Digital Article / AI and Machine Learning
Overcoming the Organizational Barriers to AI Adoption
People, processes, and politics determine whether AI creates value.
By Jin Li, Feng Zhu, and Pascal Hua
Published on HBR.org / November 11, 2025 / Reprint H08YUXPeople, processes, and politics determine whether AI creates value
This document is made available to you with compliments of Harvard Business Review for your personal use only. Copying or posting is an infringement of copyright. Please contact
customerservice@harvardbusiness.org or 800-988-0886 for additional copies.
HBR / Digital Article / Overcoming the Organizational Barriers to AI Adoption
The hype around AI has dominated conversations among top executives. Yet most organizations are still struggling to generate meaningful returns from their AI initiatives.
To explore these challenges in greater depth, we conducted extensive research, combining a survey of over 100 C-suite executives with more than two dozen interviews across industries. The survey revealed that 45% of executives found the ROI of AI adoption to be below expectations, while only 10% reported results exceeding expectations.
It also highlighted that the most significant barriers are organizational, rather than technical. Building on these findings, we identify a set of interlocking obstacles rooted in three areas: people, processes, and politics.
In this article, we explore these interlocking obstacles and look at how companies are addressing them.
Cultivating People’s Readiness for AI
When thinking about employees’ readiness for AI, our research identified three problems: uncertainty, fear of replacement, and the self-image problem.
The Uncertainty Problem: “What will this AI actually do?”
Slack’s 2024 global survey of more than 17,000 office workers found that 61% of employees had spent less than five hours learning about AI and 30% had received no training at all. In the absence of knowledge, opinions can be polarized. Some employees dismiss AI as mere hype, while others assume it can do everything.
Companies’ uncertainty about AI extends beyond technical capability. One audit firm, for example, identified AI opportunities across its workflow, but both clients and auditors resisted, citing regulatory risk. In the end, the firm abandoned many of its AI-based approaches.
To address these concerns, firms need to embed AI governance into their daily work and make it intuitive for every employee. Effective governance not only safeguards against unintended consequences but also helps demystify AI. As an example, in 2018 DBS Bank introduced the PURE framework—Purposeful, Unsurprising, Respectful, and Explainable—to evaluate every AI use case. Instead of relying on lengthy policy documents, employees are guided by four simple questions: Is the use purposeful and meaningful? Will the results surprise customers? Does it respect customers and their data? Can the outputs be explained? This approach reduces uncertainty while ensuring responsible use. DBS also established a Responsible Data Use Committee to review projects that do not meet PURE requirements.
With an easy-to-grasp framework and clear human oversight, the bank empowered employees at all levels to innovate responsibly. By 2023, AI had already generated $274 million in value for DBS.
Fear of Replacement: “Will I keep my job?”
When employees suspect they are training a system that will replace them, they comply minimally. They drag their feet when asked to “label data” or “teach the model.” This “training trap” slows down the adoption of AI in service, retail, and manufacturing firms.
Companies can counter this by sharing the upside—offering training royalties for data and labeling work, productivity bonuses tied to realized gains, and career guarantees that channel efficiency gains into reskilling rather than layoffs.
Because many of these benefits are based on future promises that are easy to break, firms must make promises credible and easy to verify. One e-commerce company, for instance, pledged to increase total labor spending by 1% annually to demonstrate its commitment to investing in employees. This 1% number is easy to check and hard to manipulate, so it helped build trust with workers. It also gave workers formal seats on the AI steering committee and greater influence over personnel decisions. These enhanced power of the workers further reinforce trust.
Another form of resistance is fault-finding: holding AI to much higher standards than humans. At a leading insurance company, employees’ fault-finding missions led to demand for unrealistically high levels of accuracy from AI systems, which in turn slowed deployment and drove up investment costs. Independent studies and external audits comparing AI and human outputs helped restore realism.
Finally, replacement fears ease when AI fuels growth rather than contraction. If technology expands the business, efficiency gains feel like opportunity, not threat.
The Self-Image Problem: “Will I appear competent?”
Fear of status loss can be even more powerful than fear of job loss. We’ve observed engineers who quietly use AI tools but conceal it to avoid appearing less skilled. Many worry that admitting to using AI could make them seem lazy, incompetent, or even dishonest. Similar image concerns result in radiologists ignoring AI recommendations to protect professional pride.
One financial services firm flipped this stigma by launching an “AI Masters” program that fast-tracks employees who demonstrate exceptional AI skills, regardless of their seniority. This celebrates the
proficiency with AI as sophistication and forward-thinking, not laziness or incompetence. By broadcasting this message, organizations can create positive incentives for employees to embrace AI.
Equally important is designing AI use for professional dignity.
Companies can position AI as a tool that presents facts without judgment, while leaving final conclusions to professionals. This framing reinforces expertise rather than undermining it. Some firms have created private “second-opinion consoles” where employees can consult AI without fear of embarrassment or reputational risk.
Processes: Redesigning Workflows
AI adoption often falters when organizations treat it as a simple overlay on existing processes. True transformation demands systematic change at three levels: individual workflow (nodes), cross-functional connections (edges), and system-wide coordination (networks).
The Node Level: Transforming How Individuals Work
A consulting firm’s legal team initially used AI like a spell-check tool—running it at the very end of traditional reviews. The approach produced negligible benefits, as AI was only 100% accurate for 40% of error types. By restructuring the workflow so that AI conducted the first pass —checking only error types it handled best—lawyers could focus solely on the remaining ones. This redesign demonstrated how rethinking workflows unlocks AI’s value.
Some firms accelerated the change by setting “mission impossible” goals that force teams to abandon old habits and discover new ways of working. One Boston startup, for instance, faced resistance to using AI in document preparation. To break through, it required that documents
—previously completed in a week—be finished within a single day. The extreme time pressure left employees no choice but to integrate AI from the start and redesign their processes around it.
The Edge Level: Redesigning Connections
The edge level focuses on how improved local judgment and data can transform inter-departmental processes and decision-making flows.
At a Japanese cosmetics company, beauty advisors in stores once supplied untrusted, anecdotal feedback. Generative AI helped them analyze customer conversations and traffic patterns, giving structured insights. Headquarters, now confident in the data, built a two-way loop: Campaigns could launch faster and be tweaked in real time based on credible field intelligence. The edge between local operations and central planning became a responsive circuit rather than a one-way
command.
The Network Level: Orchestrating System-Wide Impact
To generate real business impact from AI, companies must consider the network level—how improvements across multiple nodes and edges interact within the broader system. Without this perspective, AI can simply shift bottlenecks from one part of the network to another, limiting overall performance gains. This phenomenon is common because many organizations concentrate their gen AI efforts in a few high-impact areas—such as marketing, customer service, or software development—while overlooking the interdependence across business units.
A major car manufacturer discovered this when it adopted generative AI to boost productivity in automotive software development (enhancing one set of nodes), enabling faster design iterations, code generation, and feature testing. Yet the overall vehicle production network showed little improvement, as hardware manufacturing became the primary bottleneck. The enhanced software development nodes were now waiting on unchanged hardware nodes, and the edges connecting them couldn’t handle the increased pace of software output.
Addressing such network-level challenges requires coordinated action across all nodes and edges. Organizations should begin by mapping the entire network topology—understanding how workflows between teams and identifying potential bottlenecks. AI adoption should be synchronized across interconnected nodes so that capacity improvements are matched throughout the network.
Politics: Navigating Power and Influence
AI shapes who gains and who loses inside organizations. The resulting politics—over data, hierarchy, and accountability—often prove more formidable than technical issues. Successful AI adoption often requires redesigning governance structures, adjusting incentive mechanisms, and, in some cases, relying on senior leadership to broker agreements and remove barriers. Here are three specific problems we observed in our research:
Resource Hoarding
Organizations quickly discover that AI’s hunger for data and knowledge collides head-on with deeply ingrained competitive instincts. At a large Chinese IT firm, researchers uncovered that programmers were 16–18% less likely to recommend AI access to their own teammates, effectively hoarding knowledge to preserve their personal edge.
Across business units, larger, more successful divisions that own sophisticated AI models and valuable datasets often see little incentive to share them with smaller units that could benefit most. Sharing can feel like enabling potential internal competitors while diluting their own performance metrics. In the Deloitte-HKU survey we conducted, the C-suite level participants identify “siloed departments preventing cooperation” as the top barrier for AI adoption.
DBS Bank confronted this resistance by designing an incentive structure that rewarded units for converting proprietary datasets into reusable assets on the central platform. A key metric tracked the
percentage of each unit’s use-case-specific datasets that had been transformed into shareable resources. This approach breaks down silos by motivating both large and small units to contribute high-quality, accessible data.
Hierarchy Disruption
AI unsettles the traditional hierarchy built on two pillars: experience and headcount. The first weakens when junior employees armed with AI outperform seasoned veterans. In one software firm, programmers with only two years’ experience began producing more and cleaner code than colleagues with five years’ tenure. Juniors felt that they were doing more for less.
Some companies responded by expanding competency models to explicitly include AI mastery and by shortening promotion ladders. When advancement cycles shrink from five years to one or two, and mastery of new tools is rewarded, young employees see immediate payoff for learning.
The second pillar, power through headcounts and control over resources, creates even stronger resistance. Managers are the gatekeepers of AI adoption, yet their authority often depends on the size of their teams. When efficiency threatens to shrink those teams, their self-interest can quietly derail otherwise valuable AI initiatives. In one translation department, leaders hesitated to automate because doing so would shrink their headcount, bonuses, and prestige.
OPPO, the smartphone maker, tackled this by staging an AI tournament where every employee had equal access to tools, and results were ranked by department. Suddenly, managers had to champion AI adoption or risk public embarrassment if their teams lagged. The contest reframed success: status no longer came from managing large teams but from enabling them to achieve more with AI.
Accountability Attribution
AI also disrupts the traditional balance of blame and discretion within organizations. Its precision turns fuzzy responsibility into hard data—and that can create new political friction.
At Dingdong Maicai, a Chinese grocery e-commerce company, AI systems began tracing every customer complaint back to the exact department at fault. When a customer received spoiled fruit,
algorithms could pinpoint whether procurement bought poorly, storage mishandled goods, or delivery caused the damage. What had once been shared uncertainty became explicit accountability.
With this change, departments that had long operated under ambiguity now found themselves publicly exposed. The binary nature of algorithmic judgment—assigning full responsibility to one side—ignored the grey areas of real-world operations. This led to escalated disputes and complaints from department heads.
The lesson is that perfect accountability can undermine organizational harmony. Dingdong eventually changed its system of attribution, allowing final judgment to humans. The goal was not to reject transparency but to buffer it with trust. Effective AI adoption requires knowing when precision helps performance—and when it merely fuels internal politics.
Pulling Multiple Levers for AI-Driven Transformation
A professional services firm with 2,200 practitioners—primarily software developers and product managers—began piloting GenAI initiatives in mid-2023. Within weeks, individual productivity rose
by 30–40%, yet by mid-2024, overall performance—measured by productivity and time-to-delivery—remained flat.
Several factors explained the gap. Developers lacked incentives to boost output, fearing efficiency gains might trigger layoffs. The flood of new AI tools created inconsistent practices across teams, disrupting established standards and complicating project management. Meanwhile, junior developers often outperformed senior ones, but work assignments and recognition still followed traditional hierarchies.
To address these challenges, the firm pulled levers across people, processes, and politics. On the people front, it redefined its competency model to explicitly reward AI proficiency, making expertise visible across the organization and turning mastery into a source of pride. To counter fears of replacement, the compensation structure was overhauled: Base salaries were reduced to 80%, while performance-based incentives of up to 40% were added, directly linking efficiency gains to individual rewards.
The process dimension was overhauled to embed AI throughout the workflow. Developers became data and process stewards, responsible for following standardized data definitions, coding practices, and AI protocols while participating in training to strengthen process consistency. A unified end-to-end framework harmonized AI integration across development stages, with updated SOPs incorporating AI-augmented steps for easier training and compliance. At the organizational level, a centralized governance model with defined checkpoints and Business Process Stewards ensured alignment across data, AI, and workflows.
Political barriers were confronted head-on. Job grades expanded from six to 14, with biannual reviews enabling rapid promotion or demotion. This system rewarded AI adopters with greater responsibility and influence, realigning incentives that once favored tenure over capability.
By mid-2025, these changes began to pay off. Productivity rose by 22%, enabling a 10% price cut that boosted sales by 20%. Labor costs grew by 5% as the firm reinvested in its workforce, reinforcing its commitment to employees. Overall profitability improved by 3%, demonstrating that AI-driven transformation translated into tangible business value. Building on this foundation, the firm expanded into markets that had previously been too price-competitive to enter. AI-supported development also shortened the learning curve for new programming languages, enabling the company to broaden its offerings.
This case shows that true AI transformation goes beyond technology. By aligning incentives, redesigning processes, and reconfiguring organizational power, the firm turned AI adoption into lasting business value.
…
Ultimately, the challenge is not adopting AI but evolving alongside
it. The true advantage lies in building an organization that can fully
harness AI’s power. Firms that see it merely as a technical upgrade will
inevitably fall short.
This article was originally published online on November 11, 2025.
Jin Li is Zhang Yonghong professor in economics and strategy, JL director of the Centre for AI, Management and Organization (CAMO), and area head of management and strategy at Hong Kong
University Business School.
Feng Zhu is the MBA Class of 1958 Professor of Business Administration at Harvard Business School. He is a coauthor of Smart Rivals: How Innovative Companies Play Games That Tech
Giants Can’t Win (Harvard Business Review Press, 2024).
Pascal Hua is national managing partner of technology and PH transformation at Deloitte China.
Copyright © 2025 Harvard Business School Publishing. All rights reserved.
Overcoming the Organizational Barriers to AI Adoption
Plan de cours — 3 heures
| Horaire | Activité | § |
|---|---|---|
| 0:00–0:15 | Warm-up : "Is your company ready for AI?" | — |
| 0:15–0:55 | Questions de compréhension (individuel → binômes → correction) | §1 |
| 0:55–1:30 | Vocabulaire : définitions en contexte, exercice en binômes | §2 |
| 1:30–1:35 | Pause | — |
| 1:35–2:20 | Points de grammaire : identification + exercices de réécriture | §3 |
| 2:20–2:45 | Débat : pertinence du document pour un data scientist | §4 |
| 2:45–3:00 | Synthèse orale : résumé en 1 min par étudiant | — |
Questions de compréhension
12 questions en trois niveaux. Tier A — repérage factuel ; Tier B — inférence ; Tier C — esprit critique. Cliquer sur Show Answer pour révéler la réponse.
Tier A — Factual Retrieval
What percentage of C-suite executives found the ROI of AI adoption below expectations?
Show Answer
45% reported ROI below expectations; only 10% said results exceeded them.
p. 2 — "45% of executives found the ROI of AI adoption to be below expectations…"What do the four letters of DBS Bank's PURE framework stand for?
Show Answer
Purposeful · Unsurprising · Respectful · Explainable — four questions applied to every AI use case.
p. 2 — "DBS Bank introduced the PURE framework…"According to Slack's 2024 survey, how much time had most employees spent learning about AI?
Show Answer
61% had spent fewer than five hours; 30% had received no training at all.
p. 2 — "Slack's 2024 global survey of more than 17,000 office workers…"Name the three levels at which AI adoption must transform processes.
Show Answer
Node (individual workflow) · Edge (cross-functional connections) · Network (system-wide coordination).
p. 4–6 — "True transformation demands systematic change at three levels…"Tier B — Inference & Interpretation
Why do the authors call the barriers to AI adoption "interlocking"? What does this imply for organisations?
Show Answer
The barriers reinforce each other: fear of replacement (people) resists process change, which fuels political struggles over data. Acting on only one dimension fails. The professional services firm case (p. 9–11) shows that productivity gains were neutralised until all three dimensions were tackled simultaneously.
p. 2 — "…a set of interlocking obstacles rooted in three areas: people, processes, and politics."What is the "training trap" and why is it hard for organisations to solve?
Show Answer
Employees asked to label data or "teach the model" believe they are building their own replacement, so they comply minimally. The difficulty: those whose expertise is most needed have the strongest incentive to withhold it, and management promises of job security are structurally hard to make credible.
p. 3 — "When employees suspect they are training a system that will replace them…"In the car manufacturer case, why did AI adoption in software development not improve overall vehicle production?
Show Answer
Accelerating software (nodes) merely shifted the constraint to hardware manufacturing, which became the new bottleneck. The edges connecting the two could not absorb the faster pace. Improving one node without considering system interdependencies relocates, rather than eliminates, the bottleneck.
p. 6 — "…hardware manufacturing became the primary bottleneck."How did Dingdong Maicai's AI system change internal political dynamics, and how did the company respond?
Show Answer
The AI pinpointed the exact department responsible for each complaint, turning formerly ambiguous blame into hard binary judgment. This triggered inter-departmental disputes. Dingdong reintroduced human oversight for final attribution — "not to reject transparency but to buffer it with trust."
p. 8–9 — "Dingdong eventually changed its system of attribution, allowing final judgment to humans."Tier C — Critical Evaluation
"The challenge is not adopting AI but evolving alongside it." Do you agree — from a data scientist's perspective?
Show Answer
In favour: technical barriers have largely been overcome; the real bottleneck is now organisational adaptability. Nuance: for cutting-edge research domains (LLM alignment, causal inference), technical challenges remain central. A strong answer positions the student personally: as a future data scientist, they are the technical architect but must also communicate value upward — exactly the gap this article addresses.
p. 11 — "Ultimately, the challenge is not adopting AI but evolving alongside it."The firm reduced base salaries to 80% and added performance bonuses of up to 40%. Evaluate risks and benefits.
Show Answer
Benefits: aligns incentives with AI adoption; rewards early adopters. Risks: increases income precarity; disadvantages roles less amenable to AI augmentation. European angle: in France, unilateral modification of salary structure requires employee consent (Code du travail) — a regulatory gap the article ignores entirely.
p. 10 — "Base salaries were reduced to 80%…"The article mentions no European regulatory context. How do GDPR and the EU AI Act change the picture?
Show Answer
GDPR constrains centralised data platforms (cf. DBS model) and Article 22 restricts automated individual decision-making — directly relevant to the Dingdong case. The EU AI Act creates tiered compliance obligations for high-risk AI (hiring, credit, critical infrastructure). European data scientists must navigate three layers simultaneously: technical, organisational, and legal.
p. 2 — audit firm cites "regulatory risk" with no European detail.The authors are affiliated with HBS, HKU, and Deloitte China. How might this shape their analysis?
Show Answer
Geographic bias: all case studies are Asia-Pacific. Managerial bias: employee interests appear as "problems to manage." Consulting interest: Deloitte's involvement may favour complex, paid-for transformations over simpler approaches. Lesson: read any text — including HBR — with awareness of authorial positionality.
p. 11 — author bios.Vocabulary List
31 termes sélectionnés pour leur pertinence académique et professionnelle.
| # | Terme / Expression | Définition (anglais) | Cat. | Page |
|---|---|---|---|---|
| 1 | ROI | Financial benefit from an investment relative to its cost; in AI, whether deployment generates sufficient business value. | Tech | p.2 |
| 2 | AI governance | Policies and oversight mechanisms regulating how AI is developed, deployed, and monitored to ensure responsible use. | Tech | p.2 |
| 3 | Gen AI / Generative AI | AI models that generate new content (text, code, images) by learning from large datasets; e.g. LLMs. | Tech | p.5–6 |
| 4 | data labeling | Annotating raw data with tags so ML models can learn from it during supervised training. | Tech | p.3 |
| 5 | workflow (node / edge / network) | Structured task sequence. In systems thinking: a node = individual unit; edge = connection; network = full system. | Tech | p.4–6 |
| 6 | bottleneck | Point where work flow is restricted, causing downstream delays. Removing it is central to operational efficiency. | Tech | p.6 |
| 7 | algorithmic accountability | Attribution of decisions and consequences to an algorithm rather than a human — raising fairness and transparency issues. | Tech | p.8–9 |
| 8 | SOP | Standard Operating Procedure: documented step-by-step instructions for a routine activity, updated to embed AI-augmented steps. | Tech | p.10 |
| 9 | to pilot | To run a small-scale test of a system before full deployment, to validate assumptions and identify problems. | Tech | p.9 |
| 10 | time-to-delivery | Total elapsed time from project start to delivery. A key KPI in software development. | Tech | p.9 |
| 11 | C-suite | The most senior executives in a company whose titles begin with "Chief" (CEO, CFO, CTO, CIO…). The "C" stands for Chief. They constitute the highest level of corporate decision-making. | Biz | p.1, 7 |
| 12 | hype | Exaggerated promotion of something. In AI: overenthusiastic expectations that often precede realistic assessment. | Biz | p.1 |
| 13 | to generate meaningful returns | To produce gains significant enough to justify an investment — measurable and substantial, not merely symbolic. | Biz | p.1 |
| 14 | interlocking obstacles | Barriers that reinforce each other; overcoming one alone is insufficient — they must be addressed together. | Biz | p.2 |
| 15 | to demystify | To make something complex or threatening easier to understand, reducing fear or resistance. | Biz | p.2 |
| 16 | training trap | Employees resist contributing to AI training because they believe they are helping build their own replacement. | Biz | p.3 |
| 17 | reskilling | Training employees in new skills after their existing role has been automated or significantly changed. | Biz | p.3 |
| 18 | fault-finding | Deliberately searching for errors in a system to justify resistance; holding AI to unrealistically high standards. | Biz | p.3 |
| 19 | to pull levers | To activate available tools, policies, or resources to influence an outcome. "Pulling multiple levers" = coordinated multi-dimensional action. | Biz | p.9 |
| 20 | to flip a stigma | To reverse a negative social perception into a positive one. The firm reframes AI use as expertise, not laziness. | Biz | p.4 |
| 21 | mission impossible (goal) | An extreme objective forcing employees to abandon habitual workflows and discover entirely new approaches. | Biz | p.5 |
| 22 | two-way loop | A feedback mechanism where information flows in both directions, enabling continuous adjustment rather than one-directional command. | Biz | p.5 |
| 23 | credible commitment | A promise structured through verifiable metrics or shared governance so it is genuinely believable and hard to retract. | Biz | p.3 |
| 24 | self-image problem | Fear that using AI will make employees appear less competent or original in the eyes of colleagues and managers. | Pol/HR | p.4 |
| 25 | resource hoarding | Deliberate withholding of data or knowledge to preserve competitive advantage within the organisation. | Pol/HR | p.7 |
| 26 | siloed departments | Teams that operate independently with minimal communication or data sharing — one of the top cited AI adoption barriers. | Pol/HR | p.7 |
| 27 | hierarchy disruption | Unsettling of traditional power structures based on seniority when AI enables juniors to outperform seniors. | Pol/HR | p.7 |
| 28 | accountability attribution | Assigning responsibility for an outcome to a specific entity. AI's precision turns ambiguous blame into hard, traceable data. | Pol/HR | p.8 |
| 29 | to derail | To cause a project to fail or be abandoned, often covertly. Managers threatened by headcount loss may quietly derail AI initiatives. | Pol/HR | p.8 |
| 30 | to drag one's feet | Idiom: to act slowly or reluctantly, deliberately delaying progress. | Pol/HR | p.3 |
| 31 | headcount | Compound noun: total number of employees in a team or organisation; a measure of a manager's power. Shrinking it signals loss of authority. | Pol/HR | p.7–8 |
Key Grammar Structures
Six structures récurrentes dans l'article, illustrées par des exemples tirés du texte. Chaque exercice comporte un corrigé enseignant (🔒).
Passive Voice — Reporting Results & Findings
Form: to be + past participle (am/is/are/was/were/has been/had been/should be…). Shifts focus from agent to result; foregrounds data over individuals.
Exercise: Transform into the passive (keep the same tense).
1. "The firm abandoned many AI-based approaches."
2. "The committee reviews projects that fail PURE."
3. "The company has generated $274M in value."
Teacher's Answer Key
- "Many AI-based approaches were abandoned (by the firm)." — simple past passive; agent optional.
- "Projects that fail PURE are reviewed (by the committee)." — present simple passive.
- "$274M in value has already been generated." — present perfect passive; "already" sits between the two auxiliaries.
Erreur fréquente : "was been generated" — rappeler que has been est l'auxiliaire du passif au present perfect et ne se combine jamais avec was.
Modal Verbs — Possibility, Necessity & Recommendation
True modals: can, could, may, might, will, would, shall, should, must, ought to. Invariables — jamais de -s à la 3e personne. Suivis d'un infinitif sans to.
⚠ Semi-modal : need to / needs to — se conjugue comme un verbe ordinaire ("the firm needs to embed" — avec -s). Ce n'est pas un vrai modal.
| Modal | French | Type | Example (text) |
|---|---|---|---|
| might | pourrait | Possibility — weak | — |
| may | peut / est susceptible de | Possibility — neutral | "resistance may come to pass" |
| can | peut / est capable de | Capacity / possibility | "AI can simply shift bottlenecks" p.6 |
| will | va / fera | Certainty / future | "Firms… will inevitably fall short." p.11 |
| should | devrait | Recommendation | "Organizations should begin…" p.6 |
| must | doit (interne) | Strong obligation | "firms must make promises credible" p.3 |
| need to/needs to | a besoin de / doit | ⚠ Semi-modal | "Firms need to embed AI governance" p.2 |
Exercise: Replace the modal in "Firms should redesign workflows" with must, could, might. Explain in French how the meaning changes.
Teacher's Answer Key
- must — obligation forte et interne : les entreprises n'ont pas le choix. Plus péremptoire que should.
- could — possibilité conditionnelle ou suggestion prudente : c'est une option envisageable, sans obligation. Ton nuancé, non prescriptif.
- might — éventualité faible : la probabilité est incertaine. Hypothèse prudente, souvent utilisée dans des prédictions non garanties.
Point discussion : Pourquoi les auteurs préfèrent-ils should à must ? (Should est plus respectueux de l'autonomie du lecteur — ton consultatif approprié pour HBR.)
Conditional Sentences — Hypothetical Scenarios
Type 1 (réel / probable) : If + present → will/can + infinitive. L'article utilise aussi when et without comme conditionnelles implicites.
Exercise: Transformer le dernier exemple en Type 2 (hypothétique) : "If efficiency threatened to shrink those teams, their self-interest _____ quietly derail…"
Teacher's Answer Key
"If efficiency threatened… their self-interest could / would quietly derail…"
Type 2 = simple past dans la subordonnée + would/could/might + infinitif dans la principale. Le glissement de Type 1 à Type 2 réduit la probabilité implicite : les auteurs traitent ce scénario comme probable (Type 1) ; le Type 2 n'en ferait qu'une hypothèse théorique.
Noun Phrases & Compound Nouns
Noun Phrase (NP) : unité syntaxique organisée autour d'un head noun. Elle peut inclure des pré-modifieurs (déterminants, adjectifs) et des post-modifieurs (groupes prépositionnels, relatives). La NP fonctionne comme un nom dans la phrase.
Ex : "a set of interlocking obstacles rooted in three areas"
Det: a · Head noun: set · Post-mod.: of interlocking obstacles rooted…
Exemples du texte (NPs surlignées en rouge) :
⚠ unsettles = verbe (3e pers. sg.) — jamais surligné comme NP.
NP vs. Compound Noun — Differences :
| Feature | Noun Phrase | Compound Noun |
|---|---|---|
| Nature | Unité syntaxique (structure) | Unité lexicale (un seul concept) |
| Length | Peut être très longue | Typiquement 2–3 mots |
| Form | Toujours un head noun + modifieurs | 1 mot, avec trait d'union, ou 2 mots séparés |
| Stress | Accent sur le dernier mot : interlocking obSTACLES | Accent sur le 1er élément : HEADcount, WORKflow |
| Text examples | "a set of interlocking obstacles," "the traditional hierarchy" | headcount, workflow, fault-finding, time-to-delivery |
Exercise : Identify all NPs in : "The growing use of automated data analytics tools has fundamentally changed business decision-making processes."
Teacher's Answer Key
- "The growing use of automated data analytics tools" — head: use ; post-mod: of automated data analytics tools
- "automated data analytics tools" — head: tools ; pre-mod: automated data analytics
- "business decision-making processes" — head: processes ; pre-mod: business decision-making
Compound nouns : data analytics (noun+noun), decision-making (noun+gerund, avec trait d'union). Test phonétique : faire prononcer à voix haute pour identifier l'accent tonique.
Present Perfect vs. Simple Past
Simple past : événement complété à un moment précis (ancré par une date ou expression temporelle). Present perfect (has/have + p.p.) : action passée dont le résultat est encore pertinent maintenant — le moment exact est secondaire. Past perfect (had + p.p.) : action complétée avant un autre moment passé.
Timeline des exemples de l'article :
Exercise : Pourquoi les auteurs utilisent-ils systématiquement le simple past plutôt que le present perfect ? Que révèle ce choix sur leur posture d'auteurs ?
Teacher's Answer Key
Chaque événement est ancré à une date précise (2018, mid-2023, mid-2024, mid-2025) — condition nécessaire et suffisante pour le simple past. Le present perfect serait approprié si la date était inconnue ou non pertinente. Ce choix signal :
- Des faits vérifiables et achevés — renforce la crédibilité.
- Une narration chronologique (cause → conséquence) — plus persuasive pour un lectorat managérial.
- Une distinction implicite entre evidence (simple past, daté) et recommendations (present simple ou modal, atemporel).
Concessive Clauses — Acknowledging Counterarguments
Définition : une proposition concessive introduit un contraste ou un résultat inattendu — elle reconnaît un fait tout en affirmant quelque chose de contraire ou surprenant. Indispensable pour l'écriture académique nuancée.
| Connecteur | Français | Comportement grammatical & nuance |
|---|---|---|
| yet | pourtant / cependant | Adverbe — débute une nouvelle phrase ou clause indépendante après un point. Contraste rhétorique fort ; ton formel. Ne peut pas relier deux propositions avec une simple virgule. |
| although | bien que / même si | Conjonction de subordination — peut être en tête ou au milieu de phrase. Pas de subjonctif en anglais (contrairement au français). Concession neutre et factuelle. |
| while | tandis que / alors que | Conjonction de subordination — souligne la simultanéité de deux états contrastés. Aussi utilisé au sens temporel ("pendant que") — le contexte désambiguïse. |
| however | cependant / néanmoins | Adverbe — en début de phrase, après un point-virgule, ou entre virgules en milieu de phrase. Jamais seul entre deux propositions avec une virgule (comma splice). Registre académique formel. |
| even though | même si / quand bien même | Plus fort que although : insiste sur le fait que le contraste tient malgré pleine conscience du fait concédé. Emphase sur l'irréductibilité de la conclusion. |
| despite / in spite of | malgré | Préposition — suivie d'un groupe nominal ou gérondif, jamais d'une proposition sujet+verbe. ✓ "despite the resistance" · ✓ "despite resisting" · ✗ "despite they resisted". |
| but | mais | Conjonction de coordination — relie deux propositions indépendantes de poids égal. Le connecteur le plus neutre et polyvalent ; moins formel que les autres. |
Exercise : Écrire deux phrases — une avec yet, une avec although — pour décrire : un modèle très précis que l'entreprise n'utilise pas.
Teacher's Answer Key
- "The model achieves 97% accuracy. Yet the company has refused to integrate it into its decision-making." — point avant yet ; ton ironique.
- "Although the model is highly accurate, the company has chosen not to deploy it, citing employee resistance." — subordonnée en tête ; ton analytique et neutre.
Erreur classique : "Although… but…" — interdit en anglais. Un seul connecteur concessif par phrase pour marquer le même contraste. (Différent du français : bien que… mais… est courant à l'oral.)
About This Document
L'article est délibérément non technique. Le débat ci-dessous permet aux étudiants d'évaluer sa pertinence pour leur propre profil.
✔ Pourquoi il est adapté
- Conscience professionnelle : comprendre pourquoi de bons modèles ne sont jamais déployés.
- Anglais en entreprise : registre des présentations C-suite et rapports de conseil.
- Écriture fondée sur des preuves : les données (45 %, 30 %…) modélisent la rigueur académique.
- Idiomes business : "training trap," "pulling levers," "race to the bottom."
- Études de cas concrètes : DBS Bank, OPPO, Dingdong Maicai.
- Pont interdisciplinaire : Personnes / Processus / Politique complète la formation technique.
✗ Limites potentielles
- Absence de technique : à cadrer explicitement comme complément, non substitut.
- Perspective managériale : écrit pour les employeurs, pas pour les étudiants eux-mêmes.
- Biais Asie-Pacifique : RGPD et AI Act absents — contexte européen à apporter par l'enseignant.
- Idiomes opaques : nécessitent un pré-enseignement pour des lecteurs non natifs.
- Densité conceptuelle : 11 pages compactes, stratégies de lecture guidée conseillées.