# mit-nanda-genai-divide-95-percent-pilots-fail-legalio-2025-08-23

## Veille

Legal.io relay of the MIT NANDA study "The GenAI Divide: State of AI in Business 2025": 95% of enterprise AI pilots deliver no measurable ROI despite $30-40B invested. Concept of the "GenAI Divide", "shadow AI economy", four structural failure factors, back-office and build-vs-buy recommendation. Empirical justification for the HR-organizational shift.

## Titre Article

MIT Report Finds 95% of AI Pilots Fail to Deliver ROI, Exposing "GenAI Divide"

## Date

2025-08-23

## URL

https://www.legal.io/articles/5719519/MIT-Report-Finds-95-of-AI-Pilots-Fail-to-Deliver-ROI-Exposing-GenAI-Divide

## Keywords

MIT NANDA, GenAI Divide, 95% pilot failure, enterprise AI ROI, shadow AI economy, agentic AI, Model Context Protocol, NANDA, back-office automation, build vs buy, enterprise AI deployment, $30-40 billion invested, AI governance, P&L impact, enterprise paradox, Tech and Media, sales and marketing bias

## Authors

Legal.io (relais et synthèse) — étude MIT NANDA "The GenAI Divide: State of AI in Business 2025"

## Ton

**Profile**: B2B media synthesis article (Legal.io, legal talent platform), journalistic-business register with a statistical emphasis. Neutral relay voice, anchored by figures from the MIT source and direct quotes from the report. Target audience: corporate lawyers, legal ops leaders, and more broadly C-level decision-makers seeking a benchmark on the actual state of AI adoption.

**Style**: Article structured as "key points → context → 4 factors → shadow economy → recommendations → agentic AI → conclusion". High figure density (52 executive interviews, 153 surveys, 300 deployments analyzed, 95% / 5% / 80% / 40% / 60% / 20%). The document's strength lies in its **argumentative reach**: it is the study that shifts the industry debate from *"AI works"* to *"AI fails by default of adoption"*. Iconic quote from the manufacturing COO: *"The hype on LinkedIn says everything has changed, but in our operations, nothing fundamental has shifted."* Programmatic conclusion: *"from building to buying, from central labs to empowered teams, and from static tools to adaptive systems"*. Grave-realistic tone, with no indulgence for companies' official narrative. Key piece of the HR-organizational dossier: if 95% fail by default of adoption, then the bottleneck is cultural and it is up to HR to solve it.

## Pense-betes

- **Source study**: MIT NANDA, *"The GenAI Divide: State of AI in Business 2025"*, published July 2025. Methodology: **52 executive interviews + 153 leader surveys + 300 public deployments analyzed**.
- **Shock figure**: **95% of enterprise AI pilots deliver no measurable P&L impact**. Only 5% create significant value.
- Estimated total investment: **$30-40 billion** in enterprise GenAI.
- Adoption ≠ transformation: **80%+ of organizations have piloted ChatGPT/Copilot, ~40% have deployed**, but these systems **mainly boost individual productivity**, not enterprise outcomes.
- Enterprise-grade funnel: **60% evaluate → 20% pilot → 5% in production**.
- Pivotal quote from the manufacturing COO: *"The hype on LinkedIn says everything has changed, but in our operations, nothing fundamental has shifted."*
- **4 structural factors of the GenAI Divide**: 1. **Limited disruption**: only 2 of the 9 major sectors (**Tech and Media**) show material business transformation. 2. **Enterprise paradox**: large enterprises lead in pilot volume but lag in deployment. 3. **Investment bias**: AI budgets favor sales and marketing, while **operations and finance offer better ROI**. 4. **Implementation advantage**: **tools built by external vendors succeed 2× more often** than internal builds.
- **Shadow AI economy**: only 40% of companies have official LLM subscriptions, but **90% of employees** report daily use of personal tools (ChatGPT, Claude). Performance often better than corporate tools. Governance gap.
- Employee preference: they prefer their personal accounts to internal tools, **even when the underlying models are identical**. For high-stakes tasks (legal, client communication), **90% prefer human oversight** (AI's memory and context limitations).
- **Recipe for success (the 5% that succeed)**: (1) vendor partnerships with customizable, learning systems, (2) focus on workflow integration, (3) back-office deployment (document automation, procurement, risk review).
- Concrete back-office ROI cases: **$2-10M annual savings** on outsourced support and document review; **-30%** in external agency spend on marketing/content; **$1M annual savings** in financial risk monitoring.
- **Next phase = Agentic AI**: systems that remember, learn, act autonomously. Structuring protocols: **NANDA** (Networked Agents and Decentralized AI, MIT project) and **MCP (Model Context Protocol)** from Anthropic. Vision of an "Agentic Web" replacing static SaaS.
- Programmatic conclusion: *"The GenAI Divide isn't inevitable. But bridging it requires a fundamental shift — from building to buying, from central labs to empowered teams, and from static tools to adaptive systems."*
- **Link with the HR narrative**: the 95% failure rate empirically validates the Mollick and Bersin thesis — the technology works but adoption is cultural/organizational. The lever is HR (empowered teams, change management, destigmatizing shadow AI), not IT (which builds and abandons).

## RésuméDe400mots

In August 2025, Legal.io relays the MIT NANDA study *"The GenAI Divide: State of AI in Business 2025"*, which becomes one of the year's most cited empirical references on the massive failure of enterprise AI adoption. The methodology is solid: 52 executive interviews, 153 surveys, analysis of 300 public deployments.

The central figure is staggering: **95% of enterprise AI pilots deliver no measurable P&L impact**, despite $30-40 billion invested. Only 5% create significant value. This is the **"GenAI Divide"**: a gap between strong adoption and weak transformation. 80%+ of organizations have piloted ChatGPT or Copilot, ~40% say they have deployed, but these systems mainly improve individual productivity, not enterprise outcomes. On the enterprise-grade systems side, the funnel is even more brutal: 60% evaluate, 20% pilot, **only 5% reach production**.

Iconic quote from the manufacturing COO interviewed: *"The hype on LinkedIn says everything has changed, but in our operations, nothing fundamental has shifted."*

The report identifies four structural factors. **Limited disruption**: only 2 of the 9 major sectors (Tech, Media) show material business transformation. **Enterprise paradox**: large enterprises lead in pilot volume but lag in deployment. **Investment bias**: budgets favor sales/marketing while operations and finance offer better ROI. **Implementation advantage**: tools built by external vendors succeed **twice as often** as internal builds.

The study also documents the **"shadow AI economy"**: only 40% of companies have official LLM subscriptions, but 90% of employees use their personal tools daily. These shadow systems are often more performant and more quickly adopted than corporate tools — a massive governance gap. For high-stakes tasks (legal, client communication), 90% of users prefer human oversight, as AI struggles with memory and specific context.

Organizations that succeed share three traits: vendor partnerships with customizable systems, focus on workflow integration, priority back-office deployment (document automation, procurement, risk review). Documented ROI: $2-10M annual savings on outsourced support, -30% on marketing agencies, $1M on financial risk monitoring.

The next phase will be **Agentic AI**: systems that remember, learn, and act autonomously. Structuring protocols: **NANDA** (MIT) and **MCP** (Anthropic), which pave the way for an "Agentic Web" replacing static SaaS.

The report concludes: *"The GenAI Divide isn't inevitable. But bridging it requires a fundamental shift — from building to buying, from central labs to empowered teams, and from static tools to adaptive systems."* This is the study that shifts the industry debate: the bottleneck is no longer the technology, it's adoption — and so the lever is no longer IT, but HR.

## GrapheDeConnaissance

- MIT NANDA —publie→ The GenAI Divide State of AI in Business 2025 (DOCUMENT, 0.98)
- GenAI Divide —est_instance_de→ fossé entre adoption et transformation IA (CONCEPT, 0.97)
- MIT NANDA —affirme_que→ 95% des pilotes IA en entreprise n'ont aucun impact P&L (MESURE, 0.98)
- Pilotes IA enterprise —mesure→ 5% de succès (MESURE, 0.97)
- Investissement IA enterprise —mesure→ 30-40 milliards de dollars (MESURE, 0.95)
- Shadow AI economy —mesure→ 90% des employés utilisent une IA personnelle quotidiennement (MESURE, 0.95)
- Vendors externes —surpasse→ builds internes (CONCEPT, 0.93)
- Tech et Media —est_instance_de→ seuls secteurs avec transformation IA matérielle (CONCEPT, 0.92)
- Operations et Finance —surpasse→ sales et marketing en ROI IA (CONCEPT, 0.92)
- MIT NANDA —recommande→ passer du build à l'achat (from building to buying) (AFFIRMATION, 0.95)
- Back-office —permet→ meilleur ROI IA enterprise (CONCEPT, 0.92)
- IA agentique —est_instance_de→ phase suivante de l'IA enterprise (CONCEPT, 0.93)
- NANDA —est_instance_de→ Agentic Web (CONCEPT, 0.92)
- Model Context Protocol —est_instance_de→ Agentic Web (CONCEPT, 0.92)
- MIT NANDA —affirme_que→ le bottleneck de l'adoption IA enterprise est organisationnel, pas technologique (AFFIRMATION, 0.95)

---
Canonical: https://www.thekb.eu/en/fiches/mit-nanda-genai-divide-95-percent-pilots-fail-legalio-2025-08-23/