Operant Deploys Data-Centric Security to Clear 85% AI POC Failure Rate

Filling the POC-to-Production Gap With Operant Networks — Ai4

Ai4

Enterprise AI projects fail to reach production because traditional connection-based Zero Trust breaks under ephemeral multi-agent swarms, requiring cryptographic security embedded directly into data payloads.

AI-assisted summary of Filling the POC-to-Production Gap With Operant Networks — original source. Check the original for full context.

Source published: 2025-08-25

Detailed notes

## Background Published on YouTube by the Ai4 conference channel on August 25, 2025, this presentation features Keith Rose, CEO of Operant Networks. According to the session abstract, the talk explores why 85% of enterprise AI projects fail after the POC stage due to missing operational requirements and ROI measurement, introducing a dedicated pilot trust layer to bridge the gap to production.

## TL;DR Keith Rose analyzes why the vast majority of enterprise AI initiatives fail to reach production, arguing that the bottleneck stems from infrastructure, security, and governance challenges rather than flawed model outputs. Through an RFP automation case study, he introduces a 'crawl, walk, run' deployment framework that inserts a secure pilot phase to validate operational constraints and non-human identity. Rose details four core architectural criteria for pilot environments, emphasizing a strict separation between data and control planes. Finally, he outlines how ephemeral multi-agent swarms outgrow traditional connection-based Zero Trust, advocating instead for data-centric security powered by Named Data Networking (NDN) to embed cryptographic trust directly into network packets.

## Atmosphere & style The presentation carries a pragmatic, authoritative tone that speaks directly to the boardroom anxieties of enterprise technology leaders confronting high POC failure rates. Rose balances conversational candor—lightheartedly referencing 'grumpy' compliance gatekeepers—with sober, disciplined systems engineering and historical networking context. The talk feels like a reassuring, no-nonsense executive briefing that cuts through agentic AI hype to anchor complex deployments in battle-tested infrastructure realities.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=TLxd61MVdlU&t=4s) Roughly 85% of enterprise AI proofs of concept (POCs) fail to reach production due to operational hurdles and trust gaps rather than model failure. - [00:04](https://www.youtube.com/watch?v=TLxd61MVdlU&t=4s) Traditional POCs yield skewed, overly optimistic ROI estimates because they test model outputs in isolation from enterprise compliance and security constraints. - [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) Successful deployment requires a three-phase progression: Crawl (POC answering 'what'), Walk (Pilot answering 'how'), and Run (Production realizing full ROI). - [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) High-value agentic workflows, such as accelerating RFP turnaround from weeks to days, regularly stall when cybersecurity and compliance teams intervene over data access and legacy integration. - [04:05](https://www.youtube.com/watch?v=TLxd61MVdlU&t=245s) An enterprise AI pilot layer requires a friction-free internal deployment model and dedicated non-human identity authentication, since standard Zero Trust is built for humans. - [04:05](https://www.youtube.com/watch?v=TLxd61MVdlU&t=245s) A secure AI control plane must enforce dynamic inter-agent access rules while remaining entirely invisible and immutable to the AI agents operating on the data plane. - [04:05](https://www.youtube.com/watch?v=TLxd61MVdlU&t=245s) Early agent connectivity standards like Model Context Protocol (MCP) enable rapid prototyping but currently lack native, dynamic enterprise-grade security controls. - [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) Ephemeral, multi-agent swarms that dynamically summon peer agents render static perimeter firewalls and point-to-point connection security obsolete. - [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) Named Data Networking (NDN) shifts trust from network connections to the data itself by cryptographically signing, encrypting, and binding access policies directly to every data packet. - [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) Data-centric NDN architecture is already proven in mission-critical infrastructure, securing roughly 2% to 3% of US power generation including nuclear facilities.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=TLxd61MVdlU&t=4s) The AI POC Failure Metric and the Trust Gap

Keith Rose, CEO of Operant Networks, introduces the central dilemma confronting enterprise artificial intelligence initiatives: the well-documented statistic that 85% of AI proofs of concept (POCs) never reach production. While understandable given that the industry is still learning and experimenting in real time, Rose highlights how agonizing this reality is for AI leadership at Fortune 500 enterprises. Having to stand in front of a board of directors or a CEO to report that nearly nine out of ten AI initiatives fail to operationalize creates a "pretty uncomfortable situation."

To understand why this failure rate persists and how to fix it, Operant Networks deconstructed the problem into three fundamental pillars: 1. **Incomplete Business Data & Inaccurate ROI**: POCs evaluate model output in isolation rather than operational reality, leading to skewed return-on-investment calculations. 2. **Borrowing Proven Frameworks**: Examining established methodologies from cybersecurity and network architecture to address deployment challenges. 3. **Underlying Technological Shifts**: Identifying the emerging paradigms in AI interaction that make existing security approaches obsolete.

``` CRAWL (POC) WALK (Pilot) RUN (Production) ┌────────────────────────┐ ┌────────────────────────┐ ┌────────────────────────┐ │ Answers: WHAT │ │ Answers: HOW │ │ Full ROI Loop │ │ Model output & task │─▶│ Security, compliance, │─▶│ Scaled, trusted agent │ │ optimization │ │ & legacy integrations │ │ execution in enterprise│ └────────────────────────┘ └────────────────────────┘ └────────────────────────┘ ```

[01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) The RFP Scenario: Shifting from Model Problems to Trust Problems

To illustrate why promising POCs stall, Rose outlines a common enterprise workflow: an engineering or construction firm responding to complex bids and RFP proposals. Gathering cross-departmental subject-matter expertise for an RFP normally consumes three to four weeks and roughly 50 to 100 labor hours. When automated with multi-agent AI systems, that turnaround time can shrink to three days. For an enterprise handling dozens of RFPs every month, a standard POC yields astronomical projected economics.

However, the transition to production derails as soon as the project touches real-world enterprise constraints—such as bidding on a power plant in a heavily regulated industry. At this point, the "somewhat grumpy teams of cybersecurity and compliance" intervene with "Whoa, whoa, not so fast." These teams raise critical infrastructure and governance questions: how data flows across the network, who has authorization to view specific inputs, regulatory exposure, and how to extract data trapped in legacy systems. Consequently, what started as a model optimization exercise transforms into an infrastructure and trust bottleneck that traditional POCs completely fail to account for in their ROI models.

To bridge this gap, enterprises must adopt a structured "crawl, walk, run" deployment strategy: - **Crawl (POC Phase)**: Answers the *what*—demonstrating what AI capabilities can achieve for a specific functional task. - **Walk (Pilot Phase)**: Answers the *how*—validating how agentic systems operate within the existing enterprise environment across security boundaries, compliance requirements, and legacy integrations. - **Run (Production Phase)**: Combines the answers to *what* and *how* to produce an accurate ROI calculation that justifies enterprise rollout.

[04:05](https://www.youtube.com/watch?v=TLxd61MVdlU&t=245s) The Four Criteria for an Enterprise Pilot Trust Layer

Rose explains that solving the "how" requires a dedicated operational pilot environment—a secure sandbox or pilot trust layer designed specifically for testing AI agents against production-grade enterprise constraints. Instead of focusing merely on product features, Rose outlines four mandatory architectural criteria necessary to establish trust and safe experimentation inside enterprise networks:

1. **Frictionless On-Premises/Internal Deployment**: The sandbox must spin up within existing internal infrastructure without demanding massive platform upgrades, network overhauls, or complex reconfigurations. 2. **Non-Human Identity & Authentication**: Zero Trust frameworks were built around human identity verification (such as multi-factor authentication). AI agents are neither human nor traditional static software packages, requiring a new mechanism to verify that an autonomous agent is truly who it claims to be. 3. **Dynamic Inter-Agent Access Controls**: In multi-agent environments, explicit policies must govern what agents are permitted to say to one another as well as what data they can pass back to the end user. 4. **Strict Plane Separation**: A complete boundary must exist between the data plane (where information flows) and the control plane (which enforces trust, identity, and security rules). The control plane must remain entirely invisible to the AI so agents cannot alter security configurations or bypass constraints.

Rose references the Model Context Protocol (MCP) as a prime example of where these gaps appear. While MCP enables rapid prototyping between clients and servers, it currently lacks native production-grade security; changing trust rules often requires tearing down connections entirely rather than managing permissions dynamically.

[07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) The Evolution of Trust: Moving to Named Data Networking (NDN)

Rose places the current AI security challenge into a broader historical context, mapping out three distinct eras in enterprise trust architecture:

``` Era 1: Perimeter Defense Era 2: Connection Security Era 3: Data-Centric Security ┌──────────────────────┐ ┌──────────────────────┐ ┌──────────────────────┐ │ [ Secure Firewall ] │ │ [ Zero Trust VPN ] │ │ [ Signed Packet ] │ │ ┌──────────────┐ │ │ Client ──▶ Host │ │ ┌──────────────────┐ │ │ │ IP & Data │ │ ───▶ │ 1-to-1 encrypted │ ───▶ │ │Data + Trust Rules│ │ │ └──────────────┘ │ │ static connections │ │ └──────────────────┘ │ └──────────────────────┘ └──────────────────────┘ └──────────────────────┘ (Broken by Remote Work) (Broken by Agentic AI) (Self-Securing Data) ```

- **Perimeter Defense**: Historically, enterprise data was placed behind a protected physical and digital perimeter ("a wall around that"). - **Connection-Based Zero Trust**: Remote work accelerated by COVID shattered the network perimeter, forcing security teams to transition to point-to-point encrypted tunnels and Zero Trust connection policies. - **Data-Centric Trust (The AI Shift)**: Autonomous agentic systems render point-to-point connection models obsolete. When an AI agent autonomously summons four specialized peer agents to spin up an ad-hoc communication network for twenty minutes and then disbands, security teams lose control over the sheer volume and dynamic churn of connections.

To prevent security breakdowns in ephemeral agent swarms, trust must be embedded directly into the data itself. Operant Networks utilizes **Named Data Networking (NDN)**, an architectural paradigm developed in academia where every individual data packet is cryptographically signed, encrypted, and governed by embedded trust rules detailing who can access or modify it. Rose points out that this technology has already been hardened in high-consequence operational environments, securing approximately 2% to 3% of the total power generated in the United States, including nuclear power facilities.

Rose concludes that saving the 85% of AI POCs stuck in the "POC valley of death" requires three coordinated steps: adopting a data-centric security model where trust is bound to the payload, deploying a secure sandbox layer to test operational readiness in enterprise environments, and establishing an accurate ROI assessment that gives leadership the confidence to move to production.

## Notable quotes > Most AI projects never make it past the POC — not because the models fail, but because teams lack the tools to measure ROI and meet operational requirements. — Keith Rose

> [00:04](https://www.youtube.com/watch?v=TLxd61MVdlU&t=4s) pretty uncomfortable situation — Keith Rose

> [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) somewhat grumpy teams of cybersecurity and compliance — Keith Rose

> [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) Whoa, whoa, not so fast. — Keith Rose

> [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) crawl, walk, run — Keith Rose

> [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) a wall around that — Keith Rose

> [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) POC valley of death — Keith Rose

> If you're responsible for turning AI demos into deployed systems, this session will give you a clear roadmap for success. — Keith Rose

## Mentioned Keith Rose, Operant Networks, Ai4, Fortune 500, Model Context Protocol (MCP), Zero Trust, Named Data Networking (NDN)

## Video description Keith Rose, CEO, Operant Networks

Abstracts

Most AI projects never make it past the POC — not because the models fail, but because teams lack the tools to measure ROI and meet operational requirements. This talk will explore why real-world deployment is so difficult, and what’s missing from most efforts to move from POC to production. I’ll describe the missing “pilot” layer and how to create the right environment to test AI with real data, enforce guardrails, and evaluate ROI more accurately. If you're responsible for turning AI demos into deployed systems, this session will give you a clear roadmap for success.

[Source](https://www.youtube.com/watch?v=TLxd61MVdlU)

中文简报

Operant 部署以数据为中心的安全方案，破解 85% 的 AI POC 失败率难题企业 AI 项目难以进入生产阶段，是因为传统基于连接的零信任架构在瞬态多智能体集群下失效，亟需将密码学安全机制直接嵌入数据载荷中。 Keith Rose 分析了绝大多数企业 AI 项目未能进入生产阶段的原因，指出瓶颈主要源自基础设施、安全和治理方面的挑战，而非模型输出存在缺陷。通过一个 RFP 自动化的案例研究，他提出了“爬行、行走、奔跑”（crawl, walk, run）的部署框架，主张引入安全的试点阶段以验证运行约束和非人类身份。Rose 详细阐述了试点环境的四大核心架构标准，并强调数据平面与控制平面的严格分离。最后，他指出瞬态多智能体集群已超出传统基于连接的零信任架构的承载能力，进而提倡采用由命名数据网络（NDN）驱动的以数据为中心的安全方案，将密码学信任直接嵌入网络数据包中。约 85% 的企业 AI 概念验证（POC）因运营障碍和信任断层无法进入生产阶段，且孤立测试模型导致前期 ROI 估算普遍过于乐观。 (0:04) 企业 AI 落地需践行“爬（POC 明确目标）、走（Pilot 解决安全合规与系统集成）、跑（生产阶段实现全闭环 ROI）”的三阶段演进策略。 (1:26) 企业级 Pilot 信任层需具备四大架构要素：无摩擦内网部署、非人身份认证、动态 Agent 间访问控制，以及对 AI 隐形且不可篡改的控制平面与数据平面隔离。 (4:05) 自主多 Agent 集群使传统网络边界与点对点零信任连接失效，推动安全架构转向命名数据网络（NDN），将加密与信任规则直接绑定并嵌入数据包自身。 (7:00)

Reading summary

Keith Rose, CEO of Operant Networks, asserts that roughly 85% of enterprise artificial intelligence proofs of concept (POCs) fail to reach production because organizations evaluate model outputs in isolation while neglecting enterprise infrastructure and governance constraints. Rose argues that crossing this divide requires shifting from connection-focused security to data-centric networking and implementing dedicated pilot environments to validate operational feasibility. Bridging the Gap from Capability to Production (1:26) Enterprises frequently demonstrate substantial productivity gains during early AI trials. Rose highlights an engineering and construction workflow where multi-agent systems reduced request-for-proposal (RFP) turnaround times from three to four weeks (representing 50 to 100 labor hours) down to three days. However, these initiatives typically stall when confronting operational realities, such as bidding on regulated power facilities. At that stage, cybersecurity and compliance teams raise unresolved challenges concerning data ownership, network routing, and legacy system extractions, exposing the fact that initial return-on-investment (ROI) models failed to account for production-grade constraints. To bridge this gap, Rose advocates a three-stage deployment model. The initial 'Crawl' phase answers what AI can functionally achieve in an isolated task. The subsequent 'Walk' phase addresses how the system operates across enterprise security policies, compliance boundaries, and legacy integrations. Only after answering both questions can an enterprise enter the 'Run' phase, deploying autonomous agents at scale with an accurate, defensible ROI calculation. Core Requirements for Enterprise Pilot Sandboxes (4:05) Validating the operational feasibility of agentic workflows requires a secure pilot sandbox before full production rollout. Rose outlines four mandatory architectural criteria for this layer: frictionless deployment inside existing on-premises or internal environments without platform overhauls; non-human authentication capable of verifying autonomous agent identities rather than human users; dynamic access controls to govern peer-to-peer agent communications and user-facing data handoffs; and strict architectural separation between the data plane and the control plane. Rose emphasizes that the control plane must remain entirely invisible to autonomous agents to prevent them from modifying or bypassing security configurations. He notes that emerging standards such as the Model Context Protocol (MCP) facilitate rapid prototyping between clients and servers, but lack production-grade security, often requiring entire connection teardowns to alter permissions rather than adjusting policies dynamically. The Shift to Data-Centric Security via NDN (7:00) Rose contextualizes current AI integration hurdles through three eras of network security: perimeter defense, connection-based Zero Trust, and data-centric trust. While Zero Trust point-to-point tunnels solved challenges created by remote work, Rose contends they cannot support autonomous AI swarms. Because AI agents can spontaneously summon multiple peer agents to form transient, ad-hoc networks for short durations and then disband, the dynamic churn overwhelms connection-oriented security models. To solve this vulnerability, Operant Networks employs Named Data Networking (NDN), an architectural approach where cryptographic signatures, encryption, and access policies are bound directly to each individual data packet rather than to the network conduit. Rose reports that NDN is already hardened in critical infrastructure, securing an estimated 2% to 3% of United States power generation, including nuclear power installations. Rose concludes that overcoming the enterprise AI failure rate requires adopting data-centric security architectures like NDN, validating real-world operational constraints within dedicated pilot sandboxes, and establishing realistic ROI metrics that give corporate leadership the confidence to move beyond experimentation into production.

精炼阅读摘要

Operant Networks首席执行官Keith Rose指出，高达85%的企业AI概念验证（POC）未能进入生产阶段，核心瓶颈在于现存安全与网络架构无法适应自主智能体交互；企业必须通过引入以数据为中心的安全模型与沙箱信任层，填补从概念验证到生产落地的鸿沟。 AI概念验证停滞与投资回报率失真 (0:04) Rose提出，目前财富500强企业中有85%的AI概念验证项目无法落地，这给企业AI领导层带来了严峻的汇报压力。概念验证停滞的关键原因在于企业仅在隔离状态下评估模型输出，而忽略了现实业务约束，从而得出了严重失真的投资回报率（ROI）预估。以工程承包或建筑行业处理复杂招标文件（RFP）为例，传统跨部门协同通常需要3到4周及50至100个工时，而多智能体系统可将周期压缩至3天，在概念验证阶段展现出极高的经济效益。然而，一旦项目涉及高度受管制的现实环境（如核电站竞标），网络安全与合规团队便会因网络数据流向、权限控制、合规暴露以及遗留系统数据提取等问题阻断项目上线。原本针对模型任务的优化测试，随即演变为传统ROI模型未能预料的基础设施与信任瓶颈。渐进式落地路径与四项沙箱标准 (4:05) 为打破僵局，企业需要实施“爬、走、跑”三阶段策略。“爬行（POC阶段）”仅回答“是什么”，即验证AI完成特定功能任务的能力；“行走（试点阶段）”重点解决“如何做”，即验证智能体如何在安全边界、合规要求和遗留集成环境中实际运行；“跑步（生产阶段）”则综合两者，得出具有现实说服力且能够指导规模化部署的精准ROI。Rose强调，解决“如何做”需要在企业内部构建具备操作级的试验沙箱或试点信任层。该信任层必须满足四项技术准则：首先，支持在现有内部基础设施中无缝部署，无需进行破坏性的网络重构；其次，建立非人类身份验证机制，因为传统的针对人类多因素认证的零信任体系无法适配自主智能体；第三，提供动态的智能体间访问控制，严格规范智能体间的信息交互和向用户传递的数据；第四，实现控制平面与数据平面的绝对隔离，使控制平面在AI智能体面前完全不可见，防止智能体篡改安全策略或绕过约束。Rose补充称，如Model Context Protocol（MCP）等工具虽然便于快速原型开发，但目前缺乏原生生产级安全性，调整信任规则时往往需要中断连接。网络信任演进与以数据为中心的安全 (7:00) Rose从网络信任架构的历史演进解释了当前AI安全挑战的根源。第一代企业安全依赖“边界防御”，通过在物理和数字边界建立防火墙保护内部数据；第二代转为“基于连接的零信任”，远程办公的兴起打破了网络边界，迫使企业采用端到端加密通道和静态连接策略。然而，自主智能体系统的普及正在摧毁连接安全模式。当一个主智能体能够随时自主调用四个对等智能体组建通信时长仅20分钟的临时网络并随后解散时，动态且庞大的连接流将让安全团队彻底失去掌控。为此，信任机制必须直接嵌入数据本身。Operant Networks采用起源于学术界的“命名数据网络（NDN）”架构，使每个数据包都经过独立的加密签名，并内嵌访问与修改规则。据Rose说明，NDN技术已在关键基础设施中经受验证，目前正保障着美国约2%至3%的总发电量，其中包括多座核电设施的安全运行。 Rose总结认为，要拯救深陷“死亡之谷”的85%的AI概念验证项目，企业必须协同推行三项举措：采用将信任直接绑定在载荷上的以数据为中心的安全架构、部署生产级沙箱验证运行可行性，以及建立反映真实基础设施成本的ROI评估，从而为企业高层推进全面生产化提供决策支撑。

Detailed notes — zh

## Background 该内容由行业峰会频道 Ai4 于 2025 年 8 月 25 日发布在 YouTube 平台。视频记录了 Operant Networks 首席执行官 Keith Rose 的主题演讲，聚焦探讨为何绝大多数企业 AI 项目难以跨越 POC（概念验证）阶段，并详细介绍了如何通过构建缺失的“试点（Pilot）”信任层、利用真实数据测试与安全护栏来精确评估 ROI 并实现正式投产。 ## TL;DR 演讲直击企业级 AI 高达 85% 的 POC 无法落地的行业痛点，指出失败根源并非模型算法本身，而是传统验证流程忽略了合规、网络安全与遗留数据接入等隐性基建开销，导致理论 ROI 严重失真。Keith 强调企业必须在“爬（POC 验证功能 What）”与“跑（投产）”之间补齐“走（Pilot 试点验证落地架构 How）”这一关键层级。面对多智能体协同导致的动态网络连接失控，他提倡采用已在核电与电网验证的 NDN（命名数据网络）底层架构，实现以数据本体为中心的原生加密签名与权限绑定。最后，他总结了通过控制面与数据面强隔离、非人类身份鉴权及精准成本核算来挽救 AI 项目跨越“死亡谷”的实施路径。 ## Atmosphere & style 整场演讲风格冷静干练、直击要害且充满工业级的严谨质感。Keith 以财富 500 强 AI 负责人在董事会面前的汇报窘境切入，瞬间拉满了直面残酷现实的紧迫感与行业共鸣。叙事不带浮躁的 AI 泡沫狂热，而是借助核电工程标书、MCP 安全缺陷以及关键基础设施调度等高要求场景，层层剥离出从算法 Demo 到生产系统之间的底层网络信任鸿沟，带给观众一种从虚幻预期回归硬核工程落地的清醒与笃定。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=TLxd61MVdlU&t=4s) 全行业高达 85% 的 AI POC 无法进入生产环境，核心瓶颈在于传统 POC 测算的 ROI 极不精确，漏算了大量后续工程与合规成本。 - [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) 以复杂的工程 RFP 标书为例，AI Agent 理论上可将周期从数周压缩至 3 天，但进入生产时会立刻撞上数据安全、合规与遗留系统的基础设施壁垒。 - [02:57](https://www.youtube.com/watch?v=TLxd61MVdlU&t=177s) 企业引入 AI 需遵循“爬、走、跑”策略：POC 仅属于回答“What”的爬行期，企业亟需处于“走”阶段的 Pilot（试点）层来回答“How”的工程落地问题。 - [04:05](https://www.youtube.com/watch?v=TLxd61MVdlU&t=245s) 构建生产级 Pilot 安全试验环境需满足零侵入部署、机器实体身份认证、Agent 间细粒度访问控制以及控制面与数据面彻底隔离四项准则。 - [04:05](https://www.youtube.com/watch?v=TLxd61MVdlU&t=245s) 当前主流的 MCP（Model Context Protocol）缺乏细粒度安全机制与动态变更能力，信任规则僵化，极难直接上线生产环境。 - [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) 在多 Agent 协同且动态创建/销毁临时子网络的场景下，传统的“网络边界”和“点对点零信任连接”安全模型将彻底失效。 - [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) Operant Networks 采用 NDN（命名数据网络）架构，使每个数据包自带数字签名、加密载荷与权限规则，实现以数据为中心的安全模型。 - [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) NDN 技术并非纯理论架构，已被验证并应用于承载全美 2%–3% 的电网与核电站发电调度数据中。 - [09:07](https://www.youtube.com/watch?v=TLxd61MVdlU&t=547s) 拯救陷入“死亡谷”的 AI 项目需要三步闭环：确立数据级信任、部署受控 Pilot 沙箱，并输出包含合规与基建成本的真实 ROI。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=TLxd61MVdlU&t=4s) 85% 的 AI POC 死亡率与投资回报率（ROI）信息缺失 Operant Networks 的 CEO Keith Rose 开场指出，整个行业普遍流传着一个令人震惊的数据：“85% 的 AI POC（概念验证）从未真正进入生产环境”。起初这一数字让人错愕，但退一步看，AI 行业目前仍处于实时摸索与大规模实验的早期阶段，高失败率在某种程度上是探索期的客观反映。然而，对于大型企业或财富 500 强的 AI 负责人而言，向董事会或首席执行官汇报这一指标无疑是极其尴尬和难堪的境地。为了破解这一困局，Operant Networks 团队将问题拆解为三个核心维度：首先，将庞大的行业痛点聚焦为具体的业务瓶颈——他们发现，大多数 POC 无法进入下一阶段，根本原因在于其未能反映出精准的 ROI（投资回报率）数据，导致决策依据不完整；其次，能否借鉴其他相邻行业或已有方案来攻坚；最后，是否存在某种底层的根本性技术范式转移，使得该问题呈现出前所未有的特质。 - 行业痛点：高达 85% 的企业级 AI POC 停滞于概念验证阶段，沦为未落地的“死项目”。 - 核心病灶：传统 POC 测算得出的 ROI 极不精确，漏算了大量后续工程化成本，导致信息失真。 --- ### [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) 复杂的工程投标用例：从“模型优化”到“信任与基建危机” Keith 举了一个大型工程或建筑公司的典型业务场景：撰写一份高度复杂的投标书或 RFP（采购需求建议书）。在传统流程下，这项工作需要企业内部多个跨领域专家深度介入，耗时往往长达 3 到 4 周，累计投入 50 至 100 个工时。若引入 AI Agent（智能体系统）协同处理，整个交付周期有望大幅缩减至短短 3 天。对于每月需处理数十份标书的企业而言，这在经济账面上展现出极具诱惑力的 ROI，管理层自然倾向于全力推进。然而，一旦该项目涉及高度受监管的敏感领域（例如承建核电站或传统发电厂），网络安全与合规团队便会立刻介入并强行踩下刹车。安全团队会发出严肃质询：“AI Agent 在网络中四处调用与传输数据，是否触犯合规红线？”、“谁能真正访问这些数据？”此外，大量核心工程数据仍封存在老旧的遗留系统（Legacy Systems）中，如何安全抽取同样悬而未决。瞬息之间，原本看似单纯的“模型与算法优化问题”，演变成了一场棘手的“信任与底层基础设施危机”；而解决这些隐形成本在最初的 POC 算账中根本未被纳入。 - 传统工作流瓶颈：编制大型 RFP 标书需要 50–100 工时、耗时 3–4 周。 - 引入 Agentic AI 的理想效果：理论上可将交付周期压缩至 3 天，展现出极佳的单点理论 ROI。 - 生产落地的现实障碍：合规审查、网络安全权限控制、遗留系统数据整合等底层基础设施开销，彻底打破了早期的成本估算。 --- ### [02:57](https://www.youtube.com/watch?v=TLxd61MVdlU&t=177s) 从“爬、走、跑”策略看 POC 与 Pilot（试点）的职责分工在追求极致迭代速度的 AI 浪潮下，人们往往忽视了大型企业引入新技术的传统演进路径——“爬、走、跑”（Crawl, Walk, Run）。Keith 强调，企业级 AI 部署必须重新拾起这种稳健的阶梯式方法。单纯的 POC 仅仅处于“爬”的阶段，它只回答了“What”层面的问题：AI 到底能在该具体场景中发挥什么功能？但从 POC 直接跳跃到大规模投产（“跑”）必然断层。企业真正缺失的是处于“走”阶段的 Pilot（试点层）。Pilot 阶段聚焦于回答“How”的问题：AI Agent 或智能体集群究竟该如何在真实的企业生产环境中运转？这涵盖了安全隔离、合规审计以及对底层遗留数据源的接入方式。只有把“What（功能价值）”与“How（落地基建成本）”两项要素叠加考量，企业才能得出真正全面、准确的 ROI 预测，从而跨越横亘在 POC 与生产环境之间的深渊。 ```mermaid flowchart LR accTitle: AI 项目从验证到投产的三阶段演进流程 subgraph Crawl["1. 爬行阶段 (Crawl)"] direction TB A["POC 概念验证"] --> B["回答 What：<br/>AI 具备哪些业务能力？"] end subgraph Walk["2. 行走阶段 (Walk)"] direction TB C["Pilot 信任试点层"] --> D["回答 How：<br/>安全、合规与遗留数据如何落地？"] end subgraph Run["3. 奔跑阶段 (Run)"] direction TB E["生产环境上线"] --> F["精确可信的 ROI<br/>全面规避合规与基建风险"] end Crawl --> Walk --> Run ``` --- ### [04:05](https://www.youtube.com/watch?v=TLxd61MVdlU&t=245s) 构建生产级测试环境的四项准则与控制面/数据面分离针对 Pilot 阶段的空白，Operant Networks 推出了一款被定义为“安全沙箱（Secure Sandbox）”或“试点信任层（Pilot Trust Layer）”的技术架构。Keith 指出，比起特定产品，更重要的是在企业内网构建一个可信安全试验环境时必须满足的四项基本准则： 1. **零侵入快速部署**：该环境必须能够无缝嵌入企业现有内部基础设施，无需对存量系统进行大规模升级、改动或繁琐的重新配置。 2. **非人实体的双因素与身份认证**：传统“零信任（Zero Trust）”理念已严重泛化，且其核心设计是针对“人类用户登录系统”。面对 AI 这类“既非人类、又不同于传统静态软件”的全新实体，必须解决机器身份认证（证明它确实是宣称的模型/Agent）的技术难题。 3. **Agent 间动态细粒度访问控制**：在多 Agent 协同体系中，必须建立严格的访问控制边界，明确规定智能体之间可以交互什么、不可以传递什么，以及它们最终能向发起提问的人类反馈何种级别的信息。 4. **数据平面与安全控制平面强隔离**：数据流转通道（Data Plane）与负责认证、安全策略下发的控制平面（Control Plane）必须彻底剥离。至关重要的是，“AI 必须对控制平面完全隐形”，绝不允许 AI 自行向上触碰或接管底层安全策略。 Keith 现场拿当前爆火的 MCP（Model Context Protocol，模型上下文协议）作为典型痛点进行佐证：许多企业正兴奋地测试 MCP，但很快会撞墙——MCP 几乎没有内建安全机制，极难直接上线生产。在当前机制下，MCP 的信任规则处于“配置后即固化（set it and forget it）”的僵化状态；一旦客户端与服务端的交互逻辑需要微调，运维人员只能被迫将整个系统彻底下线，重新修改信任规则后再重新拉起。 - 零信任在 AI 时代的局限：原有的基于人类账号的鉴权体系无法直接套用于自主通信的 AI 智能体集群。 - MCP 生产落地软肋：缺乏细粒度安全机制与动态策略变更能力，任何权限调整都需要关停重启服务。 --- ### [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) 信任边界的历史演变：从网络连接到 NDN（命名数据网络） Keith 梳理了安全信任架构的代际更迭。最初的模式是“边界防御”，即将所有机密数据与核心资产放置于严密的内网中，周围筑起坚固的防火墙边界；疫情暴发后，远程办公常态化打破了物理边界，行业被迫演变为基于“零信任的点对点安全连接”（Secure Connections），这也是当下企业试图约束 AI 的主流手段。然而，Keith 尖锐地指出，在真正的 Agentic AI 时代，基于连接的防护体系必将彻底崩溃。试想这样一个场景：一个 AI Agent 在执行任务时瞬间拉来 4 个智能体伙伴协同，动态生成一个临时子网络，该网络仅存在 20 分钟便销毁无踪；紧接着又有成百上千个类似的瞬态网络在系统中不断涌现。在这种极度动态、连接爆炸的环境下，安全人员根本无法管理连接状态。因此，底层范式必须再度下沉一级——直接聚焦于“安全数据本身（Trusted Data）”。为解决这一难题，Operant Networks 引入了一项源自学术界的底层网络技术：NDN（Named Data Networking，命名数据网络）。NDN 彻底重构了数据存储、传输与安全机制——在 NDN 架构中，每一个独立的数据包本身都自带数字签名、加密载荷以及细粒度的信任规则（严格限定了谁可以对该数据执行何种操作）。该技术不仅停留在理论阶段，此前在能源行业已被验证多年，当前全美约 2% 到 3% 的电网发电系统（包括核电站）都依托该架构运行，如今被正式引入 AI 机器间协同通信场景。 ```mermaid flowchart TD accTitle: 企业安全信任架构的三代演进路径 subgraph Gen1["第一代：物理边界防御"] G1["集中式内网机房"] --- W1["静态网络外墙 / 防火墙"] end subgraph Gen2["第二代：点对点连接防护"] G2["远程办公 / 零信任连接"] --- W2["动态点对点加密通道<br/>(AI 高并发下必将失控)"] end subgraph Gen3["第三代：以数据为中心 (NDN)"] G3["数据包级自包含安全"] --> D1["数据自带签名"] G3 --> D2["数据自带加密"] G3 --> D3["内嵌权限控制规则"] end Gen1 --> Gen2 --> Gen3 ``` - 信任模型演进轨迹：企业网络边界防护 $\rightarrow$ 点对点安全连接（零信任通道） $\rightarrow$ 动态自签名数据层（NDN）。 - NDN 工业应用背景：已在极其苛刻的核电及关键能源基础设施中承载 2%–3% 的全美发电调度数据。 --- ### [09:07](https://www.youtube.com/watch?v=TLxd61MVdlU&t=547s) 拯救陷入死亡谷的 AI 项目：三步闭环在演讲尾声，Keith 重新回应了如何将 85% 挣扎在“死亡谷（Valley of Death）”的企业 AI 项目挽救上岸。他总结了贯通 POC 到生产环境的完整闭环框架： 1. **确立以数据为中心（Data-Centric）的底层信任**：摆脱脆弱的动态连接追踪，直接让安全策略与数据包本体绑定，实现细粒度的自包含信任。 2. **部署隔离的生产级 Pilot 信任层**：在不侵入改造旧有基建的前提下，通过数据面与控制面的物理隔离，为 AI 构筑能够直接触达真实业务数据的受控沙箱。 3. **输出真实精准的可信 ROI**：将合规、遗留系统适配与安全维护成本完全显性化，向管理层与董事会交出一份具备充分确定性的投资回报测算，推动业务顺利从实验跨入生产交付。 Keith 总结道，唯有将功能层面的“What”与工程落地的“How”有机拼合，企业才能彻底跨越 POC 阶段的信任断层，让 AI 技术真正落地生根。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=TLxd61MVdlU&t=4s) 85% 的 AI POC（概念验证）从未真正进入生产环境 — Keith Rose > [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) AI Agent 在网络中四处调用与传输数据，是否触犯合规红线？ — 安全与合规团队 > [01:26](https://www.youtube.com/watch?v=TLxd61MVdlU&t=86s) 原本看似单纯的“模型与算法优化问题”，演变成了一场棘手的“信任与底层基础设施危机” — Keith Rose > [02:57](https://www.youtube.com/watch?v=TLxd61MVdlU&t=177s) 单纯的 POC 仅仅处于“爬”的阶段，它只回答了“What”层面的问题：AI 到底能在该具体场景中发挥什么功能？ — Keith Rose > [02:57](https://www.youtube.com/watch?v=TLxd61MVdlU&t=177s) Pilot 阶段聚焦于回答“How”的问题：AI Agent 或智能体集群究竟该如何在真实的企业生产环境中运转？ — Keith Rose > [04:05](https://www.youtube.com/watch?v=TLxd61MVdlU&t=245s) AI 必须对控制平面完全隐形，绝不允许 AI 自行向上触碰或接管底层安全策略。 — Keith Rose > [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) 在真正的 Agentic AI 时代，基于连接的防护体系必将彻底崩溃。 — Keith Rose > [07:00](https://www.youtube.com/watch?v=TLxd61MVdlU&t=420s) 每一个独立的数据包本身都自带数字签名、加密载荷以及细粒度的信任规则 — Keith Rose > [09:07](https://www.youtube.com/watch?v=TLxd61MVdlU&t=547s) 唯有将功能层面的“What”与工程落地的“How”有机拼合，企业才能彻底跨越 POC 阶段的信任断层，让 AI 技术真正落地生根。 — Keith Rose ## Mentioned Keith Rose, Operant Networks, Ai4, POC (概念验证), Pilot (试点), Agentic AI, Zero Trust (零信任), MCP (Model Context Protocol), NDN (Named Data Networking, 命名数据网络) ## Video description Keith Rose, CEO, Operant Networks Abstracts Most AI projects never make it past the POC — not because the models fail, but because teams lack the tools to measure ROI and meet operational requirements. This talk will explore why real-world deployment is so difficult, and what’s missing from most efforts to move from POC to production. I’ll describe the missing “pilot” layer and how to create the right environment to test AI with real data, enforce guardrails, and evaluate ROI more accurately. If you're responsible for turning AI demos into deployed systems, this session will give you a clear roadmap for success. [Source](https://www.youtube.com/watch?v=TLxd61MVdlU)

Source quotes

[0:04] Keith Rose: 85% 的 AI POC（概念验证）从未真正进入生产环境 [1:26] 安全与合规团队: AI Agent 在网络中四处调用与传输数据，是否触犯合规红线？ [1:26] Keith Rose: 原本看似单纯的“模型与算法优化问题”，演变成了一场棘手的“信任与底层基础设施危机” [2:57] Keith Rose: 单纯的 POC 仅仅处于“爬”的阶段，它只回答了“What”层面的问题：AI 到底能在该具体场景中发挥什么功能？ [2:57] Keith Rose: Pilot 阶段聚焦于回答“How”的问题：AI Agent 或智能体集群究竟该如何在真实的企业生产环境中运转？ [4:05] Keith Rose: AI 必须对控制平面完全隐形，绝不允许 AI 自行向上触碰或接管底层安全策略。 [7:00] Keith Rose: 在真正的 Agentic AI 时代，基于连接的防护体系必将彻底崩溃。 [7:00] Keith Rose: 每一个独立的数据包本身都自带数字签名、加密载荷以及细粒度的信任规则 [9:07] Keith Rose: 唯有将功能层面的“What”与工程落地的“How”有机拼合，企业才能彻底跨越 POC 阶段的信任断层，让 AI 技术真正落地生根。

Source quotes

Keith Rose: Most AI projects never make it past the POC — not because the models fail, but because teams lack the tools to measure ROI and meet operational requirements. [0:04] Keith Rose: pretty uncomfortable situation [1:26] Keith Rose: somewhat grumpy teams of cybersecurity and compliance [1:26] Keith Rose: Whoa, whoa, not so fast. [1:26] Keith Rose: crawl, walk, run [7:00] Keith Rose: a wall around that [7:00] Keith Rose: POC valley of death Keith Rose: If you're responsible for turning AI demos into deployed systems, this session will give you a clear roadmap for success.

Original transcript

[0:04] How you doing? I'm Keith Rose um CEO of [0:07] a company called Operant Networks. Um as [0:09] we said this is talk is called filling [0:12] the PC to production trust gap. Now [0:16] everyone has seen the statistic 85% of [0:18] AI PC's never make it to production. Um [0:22] at first it was like shocking but when [0:24] you take a step back it's like okay this [0:26] this industry we're all learning in real [0:29] time and experimenting so a lot of that [0:31] is reflected in this number but if you [0:33] are the head of AI at a large enterprise [0:35] a fortune 500 company and you have to [0:38] report this metric to your board or to [0:40] your CEO that is a pretty uncomfortable [0:42] situation. [0:44] So our team started thinking about how [0:46] do you address this problem? How might [0:48] you, you know, improve your numbers? And [0:50] we kind of broke it down into three [0:52] parts. Number one, yes, that's a big [0:53] broad industry problem. How do you uh [0:56] bring this down to a specific business [0:59] uh painoint? And we we landed on the [1:02] fact that these PC's do not reflect [1:04] accurate ROI information and therefore [1:06] it's incomplete information. Secondarily [1:08] to solve the problem, are there any [1:10] other adjacent industries or approaches [1:12] that we can borrow from to kind of solve [1:14] the problem? And then third would be are [1:17] there any uh underlying fundamental [1:19] technology shifts that make this problem [1:21] different or new in a in a unique way. [1:23] So we'll touch on all three of these [1:24] here. [1:26] So this is a use case that we've seen [1:28] actually two speakers ago has a solution [1:30] to this product. Um, but this is a use [1:32] case that we've seen is that a typical [1:34] engineering or construction company [1:36] that's filling out a large complex bid [1:38] or an RFP proposal that requires a lot [1:42] of inputs from different uh expertise [1:44] within your organization. That process [1:46] could easily take 3 or 4 weeks, maybe 50 [1:49] to 100 hours of work. Um, and if you [1:51] deployed AI agents at this problem, you [1:54] could potentially turn this uh project [1:56] around in say 3 days. Um, and that's [1:58] significant. So for a company of this [2:00] size, say they're doing dozens of these [2:02] every month, um that's a significant [2:04] project where if you ran that P, you're [2:06] saying there's a good ROI, there's good [2:07] economics there. Let's move forward. But [2:10] of course, say maybe that RFP, that [2:13] construction project was to build a [2:15] power plant, a highly regulated [2:17] industry. So then you have the somewhat [2:20] grumpy teams of cyber security and [2:22] compliance and they say, "Whoa, whoa, [2:24] not so fast." um you have a lot of AI [2:26] agents that are be you know there's [2:27] going to be data flying around our [2:29] network how does this does this violate [2:30] any regulatory concerns so cyber [2:32] security teams are saying hey who has [2:34] access to this data um and then [2:36] potentially a lot of this data is hosted [2:38] on legacy systems how do we actually [2:39] extract that data so suddenly you've [2:42] turned this P from what was a [2:44] optimization or model problem into now a [2:48] trust and infrastructure problem um [2:52] and those components often aren't [2:54] included in the ROI. So, how do you go [2:57] about thinking about this? And I know in [2:59] this industry AI, we move very very fast [3:01] at a break neck pace. But historically, [3:05] large scale tests, experiments at [3:08] enterprises would usually take a crawl, [3:11] walk, run approach to trying new [3:12] technology. And this industry might want [3:15] to take a step back and do that when [3:17] we're talking about large enterprises. [3:18] Um, where the PC [3:21] answers the what question. what can AI [3:24] do for this particular use case? But [3:26] then the second question would be the [3:28] walk phase or the pilot phase which is [3:31] the how. How does this AI agent or this [3:34] agentic system work in my enterprise [3:37] environment? specifically with regards [3:39] to security, compliance, integration [3:42] with legacy data sources. And we think [3:44] that answering both the what and the how [3:48] question now gives you a very clear [3:50] understanding and an accurate ROI [3:52] prediction. And this is how you'd be [3:53] able to move things beyond PC into [3:56] production. [3:57] So how do we do this? Are there any [4:00] other industries or approaches that we [4:01] can kind of borrow from to uh to solve [4:04] this problem? [4:05] And so what we found is a product that [4:07] our company launched actually uh earlier [4:09] this week at this event. Um we call it a [4:12] secure sandbox or a pilot trust layer. I [4:15] mean there's many different ways of [4:16] calling this but less about the product [4:18] and more about the approach we went of [4:20] deciding what is the criteria required [4:23] to instill a trusted secure layer for [4:25] experimentation and testing inside an [4:28] enterprise. Um and we came up with these [4:30] four criteria. Number one is that this [4:32] environment needs to be able to be [4:34] popped up inside your in uh internal [4:37] infrastructure without having to do [4:39] massive upgrades or changes or [4:40] configuration to your existing systems. [4:43] The next two points borrow heavily from [4:46] um the security world and and the ideas [4:49] and concepts around what's called zero [4:50] trust if you're familiar with that. Um [4:53] zero trust these days doesn't really [4:55] mean much. It's more of a philosophy [4:57] than a tech because that term has been [4:59] highly diluted. Um but the two key [5:01] principles and also zero trust is meant [5:03] for people logging into systems. So when [5:06] you have non- people like AI um which is [5:09] not quite human but not also a [5:11] traditional piece of software, how do [5:13] you do two factor authentication? How do [5:15] you prove that AI thing is who they say [5:17] they are? And when you have multiple of [5:19] these things talking to each other, how [5:21] do you in uh implement access controls? [5:24] What exactly you're allowed to say to [5:26] each other? and then also back to the [5:28] human that's actually asking the [5:29] question. So um those two things need to [5:32] be solved. And then the last one which I [5:34] think is key for um experimentation in a [5:36] production environment is the fact that [5:39] you separate what we call the data [5:41] plane. So this is where all the data and [5:42] information is flowing with the secure [5:44] control plane. So the secure control [5:47] plane is really the things that are [5:48] implementing things like authentication, [5:50] security, trust rules. And you want to [5:52] make sure that uh I'm bringing up this [5:54] word trust because many many panels you [5:56] go to at this conference bring up trust. [5:59] And it's very clear to me that most [6:00] people don't yet trust AI in their [6:03] production environment. So you want to [6:04] make sure that in your control plane the [6:07] AI [6:09] it's invisible to the control plane. So [6:11] the AI can never reach up and take [6:13] control of the security. Um and so this [6:16] is an example that uh we built for MCP. [6:18] I'm sure many people here at the [6:20] conference and the audience are [6:21] experimenting with MCP. If you are, [6:23] you're going to realize very quickly [6:25] that there is no security and it's very [6:27] difficult to implement MCP in a [6:29] production environment. Um, just the [6:31] nature of changing a trust rule. It's [6:34] like right now it's like uh set it and [6:36] forget it. This thing can talk, you [6:37] know, a server can talk, a client can [6:39] talk to a server and you can never [6:41] change that interaction without taking [6:42] it all down, changing the the uh trust [6:45] rules of security and then bringing it [6:46] back up. Um anyway, so this is a great [6:49] example of things we've used. So why is [6:51] this difficult? I mean it sounds [6:52] reasonable. Okay, create a secure [6:54] environment within your enterprise [6:55] network. Um is there something new or [6:58] different here? [7:00] Um and I'd say yes, there is something [7:02] majorly different here. So um when we [7:04] talk about trust and the teams and [7:07] departments that are in charge of [7:08] implementing trust, um we saw this [7:12] happen once before. So the old way of [7:14] implementing trust is you put all your [7:16] secure IP and data and privacy in a [7:18] secure place and you build a wall around [7:21] that a secure perimeter and then the [7:23] last major inflection point in trust [7:25] happened with co where everyone started [7:27] working remotely from home and these [7:29] professionals basically lost control of [7:31] the perimeter. So then they move to [7:33] secure connections or zero trust [7:35] connections where you have all these [7:36] onetoone secure connections and that is [7:39] where our industry is right now and [7:41] that's what people are trying to do with [7:43] AI. However, I'd argue that very soon [7:46] when you really start deploying a true [7:48] agentic AI where you have an AI agent [7:51] that can call in four of their friends [7:54] and you're trying to do things like [7:55] authentication, trust rules, access [7:57] controls and that little pop-up network [8:00] only exists for 20 minutes and then goes [8:03] away and there's another one and another [8:04] one and another one that the trust [8:06] professionals, security compliance [8:08] professionals, they will lose control of [8:10] the secure connections. there will be [8:12] too many connections and they will be [8:14] too dynamic in nature. So we have to [8:17] move one level deeper again to secure [8:20] data um trusted data and so this is the [8:24] major fundamental technology shift [8:26] underneath all of this. Uh our company [8:28] has been working with uh this [8:30] interesting technology for many years [8:31] that comes out of academia called NDN [8:34] named data networking. It is rethinks [8:37] really how data is stored, transported [8:40] and secured in NDN. Every piece of data, [8:43] every data packet itself is signed, [8:46] encrypted and contains trust rules. Who [8:48] can do what with that data? Um, our old [8:52] company, we used to deploy this in the [8:53] energy sector. So this right now, this [8:55] technology is deployed on about 2 or 3% [8:58] of all the power generated in the United [9:00] States, including nuclear power plants. [9:01] And we are now deploying this into the [9:03] AI world when you have lots of machines [9:05] all communicating with one another. [9:07] Okay, so back to our question. So how do [9:10] you save 85% of the AI PC's from falling [9:14] into this valley of death, PC death? Um, [9:17] and we think there's three pieces to it. [9:19] Start with a datacentric approach. How [9:22] do you have trust apply to data? [9:25] Secondarily, create a secure trusted [9:27] layer where you can experiment inside [9:29] your production environment. And then [9:31] thirdly, then you will have trusted ROI. [9:34] You have an accurate uh understanding of [9:35] what your ROI is. Therefore, enabling [9:38] many more of your projects to move on to [9:39] production. And that is