération suivante pour résoudre un problème d’aujourd’hui n’est pas une stratégie, c’est un aveu d’impuissance.

• il y a une nécessité de prendre du recul et de ne pas se précipiter par peur d’être dépassé… par qui et par quoi déjà ?

      
• l’important n’est pas d’accélérer l’adoption, mais de l’améliorer : de savoir quoi faire avec l’IA, et comment bien le faire
• Voilà pourquoi il ne faut pas céder au FOMO (« Fear of missing out« , la peur de manquer) et déployer l’IA de force et en urgence. Car la précipitation n’a jamais produit que de la dette (technique, numérique, organisationnelle, humaine…). Et celle-ci, contrairement aux outils, ne se met pas à jour en un clic.

18 more annotations...

• Développer une littératie en intelligence artificielle, c’est-à-dire une compréhension minimale de son fonctionnement, de ses limites et de ses effets, permet de rester acteur des usages.

• L’IA générative : un système sociotechnique, pas une intelligence autonome

• Or, les sciences sociales nous rappellent que les technologies ne sont jamais neutres. Bruno Latour (1992) explique que les objets techniques cristallisent toujours des choix humains, des priorités, des arbitrages et des valeurs. Une fois intégrés dans la technique, ces choix deviennent moins visibles : ils finissent par apparaître comme allant de soi, alors qu’ils pourraient être discutés ou remis en question.
• L'IA générative s'inscrit ainsi dans des chaînes d'approvisionnement mondialisées, dont les coûts écologiques et humains sont largement externalisés vers les pays du Sud global. Elle est donc profondément matérielle et située.
• À cette matérialité s’ajoute une dimension souvent invisibilisée : le travail humain qui rend ces systèmes possibles
• Ces dynamiques ne produisent pas uniquement de la précarisation, mais également de nouvelles formes d’emploi : métiers hybrides, fonctions d’évaluation de modèles, modération algorithmique, annotation spécialisée ou ingénierie de prompts.
• es formes de travail échappent largement aux cadres classiques du droit du travail, de la protection sociale et de la reconnaissance institutionnelle.
• L’IA générative n’est donc ni autonome ni immatérielle : elle est soutenue par une chaîne mondiale de travail, d’infrastructures et d’organisations. Rendre cette chaîne visible est une condition centrale d’une approche responsable
• Sur le plan technique, les modèles d’IA générative fonctionnent par prédiction statistique. Ils recombinent des formes apprises à partir de vastes corpus de données, sans compréhension du sens.
• Les systèmes produisent de la vraisemblance, non du jugement.
• ces systèmes automatisés intègrent et amplifient des inégalités sociales, raciales et de genre déjà présentes dans leurs données et leurs conditions de conception.
• La question n’est plus seulement ce que l’outil permet de produire, mais si cet usage justifie les ressources mobilisées.
• choisir quand recourir à l’IA générative, pour quels besoins, et quand des solutions plus simples, humaines ou techniques, sont suffisantes

• Face à ces enjeux, la réponse ne peut être uniquement technique. Elle passe par le développement d’une littératie en intelligence artificielle.

• Nous proposons donc de distinguer trois dimensions complémentaires de la littératie IA :

  - La littératie technique : comprendre les principes de fonctionnement, les limites et les capacités des systèmes

  - La littératie critique : interroger les choix incorporés, les biais encodés et les effets sociaux et environnementaux

  - La littératie organisationnelle : savoir négocier les usages, identifier les contextes où l'IA apporte de la valeur, et construire des cadres de décision collectifs
• Le prompt devient ainsi une interaction située, dans laquelle l’humain explicite, ou non, son point de vue, ses attentes et ses contraintes.
• omprendre que le prompt est toujours situé, et que l’esprit critique passe aussi par une attention aux limites et aux biais propres à chaque système.
• la manière dont les utilisateurs interagissent avec l'IA générative détermine significativement leurs performances et leur capacité d'apprentissage
• l'IA générative peut soutenir l'autonomie cognitive, mais uniquement si elle est utilisée de manière critique et réflexive.
• la sobriété suppose surtout une interrogation collective sur la pertinence même du recours à l'IA générative
• L’IA générative continuera de se diffuser dans les organisations. La question centrale n’est donc pas de savoir s’il faut l’utiliser, mais comment en garder la maîtrise.

• identifier quand l’IA apporte de la valeur, quand elle n’en apporte pas, et sous quelles conditions.
• : au niveau individuel, développer une littératie IA et des pratiques de prompting réflexif ; au niveau collectif, construire des cadres de décision partagés et des espaces de délibération sur les usages ; au niveau institutionnel, intégrer les coûts sociaux et environnementaux dans les décisions d'adoption, et développer des métriques alternatives à la productivité brute.

21 more annotations...

• Josh was practicing a specific leadership capability—one that most organizations do not yet have language for, one that I call experience intelligence, or the ability to read and shape the human experience.
• Under his leadership, cast members did not respond because they were pressured. They responded because they felt seen, trusted, capable and proud of what they were a part of.

• Experience intelligence requires leaders to recognize two key insights.

   
• The first is that the tools they tend to use to change people’s behavior—goals, feedback, and praise for employees; pricing, rewards, and loyalty programs for customers—are directive and produce only short-term results.
• The second is that the most powerful experiences are those which the person says they love—not “like,” not “respect,” not “learned a lot during,” not “really enjoyed.” While each of these are positive feelings, it turns out they do not predict what a person is going to do next. A wealth of psychometric data show only love is predictive.
• Tim Massa of Kroger has it. He changed his title from EVP of people to chief associate experience officer and led the grocery giant through a very challenging 18 months by taking both love and experience seriously.
• Susannah Frost, the president of Chick-fil-A has it, as does COO Cliff Robinson. Though they are distinct in their personal styles, they are both deeply curious about how to wrap the chicken sandwiches they sell in genuinely loving experiences for both team members and guests. Understanding that experience drives behavior, they encourage all franchise operators to “be the customer.”
• Patti Poppe, the CEO of Pacific Gas and Electric Company (PG&E) has it. You can see it in her unremitting belief that the secret to restoring trust in PG&E was to love the communities they serve. She launched a “leading with love” campaign, which included a series of town halls with local communities and towns; the redesign of service onboarding processes for both individuals and business; and the burying of more than 1,200 miles of power lines as a wildfire safety effort.
• . Many leaders rely on blunt force, chest-thumping directives, transaction, and extraction. They treat human beings as variables to be controlled rather than sources of energy to be cultivated. Or they design for process, optimized for efficiency, speed, or cost
• The most intelligent organizations are selecting leaders who understand that experiences where people say, “Love that!” are by far the most powerful force in business.

8 more annotations...

• He focuses on how the AI agents are working, but also how they are learning and adapting—much like how a traditional manager might walk the floor, check in with a struggling employee, or huddle with a team on a tricky case.
• The Agentforce platform, which Salesforce is using itself and selling to client companies, now autonomously resolves nearly 74% of the company’s inbound customer support cases. Dozens of digital “employees,” all AI agents, are resolving customer issues, drafting personalized emails, and routing more complex cases to human specialists.
• Overseeing this hybrid workforce are agent managers—people like Stauber, who runs an AI-powered human plus technology team.
• a profound shift in management is underway.
• agent managers are fast becoming the connective tissue between strategic intent and autonomous execution
• The SDR agent ensures the sales cycle never stops, “while my team is sleeping, our agents are already interacting with customers,” said Vanessa Tabbert, VP of the team.
• The evolution to agentic SDRs transformed the human seller’s role from one focused on low-connect prospecting to one centered on high-value human interaction, empathy, and creative problem-solving.
• Achieving sustained success with agentic AI requires an attitudinal shift championed by line of business (LOB) owners
• Lines of business owners should be responsible for designing, testing, and governing the agents that power their workflows, much as they would for the human workforce
• This approach creates both opportunity and complexity. To be successful, agent managers must blend deep functional expertise with operational AI literacy. They must know not only what the business wants, but how to teach an AI agent to achieve it, safely, consistently, and transparently.

• Depending on organizational maturity, agent managers may report to:

   
   
  • Digital customer success (as at Salesforce),
   
  • Sales management (with tight alignment to a centralized AI Operations team),
   
  • Or a cross-functional Digital Transformation Office
   
   

• The role typically involves:

   
   
  • Monitoring agent performance: quality, speed, escalation, and customer sentiment
   
  • Refining prompts and workflows to improve accuracy and tone
   
  • Managing handoffs to human agents and ensuring escalation when required
   
  • Conducting root-cause analysis of failed cases to drive continuous improvement
   
  • Quantifying impact through ROI analysis and executive reporting

• Their success depends on six critical capabilities:

   
   
  1. AI operational literacy: They understand how agents operate, how prompts drive outcomes, and how to diagnose system failures.
   
  2. Functional depth: They possess deep knowledge of the business process the agent supports, whether customer service, finance, or logistics.
   
  3. Systems thinking: They visualize how agents interact across workflows, departments, and even other agents, to achieve “multi-agent orchestration.”
   
  4. Change resilience: They adapt quickly to shifting models and business needs, refining agent logic in weekly “test-deploy-learn” cycles.
   
  5. Prompt craftsmanship: They excel at designing and refining the language and logic that shape agent behavior, the equivalent of employee training for machines.
   
  6. Designing work across machines and humans: They know how to create hybrid workflows, assessing limits of machine capabilities, creating human escalation routines as needed. They’re adept at motivating the human workforce in the context of hybrid AI-human work.
• Successful management in this hybrid era also requires a shift in how human performance is measured. The focus on activity-based KPIs (for instance, making 60 calls a day) is obsolete. The new model shifts KPIs to focus on outcomes that depend on orchestration and influence—a recognition that performance now depends on how well a person is tuning their agent and running workflows with it.
• The new skill requirements include “managing AI,” i.e., knowing how to command the agent effectively and “managing engagements,” i.e., developing the acumen necessary to engage successfully in the high-value human interactions.
• While agent managers can come from a variety of disciplines, the most effective ones emerged from roles already accountable for service quality, customer outcomes, and operational judgment.
• Early deployments also clarified how agent work should be structured. Agent managers focused on using natural language, shaping intent, judgment, and tone by translating complex business logic into simple, adaptive instructions an AI agent can follow
• The broader lesson is that agent management is not necessarily a technical role. Organizations that developed the role successfully treated the role as an apprenticeship, immersing managers in live operations, failure reviews, and iterative test–deploy–learn cycles, while clarifying decision rights and escalation paths early.
• the need for dedicated orchestration will only grow.
• They will need a blend of business insight, AI fluency, and ethical judgment.
• This moment therefore demands a call to action. Within 12 to 18 months, “agent manager” will likely be a standard title in AI-first enterprises, a career path for leaders who know how to scale impact through intelligent automation.

19 more annotations...

• he features that make AI coding tools useful for programmers can be equally valuable for knowledge work that involves no programming whatsoever.
• Imagine that you have an important meeting coming up with a new sales prospect.

• Since AI chatbots can visit web pages and extract information, you can use them to speed up part of this process. You go to the web interface (or desktop app) of your favorite chatbot and prompt the AI:

   

   

  Research this person [LinkedIn URL] and their company [company URL]. Find recent news about the company and industry. Summarize what I should know before meeting with them. Create a one-page briefing document.
• In other words, conversational AI tools are good at producing a one-time output, but they don’t easily support repeatable workflows.
• With an agentic AI coding tool, you start with a prompt similar to what you’d use with a chatbot, but the tool works directly with the files on your computer.1 It can read files you already have and, most importantly, edit them and create new files

• For example, you can create a folder called “meeting-prep” and prompt the tool:

   

   

  Research this person [LinkedIn URL] and their company [company URL]. Find recent news. Check my notes folder for any past correspondence. Create a briefing document and save it to this folder.
• The folder now contains not just the briefing but also an INSTRUCTIONS.txt document that captures the approach you arrived at through iteration: what sources to check, what sections to include, how long it should be, and what to emphasize. The instructions reflect the refined version of the task, not your first attempt.

• The tool has no memory of your earlier conversation, but that’s easily fixed: You can point it toward the folder and prompt:

   

   

  Follow INSTRUCTIONS.txt to create a briefing for [new LinkedIn URL] and [new company URL].
• When a colleague needs to prepare for their own meetings, you can share the folder. They’ll see the instructions and can run the same process themselves.
• A form of “memory.” The tool can read what you’ve accumulated in folders, write outputs directly where you need them, edit files incrementally based on your feedback, and work with your existing folder structure.
• Automation. When source materials change, you can re-execute the same analysis with a simple plain-language instruction, without having to manually repeat the workflow.
• Parallel execution. These tools can run multiple tasks simultaneously, with each task operating independently
• The first building block is multistep reasoning. The tools can break a complex task into a sequence of actions
• The second building block is adaptive execution. As the tools execute each step, they observe the results and course-correct when something fails
• The third building block is tool use. The tools can invoke processes that an external system executes on their behalf, such as reading a file, running a command, browsing a web page, or connecting to services like Google Drive or Slack or to databases.
• 1. Competitive intelligence. Say that a product team maintains a folder containing files on each competitor — positioning, pricing, recent announcements — alongside documentation of their own product. By pointing a coding tool at this folder and asking it to execute the INSTRUCTIONS.txt file in that folder, competitor information can be refreshed with up-to-date information,
• 2. Marketing campaign versioning. Launching a campaign across multiple audience segments and platforms typically requires dozens of copy variations. By storing brand guidelines, value propositions, and audience personas as files, platform-appropriate copy for each segment can be generated in parallel very quickly.
• 3. Due diligence review. Consultants and analysts frequently work with large volumes of documents — such as financials, contracts, and HR materials — spread across dozens of files. Because coding tools like Claude Code operate locally, they can navigate folder structures, cross-reference documents, and produce structured outputs without the manual uploading and downloading required by browser-based tools.
• An instruction to “clean up this folder” could delete files you intended to keep
• Another potential issue is that files processed by these tools can contain hidden, malicious instructions — a security risk known as prompt injection.
• The ability to work directly with the files on your computer is a seemingly modest one, but it vastly expands the potential for automation and enables the valuable accumulation and reuse of institutional best practices and knowledge.
• While these tools have historically been marketed to developers and discussed primarily in technical forums, their capabilities are changing fast, and we are at the early stages of their crossover into the nontechnical mainstream.

20 more annotations...

• Avec son modèle Rendanheyi, Haier a métamorphosé son organisation pyramidale, maintenant composée de milliers de micro-entreprises autonomes (plus de 4000). Chaque unité, constituée de 8 à 12 personnes, définit son offre de valeur, qu'elle soit interne (fonctions supports) ou externe (vers le marché)
• Elles évoluent dans une logique de marché interne concurrentiel où chacune choisit ses prestataires (internes ou externes) pour servir ses clients/utilisateurs (Externes/internes)
• Elles assument pleinement la responsabilité de leurs résultats (P&L) et disposent d'une liberté totale pour s'organiser, recruter et … délivrer du résultat.
• préfigure ce qui se dessine aujourd'hui avec l'IA : des équipes réduites, pluridisciplinaires, fortement outillées, capables de piloter un produit ou un service sur toutes ses dimensions.
• L'autonomie ne sera plus une simple marge de manœuvre que l’on défend : elle deviendra l'ossature de l’organisation du travail.
• le collaborateur devient un « entrepreneur » au sein de l'écosystème interne. Sa micro-entreprise prospère ou disparaît.
• Avec l'IA, cette même logique pourrait se généraliser : objectifs plus ambitieux, cycles raccourcis, attentes accrues, comparaison permanente des performances entre équipes, marketplaces interne
• Plus d'autonomie formelle, certes, mais davantage de responsabilité individuelle, une charge cognitive accrue, moins de zones tampon pour absorber les chocs, moins de confort, moins de possibilité de se retrancher derrière l'organisation
• cette autonomie radicale ainsi structurée a généré une accélération spectaculaire des décisions, une forte capacité d'innovation, une proximité utilisateur/client et une responsabilisation accrue avec des performances économiques assez bluffantes : +50% de CA, …
• Symptôme révélateur : 12 000 cadres ont refusé la première mouture du changement. L'autonomie et l’engagement deviennent des mécanismes de sélection implicite, produits par le système lui-même.
• Haier a dû corriger son propre modèle dans une deuxième vague de Test and Learn. Face aux effets pervers d'une concurrence généralisée interne, l'entreprise a introduit en 2019 les Ecosystem Micro Communities, réinjectant coopération et régulation collective, au sein de plateformes rassemblant par thématiques les mico-entreprises internes.
• : l’« exponentialisassion » des initiatives générées par ces équipes autonomes devra être régulée, dirigée, contenue, transversalisée au risque de voir le chaos s’installer
• Cette horizontalité d’Haier n'a pas fait supprimer le management central. Elle l'a déplacé vers la conception des règles, le design des plateformes, la définition des métriques, l'arbitrage des tensions.
• Avec l’IA, ce déplacement vers un management organisationnel de régulation va s’accélérer, afin de coordonner et de faire coopérer des équipes devenues pleinement autonomes dans leur activité.
• Ce glissement du management des personnes vers le management de l’organisation impliquera le développement de nouvelles compétences : la capacité à comprendre, orienter et réguler l’organisation et ses flux

• Ces nouveaux métiers reposeront sur trois compétences clés :

  • D’abord, designer les règles du jeu : définir les métriques, les règles de coopération et les espaces de régulation collective, plutôt que dicter le “comment”.

  • Ensuite, orchestrer sans diriger, en facilitant les synergies, en arbitrant les tensions et en maintenant une vision d’ensemble.

  • Enfin, anticiper les angles morts de l’autonomie : repérer les voies d’épuisement, les compétences qui se perdent et les dérives.
• Il entrainera le développement d’une culture organisationnelle forte et une montée en compétence des managers, afin de structurer, réguler et soutenir durablement les dynamiques à l’œuvre dans cette myriade d’entités.

15 more annotations...

• rushed state legislation is creating an “increasingly unworkable regulatory environment” for employers.
• and found that only two automated employment decision tool (AEDT) complaints were filed during the audit’s two-year scope.
• show what happens “when legislation gets ahead of itself” and warned Congress against similar missteps at the federal level.
• Kelley made the case that the U.S. already has a “well-established technology-neutral legal framework fully capable of addressing most AI-related misconduct,”
• he has never encountered employers using AI to interfere with union activity or suppress organizing efforts
• used an AI‑driven “heat map” to identify stores at risk of union activity.
• helpline staff they would be laid off and replaced by an AI chatbot just weeks after those workers voted to unionize.
• he bot was shut down “after it provided harmful advice to people seeking help.”
• ederal statistical agencies are flying blind when it comes to measuring AI’s real impact on American workers. The problem: Current labor statistics count jobs rather than measuring how work is done.
• “AI now handles much of that, freeing me to focus on research design, collaboration and communication,” she said. “The occupation ‘economist’ still exists, but the work has changed. This is precisely what we need to measure.”
• here was no occupation that AI could fully handle alone. Yet, the U.S. Bureau of Labor Statistics/U.S. Census Current Population Survey includes no questions about AI use or task changes.

• “Good policy requires good data, and we currently don’t have it,” said Sharfuddin, in closing.

   
• 90% of U.S. employers use algorithmic management tools and that two-thirds of workers report such tools are used in their workplaces “across earnings in different jobs.”
• Goldman described how AI enables more invasive surveillance than any human manager could conduct.
• “Time-on-task measures can pressure workers to work constantly and make it difficult to take necessary breaks
• In hiring, Goldman said, one-third to “the vast majority” of employers use AI tools for recruitment and screening. She said these systems often lack transparency, meaning “workers often do not understand how decisions affecting their livelihoods are being made. This can leave them without meaningful recourse to challenge the decision or to correct inaccurate data that is being used.”
• “On the same day at the same hour in the same hospital, two different gig nurses can be paid different amounts by the same app.”

• Goldman argued these harms are “the product of choices, not an inevitability” and called for federal guardrails, including transparency requirements, mandatory human involvement in significant employment decisions, bias testing and a meaningful role for workers and unions in AI implementation.

   
• “They need to be able to hire technologists and really understand the data so they can get into the black box and do the kind of systemic work that needs to happen,” Goldman said.

• robust self-governance” through internal controls works better than “blanket restrictions.”

   
• arguing that when technology makes a resource more efficient, it lowers costs and expands demand so much that total consumption increases. “Every automation wave that we have had, especially since 1750, has followed this pattern,” he said, predicting AI will likely increase total jobs rather than decrease them.
• DeSaulnier noted that “history teaches us that workers don’t benefit at the same rate as investors” and urged the committee to “get ahead of that,” rather than waiting for workers to suffer.

20 more annotations...

• For the past couple of years, Gartner research has suggested that organizations have a management problem.
• 35% and 27% reported being satisfied with the effectiveness of mid-level and frontline managers, respectively.
• 38% reported satisfaction with the quality of their manager,
• When managers don’t deliver, the costs can add up, often manifesting as a drag on productivity, performance, and morale.
• 0% acknowledged that managers are overwhelmed by the expanding scope of their responsibilities—
• one in four managers would prefer not to be people managers, up from one in five just two years ago
• Many of these managers found themselves in the role without knowing much about it beforehand, with less than a third reporting exposure to simulations, mentorship, or opportunities to gauge whether they were suited to becoming a manager.
• high individual performance is a top factor for management selection, while only 22% prioritize high scores on pre-management selection assessments.
• When a manager doesn’t actually want to be a manager, they lack authentic commitment to and engagement in the role
• just 16% said their organization’s manager-selection process ensures the right people work in people manager roles.
• HR needs to identify that population before they’re selected for management roles so that management is an active choice, not a passive experience.

• First, it prepares new-to-role managers for the realities of the job, demystifying it and exposing the increased weight, complexity, and intensity of work. Second, it helps potential managers make an informed decision on whether to continue with the management-selection process.

   
• simulations and experience with the more opaque and challenging activities of management, including workload prioritization decisions, difficult performance and engagement conversations, and performance calibration sessions.
• It’s also important that managerial candidates be provided with non-punitive, non-permanent off-ramp opportunities during the selection process.
• Observing where and how suspected reluctant managers are struggling helps establish whether their reluctance is caused by deep-rooted factors, such as a true dislike of the role or disinterest in the necessary work, or more surmountable sources, such as a lack of confidence or a feeling of overwhelm in the role.

• Assessing Manager Reluctance

   

  Is the manager’s reluctance driven by deep-rooted factors or more-surmountable ones?

   

  Signs of addressable reluctance	Signs of entrenched reluctance
  Overwhelmed by the amount of work	Overwhelmed by the type of work
  Frustrated with temporary challenges	Frustrated with challenges inherent to the role
  Confident in their ability to perform certain critical tasks	Strongly dislikes performing certain critical tasks
  Averse to certain activities because they find them dull	Averse to certain activities because they find them prohibitively uncomfortable
• When reluctance is deeply rooted, interventions are less likely to have the intended impact, and the focus should turn to the best ways to off-ramp reluctant managers.
• When a lack of confidence in the role is driving manager reluctance, leaders should identify management-adjacent skills that the reluctant manager uses in their personal lives that can be transferred to how they manage at work to build their confidence for the job.
• To address the mental load of management, HR can teach habit-building practices that drive behavior change and decrease the cognitive load of management and perceived difficulty of tasks
• By proactively identifying and addressing the presence of reluctant managers, HR leaders have a pivotal opportunity to significantly improve leaders’ confidence in the management corps

18 more annotations...