# oliver-wyman-forum-ai-agenda-reality-check-2026-10-15

## Veille

Report **The AI Agenda 2026 — The reality check** from the **Oliver Wyman Forum** (the **Oliver Wyman** consultancy, **Marsh** group), a **91-page** PDF dated **October 15, 2026** (file metadata), prefaced by **John Romeo** (CEO) and **Ana Kreacic** (COO). Empirical base: two waves of the **Global Voices** survey (April, **N = 12,883**; August, **N = 12,837**; 15 markets including France), the **Oliver Wyman Forum × NYSE CEO Agenda 2026** survey (**415 CEOs**), parallel CEO / tech-leader surveys (**N = 200**), a **Sovereign AI Index** (top 100 economies), and three agent-purchase simulations. Overall thesis: *"AI is everywhere in our company except the P&L"*; AI capability is arriving faster than the decisions needed to extract value from it.

## Titre Article

The AI Agenda 2026 — The reality check

## Date

2026-10-15

## URL

https://www.oliverwymanforum.com/content/dam/oliver-wyman/ow-forum/template-scripts/templates/ai-agenda/2026/pdf/The_reality_check.pdf

## Keywords

AI ROI, return on intelligence, ROI harvester, AI investment portfolio, return thesis, shifted bottleneck, full cost, verification time, freed-up time, Great Leadership Misalignment, superworkers, unbundling of management, micro-CEO, learning crisis, junior jobs, bionic workforce, shadow AI, shopping agent optimization, agentic commerce, agent mandate, proof of authority, loyalty, AI recommendation, advertising in assistants, living governance, risk lab, human in the loop, LLM-as-a-judge, layered controls, AI sovereignty, controlled interdependence, control layer, vendor dependence, reversibility, Sovereign AI Index, chokepoints, EUV lithography, Global Voices, CEO Agenda 2026, NYSE

## Authors

Oliver Wyman Forum — préface de John Romeo (CEO) et Ana Kreacic (COO) ; cinq équipes d'auteurs par chapitre (Aaron Fine, Helen Leis, Adrian Oest, Daniel Cope, Abhishek Sharma et al.).

## Ton

A consulting-firm report aimed at executives, structured very regularly: each chapter opens with an "In brief," a list of "The reality is…," a two-page "By the numbers" spread, then five or six reality checks illustrated with charts. The register is prescriptive ("leaders must decide") and uses memorable phrases (*ROI harvester*, *Great Leadership Misalignment*, *micro-CEO*, *superworker*, *living governance*, *terms of dependence*) and extended metaphors (Formula 1 for governance: better brakes allow going faster).

The report is very explicit about its methodological limitations. Each chart carries the question asked, the base, and the wave; comparisons between groups are flagged as descriptive rather than causal; productivity figures are self-reported; the "Oliver Wyman analysis" diagrams are presented as a point of view rather than as estimates; the 2030 projection of AI-influenced commerce is a directional scenario; the Global Voices sample is 96% employed, so "consumer" results reflect a working population. The report presents Quotient, Oliver Wyman's AI offering, as one of its sources. Section illustrations are AI-generated, as the document states.

## Pense-betes

- **Three flaws in the ROI question**: asked too broadly (a single metric for four investment types), applied too narrowly (full cost underestimated, task-level gain mistaken for return), asked too late (after capital has been committed, hence *"retroactive storytelling"*).
- **Four investment types**: targeted applications (fast return, e.g. call summaries), employee tooling (diffuse gain), foundations (data, governance, paid for by what they enable), transformation (the largest and slowest return). Not mutually exclusive.
- **22 minutes** of verification, correction, or rework per hour of AI-assisted work (N = 7,845, average; the added cost relative to the prior process is not established).
- **Return thesis**: expected return + everything that must hold true to achieve it (e.g. a pricing tool: clean the data, retire the legacy system, get sales reps to follow through); staged funding with *scale / redesign / pause / stop* criteria set in advance.
- **Shifted bottleneck**: an agent processes a claim in seconds, but the cycle doesn't move faster if the approval queue stays the same. **52%** of CEOs with returns above 10% prioritize redesigning roles and workflows, versus **33%**.
- **ROI harvester**: an executive responsible end-to-end for converting capacity into results (cost reduction, reinvestment, growth). Only **8%** of users report a company rule on how freed-up time is used; **54%** decide on their own or without an explicit decision.
- **Portfolio**: start from the objective (growth, margin, customer experience, resilience); *quick wins* fund the effort but must not become the rule. CEOs with returns above 10% are **4.4×** more likely to have deployed AI at scale in new products (**40%** versus **9%**). A **ten-step** guide (A. build the portfolio, B. design each investment, C. prove, capture, decide). **Chapter 2 — The workforce**
- **Misalignment**: **78%** of leaders say they are ready for AI at scale, but only **37%** see a clear, followed-through ambition; CEOs and CFOs often diverge (cost lever versus growth lever).
- **Dark narrative**: **68%** would give up part of their salary for a guarantee against being replaced by AI (**47%** at least 10%, median 5%). If productivity rises, **31%** of leaders expect a bonus and **7%** expect layoffs; among non-managers, **9%** and **24%**. "Fewer people" ranks last among what *should* happen, first among what *will* happen.
- **Unbundling of management**: **80%** of daily users have learned as much or more from AI than from their manager; **65%** think AI could do their boss's job today or within a few years; non-managers spend ~a third of their week directing AI; **65%** trained themselves, **18%** received training.
- **Learning**: **74%** of managers see AI doing the formative tasks juniors used to do, **27%** say their employer has entirely rebuilt learning and onboarding; **43%** of CEOs plan to reduce junior roles (17% a year earlier); employment of 22-25 year-olds is **19%** below trend in exposed occupations (Stanford Digital Economy Lab, ADP data).
- **Superworkers**: **27%** of users say AI at least doubles their output; **38%** when five or six employer conditions are met, **17%** when two or fewer. Companies judged ahead of the curve: clear ambition **51%** versus **5%**, intent to stay **73%** versus **48%**. **Chapter 3 — Agentic commerce**
- **Starting point for purchases**: search engine 43.6%, AI **33.3%**, store 4.4%. In case of conflicting recommendations, **31.9%** would follow the AI, 22.6% people close to them, 10.6% a brand.
- **The third shelf**: 49% choose based on fit to the request, 46% based on independent reviews, 24% based on a known brand. **88%** of the sources cited in the simulation are independent of brands; **68%** would trust an assistant less if it displayed paid placements.
- **Shopping agent optimization**: making product information, evidence, and checkout flows readable and actionable by authorized agents. A six-step journey (selection, details, checkout, verification, payment, confirmation); **79%** of the 62 simulated journeys hit an obstacle.
- **Loyalty**: **54%** would consider an unfamiliar brand if the AI judged it the best fit; **60%** would switch from their usual brand. The effect varies by category (insurance exposed, iconic luxury protected).
- **Seller trust**: verify the agent, the customer, and the **mandate**; emerging payment protocols with a cryptographically signed mandate; **61%** would leave the company or its AI after a serious error affecting their money or data. Amazon blocked **Muse** (Meta) in September 2026 as an "unauthorized AI agent."
- **Delegation**: **49%** would let an agent spend without prior validation; more readily for groceries (28%) than for a loan (7%); no agent provider reaches 30% trust. See [[taylor-bavor-sierra-personal-agent-protocol-2026-10-06]]. **Chapter 4 — Governance**
- **Approved path**: easy access, clear policy, AI built into software, better tools. With three or four conditions met, **85%** say they are more productive (44% with one or none). Shadow AI: **63%** where governance seems too strict, **38%** where it seems about right; the top reason is that approved tools are worse (**35%**).
- **Risk lab**: bounded, reversible scope, an executive owner, limits, logging, exit conditions (example: a refund agent on one product line, one market, four weeks).
- **Risk by design**: business, tech, operations, legal, and risk involved from the design stage; decisions turned into reusable pathways; **52%** of the public hold the deploying organization responsible for a loss, **24%** the developer.
- **The end of blanket human in the loop**: **17%** always check important answers, **42%** have experienced a real problem caused by false AI information. Three layers precede the human: limits set in advance, automated monitoring (including calibrated *LLM-as-a-judge*), escalation rules.
- **Living governance**: approval opens the cycle (permission → observation → evaluation → adjustment → renewal); seven triggers for re-examination (model, data or prompt, tool, scale, autonomy, consequence, usage). **Chapter 5 — Sovereignty**
- **Dependence**: **55%** expect at least a moderate disruption if their primary provider goes down; **46%** of CEOs have logged concentration risk in the risk register; **10%** of tech leaders have a signed statement on retained assets; switching the primary model costs **21 to 60 engineer-days**, excluding memory and orchestration; **0.1%** of cloud customers switch providers each year (CMA).
- **Sovereignty stack** in six layers (energy, chips, models, data, orchestration and memory, applications) and four postures depending on severity and replacement lead time: *buy/switch*, *set terms/monitor*, *diversify/test the exit*, *build/ally*.
- **No need for a frontier model**: **4%** of CEOs expect a durable advantage from privileged access to a model, **74%** from redesigned processes, **64%** from proprietary data. Open-weight models trail the frontier by a few months; the compute needed for equal performance roughly halves every eight months.
- **States**: set the rules (public procurement, regulation) rather than spend; **74%** support public investment, only **19%** if it requires higher taxes or lower spending.
- **Index**: **17** of the 100 economies hold a hard-to-replace position, **7** a genuine *chokepoint*; the United States hosts **91%** of verified compute and leads five of six dimensions, but depends on the Netherlands (EUV), Japan, Korea, China (gallium), and Taiwan. France ranks **3rd** in model leadership and **2nd** in government readiness.

## RésuméDe400mots

The Oliver Wyman Forum publishes the second edition of "The AI Agenda," titled "The reality check." The opening finding: AI is everywhere in the company except the P&L. Budgets are rising, pilots are multiplying, but proven gains remain delayed, diffuse, or dependent on changes external to the technology. The report draws on more than 25,000 responses across 15 markets, a survey of 415 CEOs conducted with the NYSE, and a sovereignty index.

On return on investment, the authors distinguish four investment types with different trajectories: targeted applications, employee tooling, foundations, transformation. The calculation often omits costs, such as the 22 minutes of verification per hour of AI-assisted work, and conflates task-level gains with returns. A return thesis must be formulated before committing capital, the process must be redesigned when AI shifts the bottleneck, and a "ROI harvester" must be named to convert freed-up capacity into results. Only 12% of CEOs report a gain above 10%.

On the workforce, the report describes a leadership misalignment: 37% of senior executives and 11% of non-managers see a clear AI ambition. Lacking a narrative, employees write a darker one of their own: 62% of daily users hide their AI proficiency, 68% would trade salary for a job-security guarantee. Management is unbundling, junior learning is eroding, and 27% of users report being twice as productive, especially where the employer provides framework, tools, and training.

On commerce, AI is becoming the second most common starting point for purchases and the most followed source of advice. Brands must practice "shopping agent optimization": agent-readable product sheets, independent evidence, and checkout flows that accept authorized agents. The seller must verify the agent, the customer, and the agent's mandate.

On governance, a simple approved path reduces shadow AI; a risk lab allows new uses to be tested; human in the loop can no longer be the supervision model, replaced by layered automatic controls and a living governance that re-examines approval with every change.

On sovereignty, no country or company controls the entire stack. The goal is controlled interdependence: owning the control layer (data, context, memory, permissions), keeping the model replaceable, paying the sovereignty premium only where the stakes justify it, and favoring rules and partnerships over public spending.

## GrapheDeConnaissance

- Oliver Wyman Forum —publie→ The AI Agenda 2026 — The reality check (DOCUMENT, 0.98)
- Oliver Wyman Forum —fait_partie_de→ Oliver Wyman (ORGANISATION, 0.95)
- John Romeo —dirige→ Oliver Wyman Forum (ORGANISATION, 0.96)
- The AI Agenda 2026 — The reality check —est_basé_sur→ Global Voices Survey (DOCUMENT, 0.96)
- The AI Agenda 2026 — The reality check —est_basé_sur→ CEO Agenda 2026 (DOCUMENT, 0.95)
- CEO Agenda 2026 —mesure→ 12 % des CEO déclarent une baisse de coûts ou hausse de revenus liée à l'IA supérieure à 10 % (N = 415) (MESURE, 0.94)
- Global Voices Survey —mesure→ 22 minutes de vérification, correction ou reprise par heure de travail assisté par IA (MESURE, 0.94)
- Global Voices Survey —mesure→ 8 % des utilisateurs signalent une règle d'entreprise sur l'usage du temps gagné par l'IA (MESURE, 0.92)
- Oliver Wyman Forum —affirme_que→ l'IA crée de la capacité, pas du rendement : chaque investissement majeur exige un ROI harvester (AFFIRMATION, 0.95)
- ROI harvester —résout→ capture de la valeur de l'IA (CONCEPT, 0.85)
- Oliver Wyman Forum —affirme_que→ l'IA déplace souvent le goulot d'étranglement au lieu de le supprimer (AFFIRMATION, 0.93)
- Oliver Wyman Forum —recommande→ formuler une thèse de rendement et financer par étapes avant d'engager le capital (AFFIRMATION, 0.92)
- Great Leadership Misalignment —observé_dans→ Global Voices Survey (DOCUMENT, 0.88)
- Global Voices Survey —mesure→ 37 % des dirigeants et 11 % des non-managers voient une ambition IA claire et suivie (MESURE, 0.94)
- Global Voices Survey —mesure→ 62 % des utilisateurs quotidiens ont minimisé leur maîtrise de l'IA ; 68 % céderaient du salaire contre une garantie de non-remplacement (MESURE, 0.94)
- Oliver Wyman Forum —affirme_que→ l'IA dégroupe le management : les salariés deviennent managers d'IA et les managers des micro-CEO (AFFIRMATION, 0.9)
- superworker —mesure→ 27 % des utilisateurs ; 38 % avec cinq ou six conditions d'employeur contre 17 % avec deux ou moins (MESURE, 0.9)
- CEO Agenda 2026 —mesure→ 43 % des CEO prévoient de réduire les postes juniors (17 % un an plus tôt) (MESURE, 0.92)
- The AI Agenda 2026 — The reality check —référence→ Canaries in the Coal Mine (DOCUMENT, 0.9)
- Oliver Wyman Forum —a_créé→ shopping agent optimization (METHODOLOGIE, 0.9)
- shopping agent optimization —s_applique_à→ commerce agentique (CONCEPT, 0.92)
- Global Voices Survey —mesure→ 48 % de la génération Z et 40 % des millennials commencent par l'IA pour trouver un nouveau produit (MESURE, 0.94)
- Oliver Wyman Forum —mesure→ 70 % des fiches produit auditées incomplètes pour un achat ; 79 % de 62 parcours simulés bloqués après la recommandation (MESURE, 0.92)
- Oliver Wyman Forum —affirme_que→ le vendeur doit vérifier l'agent, le client et le mandat avant d'autoriser un achat par agent (AFFIRMATION, 0.92)
- Amazon —s_oppose_à→ Muse (TECHNOLOGIE, 0.88)
- Oliver Wyman Forum —affirme_que→ A human in the loop cannot be the operating model for AI oversight (CITATION, 0.95)
- gouvernance vivante —remplace→ human in the loop (METHODOLOGIE, 0.82)
- gouvernance vivante —utilise→ LLM-as-a-judge (METHODOLOGIE, 0.85)
- laboratoire de risque —permet→ gouvernance vivante (METHODOLOGIE, 0.8)
- Global Voices Survey —mesure→ 85 % se disent plus productifs avec 3-4 conditions de gouvernance contre 44 % avec 0-1 ; shadow AI 63 % si trop strict contre 38 % (MESURE, 0.92)
- interdépendance contrôlée —affine→ souveraineté IA (CONCEPT, 0.88)
- Oliver Wyman Forum —recommande→ posséder la couche de contrôle (données, contexte, mémoire, permissions) et garder le modèle frontière remplaçable (AFFIRMATION, 0.93)
- CEO Agenda 2026 —mesure→ 4 % des CEO attendent un avantage durable d'un accès privilégié à un modèle ; 74 % des processus refondus (MESURE, 0.92)
- Sovereign AI Index —mesure→ 17 des 100 premières économies tiennent une position difficile à remplacer, 7 un vrai chokepoint ; 91 % du calcul vérifié aux États-Unis (MESURE, 0.92)
- Oliver Wyman Forum —a_créé→ Sovereign AI Index (DOCUMENT, 0.95)

---
Canonical: https://www.thekb.eu/en/fiches/oliver-wyman-forum-ai-agenda-reality-check-2026-10-15/