LesenForschungRadarAnlageframework
Anmelden / Registrieren
← Zurück zur Lektüre
NewsThema · 企业AI落地瓶颈Wesentlich3 Min. LesezeitDisconfirmAI Redaktion

Accenture Chief AI Officer Lan Guan: Only 10-20% of Enterprise Workflows Justify Frontier Model Capabilities

Accenture's Chief AI and Data Officer Lan Guan, drawing on experience from over 18,000 projects, argues that scaling enterprise AI requires a seven-layer infrastructure spanning data, models, agents, governance, security, operations, and front-end—models are just one ingredient; internal research suggests only 10-20% of enterprise workflows justify frontier model capabilities.

2026-10-19 17:05 GMT+8

Lan Guan, Accenture's Chief AI and Data Officer, used a kitchen metaphor to explain why most enterprise AI projects stall at the pilot stage: most companies have baked their cake in "Grandma's kitchen," but to produce that cake consistently and safely millions of times requires an industrial kitchen—equipment, repeatable processes, and quality controls. Behind the metaphor is a concrete observation: since ChatGPT's launch, Accenture's team has completed over 18,000 AI projects, and from those projects they've seen that the vast majority of enterprises are stuck in the same place—not because models aren't powerful enough, but because the infrastructure around models doesn't exist.

Lan Guan describes enterprise AI infrastructure as a seven-layer cake: data knowledge, models, agentic core, governance, security and safety, scaled AI operations, and the user front end, all built on top of robust AI infrastructure. His core argument is that "the model is only one ingredient; you need everything else in the kitchen to bake the cake." This isn't an abstract framework but one distilled from concrete projects: he cites the example of building an AI companion for a large airline. The team initially rushed to the models like everyone else, connected to powerful general-purpose LLMs, then discovered they knew nothing about that airline's customers, routes, or operational rules—they had to integrate airline data and clean large volumes of messy standard operating procedures and policy documents to turn the model into a usable product.

Anmelden / Registrieren
LesenForschungRadarAnlageframework

From these projects, Accenture's internal research produced a figure with direct implications for cost structure: only 10-20% of enterprise workflows justify the use of frontier model capabilities. Lan Guan says many companies are just starting to realize that not every workflow deserves the most expensive model. This realization means companies don't need to default to the strongest model for every task, and Accenture's approach is to help them select the right model for the right work, while also partnering with model companies to build repeatable industry solutions and launching full-lifecycle services for open-weight models, giving clients more options for data control.

But scaling is far from complete. Accenture CEO Julie Sweet acknowledged that only 5% of organizations have an enterprise AI stack that would even allow AI to be repeatable, let alone deliver the kind of ROI the hype cycle has promised. Sweet compared the current phase to two industry resets in the past 20 years—offshoring and cloud/digital—each of which eliminated some services, but underlying technology enabled enterprises to do so much more that new demand far exceeded what was replaced. She believes the AI era will follow the same pattern, but it will take time.

It's worth noting that these observations come from a company that directly sells the services in question. Accenture's business model is helping enterprises build "industrial kitchens," so the narrative that "you need more than just a model" naturally serves its revenue. The 18,000 projects, 5% repeatable-stack figure, and 10-20% frontier-model applicability are all self-reported data without published methodology or sample composition. However, the seven-layer framework as an analytical tool—reminding readers that models are just one link in the deployment chain, with data cleaning, governance, operations, and front-end equally constituting bottlenecks—retains explanatory value independent of Accenture's commercial position.

Quellen

  • diginomica.com↗
Thema

企业AI落地瓶颈

【背景】尽管生成式AI技术快速迭代,但企业在将其转化为实际业务价值时仍面临严峻挑战。埃森哲首席AI与数据官Lan Guan指出,仅10-20%的企业工作流值得使用前沿模型能力,规模化需要七层基础设施支撑。ThinkingAI韩国区总经理서광…

5 Schritte2 BeiträgeAktualisiert 2026-10-21 00:51

Wie es dazu kam

  1. 2024-01-10IBM发布《全球AI采用指数2023》,指出42%的大型企业已部署AI,但40%仍在探索阶段,技能缺口(33%)和数据复杂性(25%)为主要障碍。newsroom.ibm.com ↗
  2. 2026-06-15新华网报道指出,企业AI应用存在“最后一公里”堵点,数据可见不可用及项目价值失踪是核心问题,需将AI嵌入核心决策环节。news.cn ↗
  3. 2026-07-20Salesforce副总裁Sumeet Agrawal撰文分析“漏水的AI管道”,指出上下文膨胀、无治理数据访问及错误模型选择导致Token浪费,呼吁建立架构级优化。fortune.com ↗

Was noch ungeklärt ist

  • 如何量化评估特定工作流是否值得使用高成本的前沿模型?
  • 在多云或多模型环境下,企业如何统一治理标准以消除“影子AI”风险?
  • 数据治理工具与业务规则嵌入的具体实施路径有哪些最佳实践?

Auch in diesem Thread

  • ThinkingAI Korea Head Seo Gwang-deok: AI Moves Fast but Business Doesn't—Bottleneck Is Operations, Not Models2026-10-20
Vollständiges Thema lesen →
Recherche fortsetzen →