Dataiku Pushes Modular AI as CIO Regret Over Vendor Lock-In Hits 74%
Ai4
Relying entirely on rented frontier models surrenders proprietary domain logic and exposes firms to sudden access cuts, requiring enterprises to decouple and own their decision layer while treating AI providers as swappable suppliers.
## Background
Published on YouTube by Ai4 on August 15, 2026, this talk features Catalina Herrera, Field CDO at Dataiku. According to the description, Herrera addresses the operational and strategic risks of relying on single AI providers and outlines how organizations can design modular, swappable architectures to protect governance and retain control of their enterprise intelligence.
## TL;DR
Herrera warns enterprise leaders against falling into the 'AI dependency trap' of relying entirely on rented intelligence that can be abruptly revoked, restricted, or altered by third-party vendors. She contends that foundational models are commoditized suppliers and that a company's true competitive advantage lies in its proprietary business judgment, domain workflows, and decision logic. By decoupling and owning the enterprise decision layer, organizations can ensure resilience against outages and model refusals while seamlessly swapping models underneath. Real-world implementations, such as SLB accelerating $10 billion in tender assessments, demonstrate the measurable value of maintaining a vendor-agnostic architecture.
## Atmosphere & style
The presentation carries a sharp, pragmatic tone that serves as an executive wake-up call, blending sobering real-world disruption stories with reassuring architectural solutions. Herrera speaks with the seasoned confidence of a field practitioner who has seen enterprise AI initiatives succeed and stumble firsthand. The narrative transitions smoothly from analytical caution regarding vendor lock-in to an empowering, actionable blueprint for enterprise autonomy.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) Many enterprise AI initiatives fall into an 'AI dependency trap' by relying heavily on rented intelligence that third-party vendors can revoke at any time without SLAs or appeals.
- [00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) Frontier AI introduces four shifting input variables that organizations must navigate: availability, price, performance, and permission.
- [04:36](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=276s) Off-the-shelf model guardrails can lead to critical operational failures, such as refusing to analyze security attack logs because they cannot distinguish attackers from legitimate security teams.
- [04:36](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=276s) A 2026 survey found that 72% to 74% of CIOs regret a major AI vendor decision made within the previous 18 months due to severe vendor lock-in.
- [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) An enterprise's competitive edge is never the underlying foundation model itself, which competitors can also rent, but rather its accumulated business judgment and domain knowledge.
- [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) Enterprises must own the decision layer—the space where subject matter experts define prompts, pricing logic, and risk rules—while keeping the underlying models swappable.
- [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) Decoupling business logic from supplier models allows organizations to benchmark token pricing, establish fallback plans, and avoid vendor outages.
- [14:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=866s) Dataiku 575 Lab features like Privacy Proxy and Inspector safeguard enterprise intelligence by automatically redacting sensitive data and tracking autonomous agent actions.
- [14:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=866s) SLB implemented a modular, vendor-agnostic tender assessment platform on Dataiku that reduced evaluation times for $10B+ in energy tenders from days to 20 minutes.
- [20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) Enterprise leaders should treat foundation models as interchangeable suppliers and maintain an agnostic architecture to protect core intellectual property.
## Detailed walkthrough
Here is a detailed, faithful walkthrough based on the transcript provided:
[00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) The "Renting Intelligence" Trap and Policy Shifts
Catalina Herrera, Field CDO at Dataiku, opens her presentation by explaining that her day-to-day role involves working directly inside Fortune 500 AI programs—observing both the projects breaking new ground and those going sideways. She warns IT leaders that many are walking straight into an "AI dependency trap" with their eyes wide open. To illustrate this vulnerability, she recalls a pivotal moment on June 12th when a major AI company received a governmental directive requiring them to block foreign nationals from accessing two of the most advanced AI models on the market. Unable to distinguish foreign from non-foreign users reliably, the vendor simply shut down model access overnight without warning, SLA, or appeal mechanism.
Eighteen days later, model access was restored, yet most enterprise teams simply redeployed their pipelines using the exact same underlying architecture from June 11th. Herrera highlights this reaction as the core trap: organizations rely heavily on rented intelligence that depends on permissions controlled entirely by third parties who can revoke them at any point.
- **4 Key Input Variables of Frontier AI:**
- **Availability:** Fluctuates unexpectedly due to outages or sudden geopolitical policy shifts.
- **Price:** Varies with usage spikes and vendor repricing.
- **Performance:** Changes continuously as newer models eclipse yesterday's best options.
- **Permission:** A new, unpredictable dial where access depends on external governance.
[04:36](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=276s) Model Refusals and Executive Regret
Moving beyond access revocations, Herrera brings up a recent incident involving Hugging Face. When security responders detected an attack, they forwarded attack logs to a hosted frontier model for analysis. Unexpectedly, the model refused to analyze the logs because its built-in guardrails could not distinguish between the malicious attackers and the legitimate security team. This scenario demonstrates that relying on off-the-shelf, rented intelligence leaves teams helpless during critical moments when they need the AI most.
These continuous operational hurdles are escalating into C-suite anxieties. She cites a 2026 survey revealing that 72% to 74% of CIOs regret a major AI vendor decision made within the previous 18 months. These leaders committed heavily to proprietary platforms before realizing how difficult and expensive it would be to migrate or update their tech stacks later on.
[06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) Preserving Enterprise Judgment as the Competitive Edge
Herrera asks the audience to consider what truly constitutes a company's competitive advantage. She argues that an enterprise's edge can never be the AI model itself, because competitors can easily rent the exact same underlying model. Off-the-shelf intelligence raises the operational floor for everyone equally; relying solely on it makes an organization identical to its market rivals.
Instead, an organization's true differentiator resides in its accumulated business judgment—the domain knowledge, operational processes, pricing logic, and collaborative workflows refined over time. Currently, these crucial assets are often fragmented across disparate tools: prompt engineering buried in inventory workflows, pricing logic locked inside vendor platforms, or risk rules trapped in isolated silos. When enterprises build directly inside someone else’s proprietary stack, they risk handing over their core intellectual property and operational identity.
[11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) Modular Architecture and Owning the Decision Layer
Organizations do not need to own the entire underlying infrastructure, but they must retain full ownership of the **decision layer**. This layer is where subject matter experts (SMEs) systematically refine prompts, encode domain-specific logic, and structure business questions. By decoupling business logic from supplier models, everything beneath the decision layer becomes completely modular and swappable.
```mermaid
accTitle: Modular AI Architecture and Decision Layer
flowchart TD
subgraph DL ["Decision Layer (Enterprise Owned)"]
direction LR
SME["Domain Experts & Business Logic"]
Prompts["Refined Prompts & Decision Rules"]
SME --> Prompts
end
subgraph OL ["Dataiku Orchestration & Safety Layer"]
direction LR
PP["Privacy Proxy (Redaction)"]
Inspect["Inspector (Agent Tracking)"]
PP --> Inspect
end
subgraph IL ["Modular Intelligence Layer (Swappable Suppliers)"]
direction LR
M1["Model A (Frontier LLM)"]
M2["Model B (Open-Source / On-Prem)"]
M3["Model C (Agentic Workflows)"]
end
DL --> OL
OL --> IL
```
Owning the decision layer enables organizations to create fallbacks (Plan B or Plan C), benchmark token pricing between competing LLM vendors, and maintain uptime during vendor outages. When Hugging Face encountered a model refusal during their incident response, having direct control over their stack allowed them to switch models swiftly and complete their investigation.
[14:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=866s) Dataiku 575 Lab Innovations and SLB Case Study
Herrera notes a broader architectural shift back toward controlling hardware, data, and core compute environments. Within this evolving setup, Dataiku and its **575 Lab** introduce two dedicated security and transparency features:
- **Privacy Proxy:** Protects sensitive internal information by automatically redacting data before it ever leaves the organization's enterprise boundary.
- **Inspector:** Provides full operational visibility into autonomous agent actions and reasoning steps behind the scenes.
To illustrate these principles in action, Herrera highlights **SLB** (Schlumberger). SLB developed an automated tender assessment system built on Dataiku that processes over $10 billion in tenders within the energy sector. By adopting a modular, vendor-agnostic framework, SLB reduced tender evaluation times from several days down to just 20 minutes—achieving measurable ROI at enterprise scale while maintaining total architectural control.
[20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) Strategic Retrospective and Core Takeaways
Dataiku's standing as a Leader in the 2026 Gartner Magic Quadrant for Data Science and Machine Learning Platforms for the fifth consecutive year reflects its long-term commitment to remaining fully vendor-agnostic amid industry consolidation, generative AI trends, and agentic workflows. Herrera concludes her presentation with a summary for enterprise leaders:
- **Acknowledge Operational Realities:** Accept that rented model access can be revoked, gated, or refused without notice.
- **Separate Decisions from Suppliers:** Treat foundational models as interchangeable suppliers while retaining full ownership of the enterprise decision layer.
- **Maintain Architectural Agnosticism:** Build flexible orchestration pipelines that allow models to be swapped easily based on cost, performance, or availability.
- **Protect Enterprise Intellectual Property:** Ensure proprietary prompts, domain expertise, and business rules remain internal assets rather than locked within vendor platforms.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) organizations rely heavily on rented intelligence that depends on permissions controlled entirely by third parties who can revoke them at any point. — Catalina Herrera
> [04:36](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=276s) relying on off-the-shelf, rented intelligence leaves teams helpless during critical moments when they need the AI most. — Catalina Herrera
> [04:36](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=276s) 72% to 74% of CIOs regret a major AI vendor decision made within the previous 18 months. — Catalina Herrera
> [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) an enterprise's edge can never be the AI model itself, because competitors can easily rent the exact same underlying model. — Catalina Herrera
> [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) Off-the-shelf intelligence raises the operational floor for everyone equally; relying solely on it makes an organization identical to its market rivals. — Catalina Herrera
> [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) When enterprises build directly inside someone else’s proprietary stack, they risk handing over their core intellectual property and operational identity. — Catalina Herrera
> [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) Organizations do not need to own the entire underlying infrastructure, but they must retain full ownership of the decision layer. — Catalina Herrera
> [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) By decoupling business logic from supplier models, everything beneath the decision layer becomes completely modular and swappable. — Catalina Herrera
> [20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) Acknowledge Operational Realities: Accept that rented model access can be revoked, gated, or refused without notice. — Catalina Herrera
> [20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) Separate Decisions from Suppliers: Treat foundational models as interchangeable suppliers while retaining full ownership of the enterprise decision layer. — Catalina Herrera
> [20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) Protect Enterprise Intellectual Property: Ensure proprietary prompts, domain expertise, and business rules remain internal assets rather than locked within vendor platforms. — Catalina Herrera
## Mentioned
Catalina Herrera, Dataiku, Dataiku 575 Lab, Privacy Proxy, Inspector, Hugging Face, SLB, Schlumberger, Gartner, Gartner Magic Quadrant for Data Science and Machine Learning Platforms
## Video description
Catalina Herrera - Field CDO, Dataiku
As AI evolves faster than governments can regulate it, enterprise leaders face growing uncertainty around access, cost, and control. Building AI on a single provider creates operational and strategic risk. Catalina Herrera shares how organizations can design AI architectures that keep models swappable, protect governance, and maintain control of enterprise intelligence—so they can adapt to change without rebuilding their AI strategy.
Enterprises risk severe operational disruption and the loss of core intellectual property by embedding proprietary workflows directly into rented foundation AI models, argues Dataiku Field CDO Catalina Herrera. She contends that long-term competitive advantage requires organizations to own their internal decision layer while treating external AI models as interchangeable commodity suppliers.
Operational Risks of Rented Intelligence (0:04)
Herrera warns that relying on commercial frontier AI exposes organizations to four volatile external variables: availability, price, performance, and permission. To demonstrate permission risks, she points to an incident on June 12th when a leading vendor abruptly revoked access to two frontier models following a government directive restricting foreign nationals. Lacking mechanisms to verify user nationalities, the provider suspended model access for 18 days without warning, service level agreements, or an appeals process. Despite the disruption, most affected enterprises redeployed their workflows onto the exact same fragile architecture once access was restored.
Operational dependencies also surface through automated guardrails. In one instance, security responders at Hugging Face forwarded attack logs to a hosted frontier model for analysis, only for the model to refuse the task because its safety guardrails could not distinguish the defensive security team from malicious actors. Herrera argues that such refusals leave enterprises vulnerable during critical incidents when external intelligence cannot be adjusted.
Misplaced Differentiation and Executive Lock-in (6:43)
These operational challenges have translated into widespread executive dissatisfaction. Herrera cites a 2026 survey indicating that 72% to 74% of CIOs regretted a major AI vendor decision made within the prior 18 months, driven primarily by the high costs and friction of migrating away from proprietary platforms. Herrera asserts that off-the-shelf foundation models cannot provide a durable competitive moat, as competitors can rent the exact same capabilities, merely raising the baseline standard across an industry.
According to Herrera, an organization's authentic differentiation lies in its accumulated business judgment, which includes specialized domain knowledge, operational processes, collaborative workflows, and pricing logic. When enterprises hardcode these rules and prompts directly into a single vendor's managed platform, they scatter their core intellectual property across isolated silos and relinquish control over their operational identity.
Decoupling Enterprise Logic from Model Providers (11:13)
To maintain autonomy, Herrera advocates establishing an enterprise-owned decision layer that separates internal business logic from underlying model suppliers. Within this architecture, subject matter experts formulate prompts, encode business logic, and structure queries inside their own environment. By isolating this decision layer, underlying large language models and open-source systems become fully modular and interchangeable.
This separation allows enterprises to create operational fallbacks, benchmark token costs across competing model vendors, and sustain uptime when a specific supplier experiences an outage or refusal. For example, when Hugging Face experienced an automated model refusal during its security incident, controlling its surrounding architecture allowed the team to switch to an alternative model and complete the log investigation.
Modular Governance and Scaled Implementation (14:26)
Herrera highlights a broader enterprise trend toward reclaiming control over compute infrastructure, internal data, and security boundaries. Within this framework, Dataiku's 575 Lab developed features such as Privacy Proxy, which automatically redacts sensitive data before queries cross enterprise perimeters, and Inspector, which audits the reasoning steps and actions executed by autonomous agents.
As an illustration of modular design at scale, energy services company SLB implemented a vendor-agnostic tender assessment platform on Dataiku to evaluate more than $10 billion in energy sector contracts. By preserving architectural independence and decoupling logic from underlying model vendors, SLB reduced tender evaluation timelines from several days to 20 minutes while retaining ownership of its decision rules.
Herrera emphasizes that sustaining enterprise AI value requires treating foundational models as interchangeable utilities rather than unique differentiators. While platform abstractions can mitigate vendor lock-in and model refusals, organizations face the ongoing task of maintaining governance and auditing agentic workflows as model options and external regulations continue to shift.
## Background
本内容由 Ai4 于 2026 年 8 月 15 日在 YouTube 平台发布，主讲人为 Dataiku 的 Field CDO Catalina Herrera。视频围绕企业在快速演进且监管不确定的 AI 浪潮中，如何通过构建模型可插拔的模块化架构、完善治理机制并牢牢掌握决策层，从而规避单一供应商绑定风险并保障企业核心智能资产展开。
## TL;DR
演讲指出了企业盲目“租用智能”和深度绑定单一闭源大模型所面临的突发断供、安全护栏误判以及高昂切换成本等核心风险。主讲人强调，开箱即用的通用模型无法构成企业壁垒，真正的护城河是沉淀在业务专家经验中的“判断力”与独特决策逻辑。企业应当将底层模型解耦为可灵活替换的“供应商”，在数据、算力和智能层建立中立自主的控制架构。通过斯伦贝谢（SLB）处理超百亿美元标书评估的实战案例，演讲展示了端到端编排如何帮助企业在保持架构独立的同时实现显著的生产力跃迁。
## Atmosphere & style
整场演讲风格沉稳、犀利且极具战略清醒感，主讲人以一线实战视角揭示了企业 AI 落地中被忽视的架构硬伤与高管焦虑。叙事从突发断供和安全误判的现实危机切入，迅速将听众带入对技术掌控权与供应商锁定的深度反思中。随着解耦思路和标杆案例的层层推导，现场氛围由警示转向务实坚定，传递出一种在技术狂热中保持自主权与理性工程思维的力量感。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) 完全依赖单一外部前沿AI供应商存在巨大的结构性断供风险，如6月12日某供应商因合规问题在夜间直接切断访问权限长达18天，暴露出“租用智能”缺乏SLA和申诉渠道的脆弱性。
- [03:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=206s) 在将前沿AI引入生产环境时，除了可用性、价格和性能这三个传统变量外，“权限（Permission）”已成为不可忽视的第四个核心变量。
- [03:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=206s) 大模型的安全护栏机制可能在关键时刻发生误判，例如在安全事件调查中拒绝分析攻击日志，导致防守方丧失对核心系统的响应与控制权。
- [03:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=206s) 调研显示74%的CIO后悔过早绑定单一AI供应商，模型迁移与技术栈重构的成本极其高昂。
- [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) 开箱即用的通用模型无法构成企业护城河，因为竞争对手同样可以租用；企业真正的壁垒在于长期沉淀的业务判断力（Judgment）与专有业务逻辑。
- [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) 企业应对外部AI波动的核心策略是牢牢掌控“决策层（Decision Layer）”，将底层LLM解耦为可灵活替换、比价与熔断的插拔式供应商。
- [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) “选择权不是一种被动态度，而是一种预先设计的技术架构”，需要支持根据任务难度与Token成本在多模型间进行动态路由与分流。
- [14:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=866s) 企业级可控架构应明确划分三层边界：数据层（永久自主掌控）、算力层（现代化弹性调度）与智能层（多模型与Agent统一治理）。
- [14:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=866s) 通过隐私代理（Privacy Proxy）脱敏与智能体巡检器（Inspector）审计，企业能够在享受前沿模型红利的同时保障数据安全与Agent行为透明度。
- [17:32](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1052s) 业务领域专家（SMEs）是AI生态落地的关键，必须打通传统机器学习、生成式AI与Agentic架构，实现端到端的流水线编排。
- [17:32](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1052s) 能源巨头斯伦贝谢（SLB）基于中立解耦架构打造标书评估系统，累计处理超百亿美元标书，将评估周期从数天大幅压缩至20分钟。
- [20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) 企业可以尽情租用和压榨外部模型的算力，但必须将“如何做决定”的编排规则与Prompt固化为自有数字资产，避免沦为数据租客。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) “租用智能”的陷阱与6月12日的突发断供
在财富500强企业的AI项目一线，既能看到开创性的成功案例，也能目睹大量项目陷入困境。许多IT领导者正“睁着眼睛走进陷阱”——完全依赖外部单一模型供应商所带来的结构性风险。回顾近期的行业震荡，6月12日发生的一起标志性事件敲响了警钟：一家头部AI公司接到政府指令，要求限制外国公民访问市场上最先进的两款前沿模型；由于该供应商无法在短时间内精准识别用户国籍，索性在夜间直接对所有相关访问进行了“一刀切”的全面断供。
```mermaid
flowchart TD
accTitle: 租用智能的脆弱链路与断供风险
subgraph External["外部供应商依赖风险"]
Gov["政策/监管突发指令"] --> Vendor["头部AI模型服务商"]
Vendor -->|"一夜之间关停访问"| Model["前沿闭源大模型"]
end
subgraph Enterprise["企业生产环境（脆弱架构）"]
Model -.->|"无SLA / 无申诉渠道断连"| Workloads["企业生产工作流"]
Workloads --> Ops["业务停摆与被动等待"]
end
```
对于全球那些基于该模型构建核心业务工作流的企业而言，模型访问权限在一夜之间因外部不可控的政策决策而彻底消失。企业是在“租用智能”——没有预警电话，没有申诉渠道，没有SLA保障，甚至找不到任何责任人来协调解决。尽管18天后访问权限得以恢复，许多团队欢呼着重新部署并继续推进生产管线，但背后潜藏的危机令人不寒而栗：底层架构没有任何改变，整个业务流程依然寄生于随时可能被第三方吊销的外部权限之上。
- 6月12日，某领先AI供应商因合规识别困难，直接关停了旗下两款顶级前沿模型的外部访问接口。
- 受影响的企业在长达18天的时间内面临核心模型不可用的真空期。
- 绝大多数企业在权限恢复后原封不动地重新上线，未对底层依赖架构做任何解耦或容灾改进。
---
### [03:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=206s) 生产环境中的第四个变量：“权限”与供应商锁定之痛
将前沿AI（Frontier AI）引入企业级生产环境时，必须将其视为业务运营中的又一个输入变量。过去，架构师通常需要权衡三个核心维度：可用性（受停机故障与政策变动影响）、价格（受算力消耗激增与厂商调价影响）以及性能（随着模型迭代，今日的最佳模型明日可能落后）。而在当下的地缘与技术环境下，系统迎来了第四个至关重要的变量——“权限（Permission）”。
这个变量的威力在近期 Hugging Face 遭遇安全攻击的事件中体现得淋漓尽致。当时应急响应团队将攻击日志提交给托管的前沿大模型进行安全分析，然而模型内置的安全护栏（Guardrails）由于无法分辨“提交日志的白帽防守方”与“恶意黑客”，直接触发了拒绝响应机制。当安全团队最需要算力与模型支持的关键时刻，模型却给出了冰冷的“拒绝”。这种在生产链路中失去控制权的现实，正引发企业高管层的集体焦虑。
```mermaid
flowchart LR
accTitle: 企业前沿AI的四维动态变量
subgraph Variables["前沿AI输入变量模型"]
A["可用性 Availability<br/>(停机 / 政策断供)"]
B["价格 Price<br/>(消耗激增 / 厂商调价)"]
C["性能 Performance<br/>(技术迭代 / 模型轮替)"]
D["权限 Permission<br/>(安全误判 / 访问吊销)"]
end
Variables --> Ops["企业级生产架构决策"]
```
- 2026年企业AI现状调研显示，高达74%的CIO后悔在过去18个月内做出的某项重大AI供应商采购决策。
- CIO们过早绑定了单一技术栈，而在面临迁移与变更时，才发现底层改造和切换的成本极其高昂。
- 安全事件中，Hugging Face 响应团队遭遇模型护栏误判，关键时刻被拒绝分析恶意日志。
---
### [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) 真正的企业护城河：从通用模型转向累积的业务判断力
面对运营与战略层面的双重不确定性，企业需要跳出具体工具，重新思考自身的核心竞争力究竟由什么构成。很多企业误以为拥有一套最先进的开源或闭源大模型就是自己的护城河，但事实恰恰相反：“凡是你能租到的模型，你的竞争对手也一样能租到。”开箱即用的通用智能（Off-the-shelf intelligence）在拉高行业平均基线的同时，也导致了严重的同质化竞争。
企业的核心壁垒绝非底层模型，而是团队在长期业务实践中积累的“判断力（Judgment）”与专有领域知识（Know-how）——即企业如何以不同于竞争对手的方式开展业务、执行独特流程以及促成跨部门协同。然而在现实中，这些宝贵的经验资产往往处于极其分散的状态：一段关键的库存管理Prompt可能写在数月前某位员工的草稿里，定价逻辑硬编码在受制于第三方的UI界面中，风控规则则深埋在某个封闭孤立的技术栈中。如果企业将这些业务认知完全托付给外部供应商的技术栈，实际上是在逐步流失最核心的企业差异化资产。
> “如果你的AI优势仅仅是模型本身，那么你根本没有任何AI优势可言。”
---
### [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) 拥有决策层：将底层模型解耦为可插拔的供应商
解决这种被动局面的核心思路，不是要求企业自研全部底层AI技术栈，而是必须牢牢掌控自己的“决策层（Decision Layer）”。决策层是业务领域专家（SMEs）提炼业务问题、持续迭代与调优Prompt、固化业务规则的地方。这一层承载着企业独有的业务逻辑，并负责向下编排与调用具体的模型与AI代理（Agents）。
```mermaid
flowchart TD
accTitle: 拥有决策层与解耦模型供应商架构
subgraph Owned["企业自主掌控：决策层 (Decision Layer)"]
SME["业务领域专家 (SMEs)"] --> Prompts["业务认知与Prompt工程"]
Rules["业务逻辑 / 风控规则"] --> Prompts
Prompts --> Orchestration["核心编排与路由引擎"]
end
subgraph Modular["可插拔模块：模型供应商层 (Suppliers)"]
Orchestration -->|"动态路由 / 成本比选"| M1["模型 A (高性能/高Token成本)"]
Orchestration -->|"熔断备用 Plan B"| M2["模型 B (私有部署/高安全性)"]
Orchestration -->|"平替切换 Plan C"| M3["模型 C (开源精调/经济型)"]
end
```
在架构设计上，必须明确“模型只是供应商，决策层才是企业本身”的分工定位。当决策层掌握在自己手中时，底层的所有模型就变成了完全模块化的可替换组件。企业可以根据任务特征灵活比对不同LLM的回答质量与Token开销，当某款模型价格过高、出现断供或拒绝响应时，可以无缝切换至备用方案（Plan B 或 Plan C）。在 Hugging Face 的案例中，团队正是通过迅速切换至本地部署且自主受控的模型，才得以绕过阻碍顺利完成安全调查。
- 保持选择权并非一种被动的态度表态，而是一种需要提前落地的技术架构（"Optionality is not a posture. It is an architecture."）。
- 架构需支持根据任务难度与成本在不同模型间进行实时路由与分发。
---
### [14:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=866s) 构建模块化控制架构：数据、算力与智能层
随着技术周期从早期的“全盘云优先”逐渐向理性可控的混合架构回归，企业应当在三层核心基础设施中明确自主掌控的边界：第一是数据层，这是企业必须永久拥有的资产；第二是算力层，随着GPU逐渐商品化，现代计算平台使得硬件资源的调度与优化更加可控；第三则是新兴的智能层（Intelligence Layer），负责对各类模型与Agent进行统一管理与性能优化。
```mermaid
flowchart TD
accTitle: 企业AI三层控制与安全治理架构
subgraph Layers["企业可控技术栈"]
L1["数据层 (Data) - 永久自主拥有"]
L2["算力层 (Compute) - 现代化引擎调度"]
L3["智能层 (Intelligence) - 统一模型与Agent治理"]
end
subgraph Gov["治理与安全机制 (Dataiku 架构)"]
Proxy["隐私代理 Privacy Proxy<br/>(敏感数据出境前自动脱敏)"]
Inspector["智能体巡检器 Inspector<br/>(全链路洞察Agent后台动作)"]
end
Layers --> Gov
Gov --> ExternalAPI["多厂商模型接入与动态切换"]
```
为了应对日益复杂的 Agentic AI 落地需求，架构还必须配备完备的治理与安全组件。例如通过隐私代理（Privacy Proxy）机制，在任何数据离开企业内网环境前完成敏感信息的自动识别与脱敏；通过巡检工具（Inspector），让架构师与合规团队能够清晰洞察AI Agent在后台的调用链条与决策动作。这种中立、解耦的模块化设计，使企业既能充分享受前沿商业模型的红利，又能随时随地行使“拔掉插头换另一家”的主动权。
- **数据层（Data）**：核心资产全生命周期自主留存与管控。
- **算力层（Compute）**：适配现代化计算引擎，动态调优基础设施性价比。
- **智能层（Intelligence）**：集成隐私脱敏（Privacy Proxy）与Agent执行透明度审计（Inspector），实现跨模型厂商的中立治理。
---
### [17:32](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1052s) 落地编排与实战范例：斯伦贝谢（SLB）百亿美元级招标评估
构建企业级AI生态需要从两大基石入手：首先是“人”，只有深谙业务痛点的业务专家才懂得如何提出正确的问题，IT部门无法凭空替代业务人员撰写Prompt；其次是“端到端的编排能力”，必须将描述性分析、预测性机器学习、生成式AI以及自主Agent等各类异构资产打通，打破企业内部普遍存在的技术孤岛，形成可复用、可替换的流水线。
在能源行业，全球巨头斯伦贝谢（SLB）通过 Dataiku 构建的模块化架构给出了标杆实践。SLB 构建了一套企业级标书评估系统（Tender Assessment System），在保持底层模型中立与架构完全独立的前提下，实现了生产环境中的大规模部署。该系统直接对接数十亿美元级别的复杂能源项目招标流程，在大幅降低对单一AI供应商依赖的同时，带来了肉眼可见的投资回报率。
```mermaid
flowchart LR
accTitle: SLB 标书评估系统落地效果
subgraph Input["业务输入"]
Tenders["全球复杂能源项目标书<br/>(累计评估规模超 100 亿美元)"]
end
subgraph Process["Dataiku 模块化编排"]
SLB["SLB 标书评估系统<br/>(与底层模型解耦 / 保持厂商中立)"]
end
subgraph Output["业务成效"]
Time["分析周期：从数天压缩至 20 分钟"]
Scale["生产级高可用与持续ROI输出"]
end
Input --> Process --> Output
```
- SLB 基于中立编排架构搭建的标书评估系统，累计评估标书总额已超过 100 亿美元。
- 标书综合分析处理时间从原先的数天直接缩短至 20 分钟。
- Dataiku 凭借跨越传统机器学习、生成式AI到 Agentic 架构的技术积淀，连续第5次入选 Gartner 数据科学与机器学习（DSML）平台魔力象限领导者（Leaders in the 2026 Magic Quadrant）。
---
### [20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) 摆脱“AI二房东”：掌控决定权与核心资产
AI 时代的核心逻辑是“以 AI 赋能人类”，但拥抱技术必须伴随着高度的战略清醒。企业可以尽情租用、测试、压榨市面上最强大的前沿模型，甚至在它们之间频繁切换；但无论外部模型如何变迁，企业必须将“如何做决定”的逻辑留存为自己的数字资产。
现实中，模型可能会突然关停，供应商可能会改变政策，安全护栏可能会产生误判。企业无需扮演为外部供应商喂养数据的“租客”，也不应将核心竞争力拱手相让。不断微调的Prompt、沉淀的业务逻辑以及对模型的编排规则，才是企业最不可替代的护城河。唯有守住决策层，才能在技术浪潮中始终掌控属于自己的竞争优势。
> “你可以租用模型，逼迫它们发挥最大潜能；但必须掌控你做决定的方式。因为‘如何做决定’才是你成其为你、而非竞争对手的核心所在。”
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) 许多IT领导者正“睁着眼睛走进陷阱”——完全依赖外部单一模型供应商所带来的结构性风险。 — Catalina Herrera
> [00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) 企业是在“租用智能”——没有预警电话，没有申诉渠道，没有SLA保障，甚至找不到任何责任人来协调解决。 — Catalina Herrera
> [00:04](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=4s) 底层架构没有任何改变，整个业务流程依然寄生于随时可能被第三方吊销的外部权限之上。 — Catalina Herrera
> [03:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=206s) 当安全团队最需要算力与模型支持的关键时刻，模型却给出了冰冷的“拒绝”。 — Catalina Herrera
> [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) 凡是你能租到的模型，你的竞争对手也一样能租到。 — Catalina Herrera
> [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) 如果你的AI优势仅仅是模型本身，那么你根本没有任何AI优势可言。 — Catalina Herrera
> [06:43](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=403s) 如果企业将这些业务认知完全托付给外部供应商的技术栈，实际上是在逐步流失最核心的企业差异化资产。 — Catalina Herrera
> [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) 模型只是供应商，决策层才是企业本身。 — Catalina Herrera
> [11:13](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=673s) 保持选择权并非一种被动的态度表态，而是一种需要提前落地的技术架构（Optionality is not a posture. It is an architecture.）。 — Catalina Herrera
> [14:26](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=866s) 这种中立、解耦的模块化设计，使企业既能充分享受前沿商业模型的红利，又能随时随地行使“拔掉插头换另一家”的主动权。 — Catalina Herrera
> [20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) 企业无需扮演为外部供应商喂养数据的“租客”，也不应将核心竞争力拱手相让。 — Catalina Herrera
> [20:30](https://www.youtube.com/watch?v=eh3GI-o_aMs&t=1230s) 你可以租用模型，逼迫它们发挥最大潜能；但必须掌控你做决定的方式。因为‘如何做决定’才是你成其为你、而非竞争对手的核心所在。 — Catalina Herrera
## Mentioned
Catalina Herrera, Dataiku, Ai4, Hugging Face, 斯伦贝谢 (SLB), Gartner, Privacy Proxy, Inspector, Frontier AI, Agentic AI, DSML
## Video description
Catalina Herrera - Field CDO, Dataiku
As AI evolves faster than governments can regulate it, enterprise leaders face growing uncertainty around access, cost, and control. Building AI on a single provider creates operational and strategic risk. Catalina Herrera shares how organizations can design AI architectures that keep models swappable, protect governance, and maintain control of enterprise intelligence—so they can adapt to change without rebuilding their AI strategy.
[Source](https://www.youtube.com/watch?v=eh3GI-o_aMs)
Source quotes
[0:04] Catalina Herrera: 许多IT领导者正“睁着眼睛走进陷阱”——完全依赖外部单一模型供应商所带来的结构性风险。
[0:04] Catalina Herrera: 企业是在“租用智能”——没有预警电话，没有申诉渠道，没有SLA保障，甚至找不到任何责任人来协调解决。
[0:04] Catalina Herrera: 底层架构没有任何改变，整个业务流程依然寄生于随时可能被第三方吊销的外部权限之上。
[3:26] Catalina Herrera: 当安全团队最需要算力与模型支持的关键时刻，模型却给出了冰冷的“拒绝”。
[6:43] Catalina Herrera: 凡是你能租到的模型，你的竞争对手也一样能租到。
[6:43] Catalina Herrera: 如果你的AI优势仅仅是模型本身，那么你根本没有任何AI优势可言。
[6:43] Catalina Herrera: 如果企业将这些业务认知完全托付给外部供应商的技术栈，实际上是在逐步流失最核心的企业差异化资产。
[11:13] Catalina Herrera: 模型只是供应商，决策层才是企业本身。
[11:13] Catalina Herrera: 保持选择权并非一种被动的态度表态，而是一种需要提前落地的技术架构（Optionality is not a posture. It is an architecture.）。
[14:26] Catalina Herrera: 这种中立、解耦的模块化设计，使企业既能充分享受前沿商业模型的红利，又能随时随地行使“拔掉插头换另一家”的主动权。
[20:30] Catalina Herrera: 企业无需扮演为外部供应商喂养数据的“租客”，也不应将核心竞争力拱手相让。
[20:30] Catalina Herrera: 你可以租用模型，逼迫它们发挥最大潜能；但必须掌控你做决定的方式。因为‘如何做决定’才是你成其为你、而非竞争对手的核心所在。
Source quotes
[0:04] Catalina Herrera: organizations rely heavily on rented intelligence that depends on permissions controlled entirely by third parties who can revoke them at any point.
[4:36] Catalina Herrera: relying on off-the-shelf, rented intelligence leaves teams helpless during critical moments when they need the AI most.
[4:36] Catalina Herrera: 72% to 74% of CIOs regret a major AI vendor decision made within the previous 18 months.
[6:43] Catalina Herrera: an enterprise's edge can never be the AI model itself, because competitors can easily rent the exact same underlying model.
[6:43] Catalina Herrera: Off-the-shelf intelligence raises the operational floor for everyone equally; relying solely on it makes an organization identical to its market rivals.
[6:43] Catalina Herrera: When enterprises build directly inside someone else’s proprietary stack, they risk handing over their core intellectual property and operational identity.
[11:13] Catalina Herrera: Organizations do not need to own the entire underlying infrastructure, but they must retain full ownership of the decision layer.
[11:13] Catalina Herrera: By decoupling business logic from supplier models, everything beneath the decision layer becomes completely modular and swappable.
[20:30] Catalina Herrera: Acknowledge Operational Realities: Accept that rented model access can be revoked, gated, or refused without notice.
[20:30] Catalina Herrera: Separate Decisions from Suppliers: Treat foundational models as interchangeable suppliers while retaining full ownership of the enterprise decision layer.
[20:30] Catalina Herrera: Protect Enterprise Intellectual Property: Ensure proprietary prompts, domain expertise, and business rules remain internal assets rather than locked within vendor platforms.
Original transcript
[0:04] All right, my name is Catalina Rueda. And I lead
the field engagements at Dataiku, and I spend my
[0:09] days inside Fortune 500 AI programs. The ones
that are going very well, breaking ground, but
[0:17] also the ones that are going sideways. So, today
I'm going to share a lot of the lessons learned,
[0:24] and we are going to talk about that trap. And
the most important thing is we're going to
[0:31] have a heads-up. This is happening. And I will
say that a lot of these IT leaders are walking
[0:39] into this trap with their eyes wide open.
So, let's have a flashback together. Ready?
[0:48] Where were you in June 12th when you heard about
this situation with the AI models? Can you feel
[0:54] it? You remember that day? Yeah, I'm seeing some
nodding. Yes, exactly. June 12th, what happened?
[1:01] A leading AI company received a governmental
directive requiring them to block the foreign
[1:08] nationals from accessing two of the most advanced
models that were available in the market.
[1:18] And the company was not able to identify foreign
versus not foreign, so they decided to shut down
[1:25] the model overnight. Do you remember that
day? Okay. What's the problem there? Well,
[1:34] think about it. Organizations around the
world that are building these workloads
[1:41] around technology, and suddenly the access
to the model disappeared because of a policy
[1:51] decision that was completely outside of their
control. You are renting that intelligence. You
[1:59] are renting that models. So, nobody call you, no
appeal, no SLA, no manager to call. It happened.
[2:11] 18 days later, the access was restored.
You have access to the model again. Yay!
[2:20] We can keep building pipelines
and deploy them to production.
[2:25] Uh-oh. Did you feel it? I am feeling it in my
stomach. Nothing changed. Nobody did anything
[2:34] to change the architecture they had behind the
scenes on June 11. And they went back to the same,
[2:43] redeployed, continued that in
production. And here is the trap.
[2:50] The structure behind the scenes didn't change.
And it depended on a permission that somebody else
[2:58] can revoke at any time. You are renting that
intelligence. I can feel my stomach already.
[3:06] Feels like anxious, okay. So, what is the
situation here? Your access is what it was before.
[3:16] And the worst thing, it depends on a permission
that somebody else controls or can revoke.
[3:26] Okay. So, that brings us to a moment of ooh, this
feels uncomfortable, yeah? So, this operational
[3:35] reality is what it takes to this frontier AI to
be deployed in production environments across the
[3:45] enterprise. And we have to treat it as any other
input variable that you have in your business, is
[3:54] Frontier AI? Yes, but it's yet another variable.
So, you have different variables. Availability
[4:04] moves with outages and now with policies
shifting. Price moves. We have consumption spikes.
[4:13] Vendor repricing is a variable. Performance moves
because today you have your best model. Tomorrow,
[4:22] that's best model is another one. It's a variable.
And guess what? Now we have a fourth dial.
[4:32] Permission. That was not there
before. Now we have permission.
[4:36] And let me show you how permission looks like in
practice. So, what happened a few weeks ago? Now,
[4:48] when attackers targeted Hugging Face this
month, the responders sent the attack logs
[4:55] to the hosted Frontier model for the analysis.
And what happened? The model refused. Why? Well,
[5:04] the guardrails of the model were working and were
unable to tell the good guys and the bad guys.
[5:12] So, when this team needed that model the most and
needed the help the most, that model said, "No."
[5:21] It was denied. That happens when you are renting
that intelligence and you don't have the control
[5:28] on when you have it available or not. And
this kind of, "No, I'm not going to help you."
[5:33] is not a one-off. It is actually echoing
across the C-suite. It's like, "Mhm,
[5:42] there are more things here that we need
to consider." And it's a starting to get
[5:47] worried, you know, when you lose control
overnight on your deployed production
[5:52] workflows because you are renting that
intelligence. So, in our 2026 survey,
[6:01] we have a 72 74% of CIOs actually saying that
they regret a major AI vendor decision made
[6:10] in the last 18 months. Why? Because they didn't
know what it was going to take to change and to
[6:21] update or to move things around. It's like that
you are stuck with that. And now you are facing
[6:27] the consequences of that when you lose access to
that overnight. So, they committed before they
[6:34] knew what moving will cost. And this makes you
think about something else. Like, okay, well,
[6:43] instead of an operational problem, let's
focus on an strategic one. And let's zoom out.
[6:52] And let's think. And I'm going to ask you, and
you can answer yourself, but you have to have
[6:57] an answer. And you have to ask yourself, what
is your company's competitive edge made of?
[7:07] And when I say this, what I mean is what
makes you different from your competitor?
[7:15] What makes you the one that the customer picks
instead of the competitor attributes far from you?
[7:25] So, and I'm going to tell you, that
is not the model. It can be the model
[7:32] because everyone can rent that model. So, it
cannot be that. Your edge cannot be the model.
[7:37] Your edge is the accumulated judgment that your
teams have been working into in the know-how,
[7:47] in the know-how of my business, how we do business
differently from our competitor, how we follow
[7:55] our own processes, how we collaborate to each
other, and how all of that is part of this AI.
[8:02] Where is that included? Because
everybody can rent the same model.
[8:07] So, how are you adding your differentiator into
this? And how you are encoding this judgement into
[8:14] this enterprise AI is the question that we need
to answer, right? And that is the risk of renting
[8:23] the edge because it's not up to you 100%. So, what
any company can rent, every competitor can rent.
[8:38] So, what makes you you unique? Where is that
workforce, those lessons learned, those know-how?
[8:46] When you have off-the-shelf intelligence, yes, we
are raising everybody's floor at the same time.
[8:54] But, what makes you faster also makes you the
same. Where is your differentiator coming from?
[9:03] What is making you the human empowered by AI
that is actually leveraging the technology
[9:12] instead of being a stuck overnight
because you rented the intelligence
[9:16] that you are using in your workflows. Mhm?
So, your AI advantage is not the model.
[9:25] If you think that your AI advantage is the
model, you have no AI advantage. Period.
[9:31] So, where does the differentiator leave? Where
do you do that? Where do you have that layer
[9:40] where your teams are actually, yes, using
the technology? Yes, phenomenal, generative,
[9:46] descriptive, predictive, generative, agentic. We
have all those those artifacts. The technology is
[9:51] there. But the reality of it is when you think
about where does your judgment live? Well,
[10:00] somebody did a prompt for my inventory months ago
or your pricing logic lives inside a UI that you
[10:10] don't control because it's a vendor or your risk
rules live somewhere a stack because it's another
[10:18] source. Everything is so siloed. Everything is
so siloed and you don't have that visibility,
[10:25] you don't have that control. So when your AI
artifacts where all of that is happening inside
[10:34] somebody else's a stack, you are losing that asset
that is the know-how of how you do your business,
[10:43] which is the differentiator that it makes you
you and not your competitor. Okay? So how do we
[10:52] fix this? How do we fix this because we are all
stuck here. Testing an agent here, an agent here,
[10:58] pilot here, pilot there. Yeah, working. But how
do you tie that with ROI? And how do you embrace
[11:04] your know-how, your business, your vision within
this ledger of technology? And that's where
[11:13] we need to think about this holistically. So
you don't need to own the entire AI stack. But
[11:22] you own you do you must own the decision layer.
And what is that decision layer? This is where
[11:32] your SMEs work. This is where you write that
first prompt and it came with something crazy,
[11:40] so you refine the prompt and it's now a little
better and now you go again and now it's a little
[11:46] better. That prompt, knowing what to ask, comes
from who? Whoever knows that business question.
[11:54] It has to come from your expert. It has to come
from your domain expert. And that knowledge-ment
[12:01] has to be embedded somewhere in that layer that
now is orchestrating the technology pieces. Yes,
[12:10] I'm going to have LLMs. Yes, I'm going to have
agents. Yes, I'm going to accelerate a lot of in
[12:14] my process. I'm going to innovate, but I'm
going to do it in a way where my signature
[12:20] is I included embedded because it's the way
that we do things different from our competitor.
[12:28] And when you think about the [clears throat]
options here, the one thing that you have to
[12:34] remember is that the model is a supplier. The
decision layer is the business and they are two
[12:45] different things. So when you own the decision
layer, everything beneath, it becomes modular.
[12:55] When the decision layer is yours, you can easily
swap a model. Oh, yeah, it doesn't work. Another
[13:02] model. Oh, I compare this answer with this LLM and
this LLM. This is charging me 10 times the amount
[13:10] of tokens for the same answer. Do I need that? I
need to compare to be able to make that decision.
[13:16] I need to realize that I have a whole ocean of
possibilities and that not necessarily everything
[13:23] works for us the way that works for them because
we have our own business, our own questions.
[13:30] So, it has to be easy to swap one another one.
So, June 12th can be a perfect moment to say, "If
[13:39] this goes down, then use the other one. Plan B,
Plan C." Why? Because it's a model. You swap it.
[13:48] Swap it as a model. So, remember the hugging face
situation. What happened? They bring the team.
[13:55] Let's investigate this. The model say, "No." And
they what they do? Oh, well, okay. Let's change.
[14:03] Let's use another model. And let's control
that model, and let's deploy it in our stack,
[14:10] and let's use it in our way. So, they shifted,
used the other model, and were able to finish the
[14:17] investigation, and to finish that job. Okay? So,
what is the highlight of this here? Optionality
[14:26] is not a posture. It is an architecture.
So, what we are talking about here is that
[14:34] is changing. Remember the old times? Everything
old is new again. We were pushing cloud first,
[14:42] cloud first, cloud first. Now, we're backwards.
No, no, no, no, no. Wait, wait, wait. Let's own
[14:48] the on-prem. Let's own the hardware. Let's own
the layers that we can control behind the scenes.
[14:56] What layers are those? Number one, data. I think
we have the opportunity to own this forever.
[15:04] Number two, compute. This one is the one that
is changing recently. Now, GPU is a commodity,
[15:11] and now you can have very good compute, and
you can have very good options to tweak that,
[15:18] and ensure that you maximize your options, because
in an infrastructure where you have 20 years ago
[15:26] engines versus last year platform engine, you
will see a difference into what happens behind the
[15:32] scenes. So, that compute ledger is controllable.
And now we have a new one. The intelligence layer.
[15:40] And now in the intelligence layer, this
is where you actually have those models,
[15:48] that intelligence, that artificial
intelligence, that is actually enhancing
[15:54] your questions, your business, what you do, but
at the same time, in a way that you can also
[16:01] optimize that layer. So, if you think about
it, there is a lot that you can control.
[16:07] If you want to have an architecture that is
agnostic to that change and to those policies
[16:13] and to somebody saying, "No, we're going to take
this model out tonight." So, you are not a stuck.
[16:19] And to make this real architecture modular and
evolving, right? We have uh now the demand for
[16:28] agentic AI. Therefore, that architecture,
modular architecture, has to evolve as well
[16:35] with that demand. And that is where Data
Co, the 575 Lab, fits with the two models
[16:43] that are now make all this architecture so easy
to maintain that agnostic position with it,
[16:51] right? So, now the privacy proxy,
keeping sensitive data uh safe by re- the
[16:59] by redacting it before anything leaves your
environment, and then you have Inspector
[17:05] that gives you a clear view into what the agents
are doing behind the scenes. And with this,
[17:10] we complement all the models from a very
agnostic architecture that is full of the pieces
[17:19] that allows you to maximize what you can do with
the technology that is available in the market.
[17:25] Renting the models when you want to, but
at the same time giving you the optionality
[17:32] to swap one, another one, and so on. So, how
do we go from here? First things first. People.
[17:42] Who are the ones that know which prompt to
write? Comes from who knows the business.
[17:49] You cannot tell somebody from IT to run a prompt
from somebody from the business. That's not
[17:54] happening. It has to come from the person that
understands what needs to accomplish. People.
[18:02] Second layer. Orchestration. Every artifact
that needs to happen for your AI to exist.
[18:09] Descriptive, predictive, machine learning,
generative, agentic, whatever piece in between
[18:17] has to be an artifact that is reusable and
that communicates with other artifacts removing
[18:26] the problem you had before of the silos
of technology because that is a still
[18:31] very prominent in enterprise. You have pockets of
technology that do not communicate to each other.
[18:38] So, if you have an orchestration layer where
you can actually ensure that these pockets of
[18:44] technology can communicate with each other. And
not only that, can be modular, replaceable. So, if
[18:52] tomorrow you have a better option within the data
or within the compute or within the intelligence,
[18:59] then you can tweak it and make it better. So,
it's a system that you can continue to optimize.
[19:05] Okay? And that is Dataiku. We have been doing this
for about 14 years. remain super agnostic to that
[19:14] ever-changing ecosystem, and allowing your people
to ask the questions the way that they know how,
[19:23] but at the same time ensuring that you own that
ledger and that you own that opportunity. So,
[19:30] who is doing this in the field? And what are good
examples? I would say Schlumberger, SLB. They land
[19:37] They actually landed at a modular architecture
here, optimizing their costs and AI dependencies
[19:44] as they are already showing ROI. So, this is a use
case, typical use case for the energy industry.
[19:52] They built a tender assessment system on Dataiku.
And this is actually evaluating over $10 billion
[19:59] in tenders and cutting analysis times from
days to 20 minutes. But even more important,
[20:06] this is happening in production. This is happening
at a scale, and in this is happening while they
[20:13] estate completely independent. So, moving the
needle, agnostic to the change in the ecosystem,
[20:21] controlling what they can control, right?
And to do that at production at a scale,
[20:30] there is a lot that needs to happen. And I will
say that Dataiku has proven experience on it.
[20:36] So, for the fifth time, we are leaders in the 2026
Magic Quadrant for AI platforms for data science
[20:43] and machine learning. This continues to change.
I will say the category has been going through
[20:50] cloud consolidation and generative
AI wave, and now a genetic wave.
[20:55] The vendor change channel of that, but
the position held, and we are still there.
[21:02] So, what are the three or four or five
things that I need you to go with today.
[21:08] To recap, first, this is happening. I mean,
we are the humans empowered by AI. As I said,
[21:15] I produce music, I do a lot of things, everything
with AI. We are the humans empowered by AI. But at
[21:20] the same time, we have to be very strategic about
this. And we have to ensure that you know that
[21:28] control what you can't control. So, you
can ask you are your access can be revoked,
[21:34] we learned that. The model can say no, we
learned that. You have to be agnostic to that.
[21:41] You can rent a model, push it, twist it,
try it, use the best AI models in the world,
[21:50] swap them, push them hard, but own how you decide.
Because the how you decide is what makes you you,
[22:04] and not your competitor. So, be the empowered
human that you are, but be strategic about it.
[22:11] Own where your business makes decisions, because
at the time that a model decides to shut down,
[22:17] it will. And you don't need to train a
landlord, because you will want to go at some
[22:24] time. And your business, the two tweaking of the
prompts, the changing the AI is the one asset
[22:32] that you need to keep with you. That is your
differentiator. All right? So, own your business,
[22:40] own the layer where you make the decision,
own what makes you you. All right, thank you.
