Pax8 Adopts Airia Platform After In-House AI Agent Pilots Fail

Pax8在内部AI Agent试点失败后采用Airia平台 企业在内部自建定制基础设施易导致AI项目崩溃，需要依托具备严格身份控制的供应商编排平台才能将Agent推向生产阶段。 Alex Bermudez和Ben Eichholz深入分析了为何在管理层压力、变革管理障碍及安全僵局的影响下，有85%至90%的企业生成式AI概念验证未能投入生产。他们结合Pax8从自研内部Agent失败与严格禁用模型、转而走向平台化策略的历程，展示了集中化AI演练场和界定用例范围如何化解影子AI与非结构化数据的挑战。该分享重点介绍了Airia的编排架构，展示了多模型接入、内置DLP护栏以及隔离式身份管理（区分自主最小权限Agent与委托用户访问）如何实现安全且可扩展的企业AI应用落地。 IDC 与 Gartner 数据显示，82% 至 90% 的企业生成式 AI 概念验证（POC）因高管预期过急、缺乏端到端培训以及创新与安全合规团队的摩擦而未能落地投产。 (2:21) Pax8 的实践表明，完全自研专有 Agent 极易因底层大模型演进过快和开发工具匮乏而停滞，促使企业从闭门自建转向采用专业编排平台的策略。 (6:25) 一刀切禁止公共大模型往往会引发严重的影子 AI 风险；Pax8 通过部署由 Airia 支持的企业级受管“AI Playground”，在保障合规审计的同时满足了业务端对前沿模型的需求。 (6:25) 企业在落地 AI 时无需等待耗时数年的全面数据清洗，而应基于高价值用例精准切分数据，并依托支持严格身份隔离和 DLP 治理的统一多 Agent 编排架构推进投产。 (11:03)

Enterprises struggle to move generative AI into production due to organizational friction, governance gridlock, and the rapid pace of model evolution. Drawing from Pax8's deployment history, Pax8 and Airia advocate for a centralized orchestration and identity-centric platform strategy over bespoke in-house builds and outright model bans. The Production Bottleneck and Custom Development Traps (2:21) According to analyst estimates cited by Alex Bermudez of Airia, between 82% and 90% of enterprise generative AI proofs of concept fail to reach production, citing IDC figures near 90% and Gartner ranges of 82% to 85%. These failures stem from aggressive executive mandates, inadequate user training, poor workflow integration, and friction between innovation teams and compliance departments. Ben Eichholz described Pax8's initial missteps, beginning with Microsoft 365 Copilot, which improved basic individual tasks like email drafting and summarization but could not be embedded into core operational processes or yield measurable business value. In response, Pax8 assembled an internal team with external contractors to develop custom AI agents. Every custom POC stalled because internal tooling lagged, development cycles were slow, and commercial models evolved faster than internal project timelines could support. Mitigating Shadow AI and Modular Protocol Vulnerabilities (6:25) Pax8 initially attempted to enforce a strict corporate ban on consumer tools such as ChatGPT, Gemini, Claude, and DeepSeek. In practice, employees bypassed controls to solve everyday workflow problems, introducing shadow AI risks. Pax8 mitigated this by establishing an internal 'AI Playground' using Airia, restoring IT visibility and compliance by granting governed access to flagship models. At the technical level, Bermudez warned that enterprise AI pipelines face expanded attack surfaces, including data poisoning in fine-tuning sets and untrusted inference infrastructure. He specifically highlighted risks in modular agent setups and the Model Context Protocol (MCP), cautioning teams against 'rug pull attacks,' where a trusted external component or tool is covertly modified or hijacked after receiving system access. Prioritizing Scoped Data Slices Over Total Cleanup (11:03) Bermudez compared enterprise repositories like SharePoint to picturesque cottages built on fragile foundations, noting that unstructured files such as PDFs and spreadsheets continuously pollute systems and cannot be handled under the same architectural assumptions as structured databases. Waiting for data teams to completely cleanse corporate data lakes before deploying AI creates an operational trap. To balance data quality against momentum, Eichholz advised against letting perfection stall deployment. Pax8 implemented a use-case-driven prioritization model: incoming requests are evaluated solely against the specific slice of data necessary for that task. If a use case demands an unmanageable data engineering cleanup, Pax8 deprioritizes it in favor of agile workflows backed by clean, accessible context. Platform Orchestration, Granular Identity, and Democratization (17:46) Airia positions its platform to shift organizations away from hardcoded, point-to-point model integrations toward multi-agent orchestration. The environment offers native access to over 150 foundation models, enabling side-by-side benchmarking, independent data loss prevention (DLP), and observability controls that function independently of external LLM vendor filters. It also supports bidirectional write-backs into operational systems across sales, finance, and operations. Eichholz cited identity management as the primary architectural reason Pax8 partnered with Airia. The design isolates autonomous background agents under distinct machine identities governed by least-privilege access, while interactive assistants operate via the user's specific delegated identity. Furthermore, no-code prototyping allowed employees across diverse business units to build functional agents without bypassing central guardrails. While moving to an external orchestration platform resolved Pax8's development bottlenecks and shadow AI risks, successfully scaling multi-agent systems over the next 6 to 18 months depends on standardizing centralized guardrails across heterogeneous model providers and reliably executing bidirectional writes into operational systems.

Airia美洲区销售负责人Alex Bermudez与Pax8人工智能创新高级总监Ben Eichholz指出，企业生成式AI落地面临高达85%至90%的概念验证（POC）失败率，单凭封禁公开模型或自行硬编码开发均难以应对快速迭代的技术与合规挑战。他们认为，企业必须转向由专业平台驱动的多智能体编排架构，通过用例级数据切分、集中式安全护栏与精细化身份委托机制，在防范新兴安全漏洞的同时实现规模化生产力转化。 企业生成式AI的落地困境与自研局限 (2:21) 引用IDC与Gartner的市场分析数据，Alex Bermudez指出当前82%至90%的企业生成式AI概念验证无法成功进入生产环境。这一鸿沟主要源于多重组织层面的阻力：董事会与管理层急于看到投资回报而施加的自上而下压力、企业对员工AI素养与技能培训的投入匮乏、将AI技术生搬硬套进现有非自然工作流的适配摩擦，以及业务创新诉求与风险合规部门严密控制之间的持续博弈。 Pax8作为拥有约1,800名员工的全球云商务平台，其自身试错经历印证了这一行业痛点。Ben Eichholz回顾称，Pax8最初部署了Microsoft 365 Copilot以提升生产力，但其表现仅局限于会议摘要整理和邮件撰写等浅层任务，管理层无法证实其对核心业务的实际价值。随后Pax8尝试通过内部专项团队结合外部供应商自研专属智能体，但由于内部缺乏配套工具链、开发周期滞后，且底层大模型的演进速度大幅超越了项目推进周期，最终导致所有自研POC无一能够成功投产。 影子AI隐患与模块化协议的安全威胁 (6:25) 在应对AI安全挑战时，粗暴推行全面禁令往往适得其反。Pax8最初曾对ChatGPT、Claude、Gemini和DeepSeek等公共模型实施严格的“铁腕”禁用政策，但因官方审批的技术栈无法解决日常工作难题，员工普遍选择绕过管控使用外部工具，导致“影子AI”（Shadow AI）风险在组织内扩散。为破除安全僵局，Pax8依托Airia平台构建了受企业监管的“AI Playground”，在保留集中审计与合规护栏的前提下向员工开放主流旗舰模型，从而有效遏制了未经授权的外部工具滥用。 与此同时，企业扩展AI架构时也面临新型技术攻击面的威胁。Alex Bermudez提醒企业警惕数据预处理与微调阶段的数据投毒、不可信推理基础设施，以及基于模型上下文协议（MCP）等模块化工具链的安全漏洞。他特别指出了诸如“地毯抽离”（rug pull）的新型攻击向量——即外部工具在获得信任权限后被悄然篡改或劫持。这要求企业建立中立的统一治理层，而不能单方面信任异构的开源协议或第三方模型供应商自带的安全策略。 告别完美数据陷阱与用例切片化策略 (11:03) 针对企业数据基础架构，Alex Bermudez提出了“英国乡村别墅”的比喻：企业的数据资产在表面上看似整洁坚固，底层却往往地基松散，大量Word、PDF、纯文本和CSV等非结构化文档散落在Microsoft SharePoint等存储系统中，且在信息治理框架下仍持续被新文件污染。大语言模型在处理结构化与非结构化数据时存在天然的技术架构差异，若数据团队执意等待所有数据湖彻底清洗完毕才推进项目，势必会陷入漫长的停滞状态。 Ben Eichholz据此强调企业切忌让“完美成为优秀的敌人”。尽管生成式AI领域公认存在“垃圾进，垃圾出”的技术规律，但企业无需通过耗时数年的全局数据治理才能提取价值。Pax8采用的解决方案是“切分大象”的用例驱动优先级模型：针对每一个智能体需求，团队仅评估其运作所必需的一小部分数据切片；一旦某项用例需要过于庞大或不可控的数据工程投入，便会立即被降级搁置，从而集中资源快速推进数据上下文干净、可行性明确的高价值业务流程。 多智能体双向编排与精细身份授权机制 (13:45) 针对未来6至18个月的企业架构演进，Alex Bermudez指出企业必须突破将单个模型硬编码至单一应用的脆弱模式，向统一的多智能体编排生态过渡。该体系不仅要通过检索增强生成（RAG）读取静态策略文档，还必须具备双向行动能力，能够直接反向写入销售、运营与财务等核心业务事务系统。由于各模型厂商自带的安全过滤与系统提示词标准极不一致，企业需要依靠集中编排平台部署中立的数据防泄漏（DLP）与全链路可观测性，以统一执行合规策略。 在落地实践中，Pax8选择与Airia合作的核心考量在于其身份认证架构设计。Ben Eichholz说明，后台自主运行的智能体必须配置为拥有最小特权的独立机器身份，而代表特定员工执行任务的交互式助手则必须使用该员工的委托访问身份执行操作。借助支持150余种基础模型无代码评测的编排平台，Pax8在确保数据驻留与推理环境主权的同时，成功将智能体构建权限下放至全公司各个业务职能部门，加速了从概念到生产的转化周期。 Bermudez与Eichholz总结指出，企业AI落地已从单一工具探索转向多智能体平台的系统化治理，核心在于通过清晰的身份隔离机制与模块化编排防范运行风险。然而，针对模型上下文协议（MCP）等开放生态的外部工具信任链验证，以及智能体在深层次双向读写事务系统时如何彻底杜绝业务误操作，仍是当前高度自主化AI部署中留待实践持续检验的关键课题。

## Background 该视频由 YouTube 频道 Ai4 于 2025 年 8 月 24 日发布，由 Airia 美洲区销售负责人 Alex Bermudez 与 Pax8 人工智能创新高级总监 Ben Eichholz 共同主讲。页面简介指出，本场演讲探讨了企业如何借助 AI 编排平台在满足严格安全与合规标准的前提下，安全地部署、管理并规模化扩展 AI Agent，将零散的 AI 实验转化为统一且安全的企业级 AI 生态系统。 ## TL;DR 演讲直面了当前企业生成式 AI 项目 POC 失败率高达近 90% 的严峻现实，深入剖析了管理层考核压力、用户采纳困难以及创新速度与风控治理冲突等核心痛点。Pax8 分享了自身从引入 Microsoft 365 Copilot 的量化困境、到过度自研定制导致项目烂尾、最终转向平台化编排架构的演进教训。针对数据投毒、未受信任推理、MCP 供应链漏洞以及影子 AI 等新型威胁，Pax8 借助 Airia 平台搭建企业级 AI Playground，并采取“小切口”用例驱动法破解数据治理困局。双方展示了未来多智能体协同、双向 RAG 与细粒度身份鉴权的目标架构，强调通过无代码能力赋能全员创新与业务落地。 ## Atmosphere & style 整场演讲氛围务实接地气，充斥着老友般的轻松幽默与坦率自嘲。Alex 幽默风趣的控场风格与 Ben 坦白踩坑经历的真诚态度形成互补，将看似枯燥的企业级安全合规与治理架构讲得生动透彻。从直面“房间里的大象”到传授破局实战经验，两位讲者的互动默契自然，给观众带来兼具技术深度与实践指导意义的临场聆听体验。 ## Key takeaways - [02:14](https://www.youtube.com/watch?v=jWWKhnPqCss&t=134s) 企业级生成式 AI 项目面临极高的概念验证（POC）失败率，IDC 与 Gartner 数据显示失败率高达 82% 至 90%，跨越 POC 进入实际生产环境是当前普遍困境。 - [02:14](https://www.youtube.com/watch?v=jWWKhnPqCss&t=134s) 企业落地 AI 的三大核心阻碍包括高管层的激进考核与投资回报率压力、员工变革管理与采纳困境，以及业务创新速度与安全合规管控之间的剧烈拉扯。 - [04:43](https://www.youtube.com/watch?v=jWWKhnPqCss&t=283s) 单点生产力工具（如 Copilot）虽能提升日常效率，但因停留在表层交互而难以向业务端提供确凿的 ROI 归因数据。 - [04:43](https://www.youtube.com/watch?v=jWWKhnPqCss&t=283s) 过度自研定制化 Agent 面临构建周期长、基础设施缺乏以及底层模型技术颠覆迭代导致成果快速过时的风险，推动企业转向平台化编排战略。 - [07:30](https://www.youtube.com/watch?v=jWWKhnPqCss&t=450s) 生成式 AI 时代的安全威胁已扩展至数据投毒、未受信任推理基础设施，以及开源生态/MCP 协议中可能改变运行逻辑的“地毯卷走攻击”（Rug Pull Attack）。 - [09:45](https://www.youtube.com/watch?v=jWWKhnPqCss&t=585s) 治理影子 AI（Shadow AI）应遵循“堵不如疏”原则，通过提供合规受控且即时集成前沿模型的企业级 AI Playground 来取代单纯的铁腕封杀。 - [11:03](https://www.youtube.com/watch?v=jWWKhnPqCss&t=663s) 数据治理无需追求一步到位，企业应采用用例驱动法（Use-case approach），针对具体需求提取最小必要数据子集，避免因追求数据完美而阻碍落地。 - [13:45](https://www.youtube.com/watch?v=jWWKhnPqCss&t=825s) 企业 AI 架构正从硬编码点对点接入，演进为具备双向读写 RAG、统一安全护栏和多模型路由的自主多智能体（Multi-Agent）协作网络。 - [15:44](https://www.youtube.com/watch?v=jWWKhnPqCss&t=944s) 企业级 AI 编排治理平台（如 Airia）通过提供内置 DLP、独立护栏、无代码原型设计及多模型集成，帮助企业平衡自研边界并加速业务部署。 - [17:46](https://www.youtube.com/watch?v=jWWKhnPqCss&t=1066s) 智能体安全的最底层支柱在于身份鉴权机制，必须严格区分自主 Agent 的独立实体身份（最小特权）与人类助手的委托访问身份（Delegated access）。 - [17:46](https://www.youtube.com/watch?v=jWWKhnPqCss&t=1066s) AI 创新的可持续性依赖于全员民主化参与，简化安全护栏与智能体构建流程能有效赋能一线员工自主创造与共享业务 Agent。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=jWWKhnPqCss&t=4s) 讲者背景与开场介绍：Airia 与 Pax8 本次演讲由 Airia 美洲区销售负责人 Alex Bermudez 与 Pax8 人工智能创新高级总监 Ben Eichholz 共同主持。Alex 开场调侃道，去年的同一会议正是 Airia 的公开亮相盛会，今天两人将带来一场务实落地的企业级 AI 部署指南，甚至开玩笑说就算内容不够实用，观众至少还能欣赏他特意穿来的“拉斯维加斯风格西装”。随后他邀请 Ben 介绍其企业背景与个人经历。 ```mermaid flowchart LR accTitle: Pax8 Global Footprint and Role subgraph Pax8["Pax8 全球云科技市场平台 (1,800+ 员工)"] direction TB A["美洲 (Americas)"] B["欧洲、中东与非洲 (EMEA)"] C["亚太 (APAC)"] end Pax8 -->|"赋能 IT 渠道与合作伙伴"| SMB["全球中小型企业 (SMB)<br/>技术与云服务交付"] ``` Ben 介绍，Pax8 是一家面向 IT 渠道生态的全球科技交易市场平台，致力于为全球中小企业（SMB）提供技术与云服务支持，目前在全球（美洲、EMEA、APAC）拥有约 1,800 名员工。Ben 的职业生涯大部分时间在销售领域度过，曾担任 Pax8 的销售管理层（在此期间结识了 Alex）。随着生成式 AI 的爆发，他敏锐意识到技术转型的必然趋势，主动转型至企业战略与业务拓展，并于今年初正式统领 Pax8 的 AI 转型办公室（AI Transformation Office）。Alex 补充道，Ben 拥有极深厚的技术底蕴；而 Alex 自己在高速增长的技术行业拥有近 15 年销售与管理经验，此前曾在全球领先的数据隐私、数据治理与 GRC（治理、风险与合规）SaaS 巨头长期任职，笑称当年互联网上烦人的“Cookie 授权弹窗”很大一部分就出自其前司团队之手。 ### [02:14](https://www.youtube.com/watch?v=jWWKhnPqCss&t=134s) 概念验证（POC）的高失败率与企业 AI 面临的核心阻碍 在进入具体议题前，Alex 抛出了一组来自权威分析机构的数据：IDC 统计显示，企业级生成式 AI 项目中有近 90% 的概念验证（POC）最终走向失败；Gartner 给出的失败率也高达 82% 至 85%。这说明当前企业面临的普遍困境并非缺乏技术尝试，而是无法快速且成功地跨越 POC 阶段进入实际生产环境。 面对这一现象，几乎每家企业内部都存在阻碍 AI 落地的“房间里的大象”（Elephants in the room）。尽管各行业的风险偏好和业务特性各有不同，但企业面临的阻力普遍可以归结为三大核心痛点： 1. **来自管理层的激进考核与汇报压力**：从业者面临来自高管（CEO、CIO、CTO）乃至董事会的直接施压，上市公司更是急于在财报与董事会汇报中展示 AI 对业务产生的实质性财务与运营影响。 2. **变革管理与用户采纳（User Adoption）困境**：许多企业尚未向员工提供足够的 AI 素养培训，或者试图用生搬硬套（Square peg in a round hole）的方式强行将 AI 塞入员工既有的工作流，导致工具与实际作业脱节。 3. **创新速度与风险控制（Risk & Control）的拉锯战**：业务部门渴望快速试错与创新，而安全、法务、隐私、合规与数据治理部门则有着严格的底线管控，两者在企业内部形成了剧烈的拉扯。 ### [04:43](https://www.youtube.com/watch?v=jWWKhnPqCss&t=283s) Pax8 的 AI 探索历程：从 Copilot 瓶颈到平台化编排战略 谈及高昂的 POC 失败率，Ben 感同身受地表示“这戳中了我们的痛处”。他坦诚分享了 Pax8 引入生成式 AI 过程中经历的阶段与弯路。两年前，Pax8 率先在全员范围内推出了 Microsoft 365 Copilot。该工具在个人生产力方面表现亮眼——不仅能自动生成会议摘要、润色邮件，还显著提升了日常内容创作的质量。然而，管理层很快遭遇了量化归因的难题：虽然主观直觉上觉得生产力提升了，却无法向业务端提供确凿的数据指标来证明实际 ROI，AI 工具仅仅停留在表层交互，未能真正嵌入核心业务流程。 为了打破这种“浮于表面”的局限，Pax8 随后走向了另一个极端：成立了专门的定制化 AI Agent 开发团队，并引入外部交付伙伴，密集立项了大量 POC 项目。然而，几乎所有自研 Agent 最终都没能成功上线投入生产——要么因为构建周期过于缓慢、企业缺乏适配的基础架构，要么在项目推进途中底层模型技术就已经发生了颠覆性迭代，导致先前的开发成果迅速过时。 ```mermaid flowchart TD accTitle: Pax8 AI Strategy Evolution subgraph Phase1["阶段一：单点效率工具"] P1["引入 Microsoft 365 Copilot"] --> R1["个人生产力提升，但价值无法量化<br/>停留在表层，未能深入业务流程"] end subgraph Phase2["阶段二：过度自研定制"] P2["自建 AI Agent 团队 + 外部伙伴"] --> R2["开发周期过长、底层技术迭代过快<br/>大量 POC 烂尾，无一成功投入生产"] end subgraph Phase3["阶段三：平台化编排架构"] P3["转向以 Airia 为代表的统一编排平台"] --> R3["紧跟前沿模型研究与工具链<br/>兼顾安全合规与快速业务落地"] end Phase1 --> Phase2 --> Phase3 ``` 在经历自研路线的受挫后，Pax8 在今年初果断转变策略，确立了**平台化战略（Platform Strategy）**：依托如 Airia 这类成熟的企业级 AI 编排平台，将跟踪底层模型前沿迭代、工程工具链集成等通用能力交给专业厂商，使内部团队能够专注于业务场景的赋能与 Agent 交付。 ### [07:30](https://www.youtube.com/watch?v=jWWKhnPqCss&t=450s) 全新安全威胁矩阵：数据投毒、未受信任推理与 MCP 的“地毯卷走”风险 随着 AI 架构向纵深发展，企业的攻击面与威胁向量（Threat Vectors）发生了根本性变化。Alex 回顾道，近十年前 GDPR、CCPA 等隐私法案的出台重塑了企业的数据收集与披露规范；而在当下的生成式 AI 时代，企业不仅要关注数据泄露，更要防范训练/微调/知识检索集中的**数据投毒（Data Poisoning）**。此外，将企业敏感数据提交给不可信的推理基础设施（Untrusted Inference Infrastructure），以及跨 LLM / 传统机器学习模型流转中的失控风险，正在成为数据治理的新防线。 在现场互动环节，Alex 询问有多少人在实际工作中使用模型上下文协议（Model Context Protocol, MCP），并进一步指出在拥抱开源工具与开放协议时必须保持高度警惕。他特别提到了**“地毯卷走攻击”（Rug Pull Attack）**——在工具调用、开源插件或 Agent 协议扩展中，恶意代码或被篡改的依赖项可能突然改变行为逻辑，从而在企业连接外部工具生态时引入巨大的供应链与运行时安全漏洞。 ### [09:45](https://www.youtube.com/watch?v=jWWKhnPqCss&t=585s) 影子 AI（Shadow AI）治理：用企业级 AI Playground 取代“铁腕封堵” 针对企业内部普遍存在的“影子 AI”（员工私自使用未经 IT 批准的第三方公共 AI 服务）现象，Ben 坦言 Pax8 过去两年也深受其扰。公司最初采取了“铁腕封杀”手段：明令禁止使用任何 ChatGPT、Google Gemini、Anthropic Claude，甚至包括 DeepSeek。然而事与愿违，严格的禁令并未阻止员工私自使用，因为业务人员在日常工作中确实需要这些先进工具来解决实际问题，而当时企业内供的 AI 栈完全无法满足这些诉求。 ```mermaid flowchart LR accTitle: Shadow AI Mitigation via Enterprise AI Playground subgraph Problem["传统治理困境：铁腕封堵 (Iron Fist)"] direction TB B1["一律禁用 ChatGPT / Claude / Gemini"] --> B2["员工私自使用 (Shadow AI)"] B2 --> B3["数据外泄与合规失控风险激增"] end subgraph Solution["Pax8 破局：Airia AI Playground"] direction TB S1["接入 Airia 统一平台"] --> S2["即时集成各家最新旗舰模型 (如最新 GPT 系列)"] S2 --> S3["内置 IT 审批、安全护栏与数据合规审计"] S3 --> S4["影子 AI 使用率大幅断崖式下降"] end ``` 意识到“堵不如疏”后，Pax8 的 AI 转型办公室成立后的第一项核心动作，就是为员工提供一套安全合规且体验优秀的内部工具。他们基于 Airia 平台搭建了名为 **Pax8 AI Playground** 的企业级统一交互环境，使员工能够直接、安全地使用业内最新的旗舰模型（例如在活动前夕团队就已第一时间将最新的前沿模型推送到平台中）。由于员工可以在 IT 批准、具备企业级安全保护的环境中自由使用顶级模型，Pax8 内部的影子 AI 风险随之大幅下降。 ### [11:03](https://www.youtube.com/watch?v=jWWKhnPqCss&t=663s) 数据治理误区：英国乡村古宅隐喻与“小切口”用例驱动法 Alex 用典型的英国乡村古宅（English Countryside House）作比喻：穿过树篱看过去，地表上的房屋结构方正、古朴优美、四壁完好；但隐蔽在地下的地基部分却杂乱无章、结构脆弱，随时可能坍塌。这生动地映射出企业的数据治理现状： 1. **非结构化数据的黑盒化**：企业在 SharePoint 等各类系统中堆积了海量的 PDF、Word、CSV、纯文本等文档，即便做过信息治理，新文件的持续写入也让数据真实内容充满了不确定性。 2. **结构化与非结构化数据的处理割裂**：LLM 无法天然以相同的方式理解和处理结构化数据库（如 ERP、CRM）与非结构化文档，如果企业在架构上采用一刀切的方式处理两类数据，必然在检索与生成环节引发严重失真。 针对许多企业陷入“数据不完美就无法开展 AI”的教条主义，Ben 强调绝不能**“让完美成为优秀的敌人”（Letting perfect be the enemy of good）**。虽然“垃圾进、垃圾出（Garbage in, garbage out）”的常识依然成立，但大多数企业现有数据所能支撑的场景远超自身想象。Pax8 目前转向了**用例驱动法（Use-case approach）**：每次评估一个新的 Agent 或 AI 助手需求时，团队仅针对该特定用例提取最小必要的数据子集（在“数据大象”身上咬下一小块可消化的肉），若某个场景依赖过于庞大繁杂的全局数据清洗，则直接降低其优先级。 ### [13:45](https://www.youtube.com/watch?v=jWWKhnPqCss&t=825s) 演进之路：从硬编码点对点架构跃升至安全多 Agent 编排体系 Alex 梳理了企业级 AI 从初级阶段迈向成熟形态的演进路线。当前多数企业所处的现状是：评估了一两个模型后，针对不同应用场景以硬编码（Hardcoding）的方式点对点接入特定模型，依靠开发人员手动跟踪模型升级，并在各应用中孤立地处理模型参数与防护。 ```mermaid flowchart TD accTitle: Target Architecture for Multi-Agent Orchestration subgraph DataLayer["统一数据源 (Data Sources)"] D1["非结构化/半结构化 (SharePoint, 文档)"] D2["静态知识 (合规政策, 历史档案)"] D3["动态事务数据 (销售、财务、ERP 系统)"] end subgraph OrchestrationLayer["企业级安全编排平台 (Orchestration Engine)"] direction TB G["统一安全护栏 & 治理策略 (Guardrails & Governance)"] RAG["双向 RAG 引擎 (Read & Write Back)"] Router["多模型智能路由 (Multi-Model Routing)"] G --- RAG --- Router end subgraph AgentsLayer["多智能体协作网络 (Multi-Agentic System)"] A1["分析决策 Agent"] A2["业务工作流 Agent"] A3["事务回写 Agent"] end DataLayer <-->|"读写双向交互"| OrchestrationLayer OrchestrationLayer <--> AgentsLayer ``` 然而，企业真正追求的目标架构是一个**高度协同的多智能体系统（Multi-Agentic System）**。该系统不仅需要利用双向 RAG 技术同时读取非结构化文档、静态政策文件以及实时动态的运营/财务/销售交易数据，更关键的是要具备“以行动为导向（Action-oriented）”的系统回写能力，使 Agent 能够闭环执行业务操作。由于各个底层模型厂商的安全护栏各不相同，企业必须在编排层构建统一的安全策略与合规标准，确保所有跨系统交互都在受控框架内运行。这正是各家企业在未来 6 到 18 个月内必须完成的核心跨越。 ### [15:44](https://www.youtube.com/watch?v=jWWKhnPqCss&t=944s) 自研边界与 Airia 企业级 AI 编排治理平台 Alex Bermudez 坦言，要真正实现企业级智能体架构落地，背后需要投入海量的工程与管理工作。这也是他们在此分享真实挑战的原因——帮助各个企业在规划第一阶段或迈向“第二阶段”（Phase 2）演进时，能提前思考如何应对这些痛点。Pax8 的 Ben Eichholz 补充道，虽然 Pax8 骨子里一直秉持“能自研就自研”（Build it ourselves）的工程师文化，并相信自己能做到最好，但面对如此庞大且复杂的 AI 基础设施与治理链条，团队清醒地认识到有些领域自研并不划算，必须与专业伙伴协同。 ```mermaid flowchart TD accTitle: Airia 平台架构与核心功能图 subgraph INGRESS["接入与开发层"] A["无代码 / 低代码快速原型设计"] B["150+ 基础大模型即插即用"] end subgraph GOV["治理、安全与可观测层"] C["原生 DLP 数据防泄漏"] D["全面可观测性与独立护栏"] E["A/B 对抗评测与基准测试"] end subgraph CONN["连接与部署层"] F["结构化 / 非结构化企业应用"] G["MCP 协议与 Agent-to-Agent 交互"] H["自主选择应用托管与推理环境"] end INGRESS --> GOV GOV --> CONN ``` Alex 随后介绍了 Airia 平台的核心定位：一个面向企业的治理、安全与编排平台。无论企业身处何种行业或探索何种业务场景，Airia 都能护航企业平稳跨入“智能体未来”（Agentic future）。其能力覆盖了从基础员工助手、AI 探索沙盒（Playground），一直到多层级复杂的自主智能体编排系统，真正帮业务团队落实行动力并创造清晰的 ROI。 - **模型广度与评测**：平台原生内置超过 150 种主流基础模型，支持直接进行对抗性场景测试（Battleground scenarios）与多用例横向评估。 - **快速原型构建**：主打 No-Code（无代码）开发能力，大幅降低业务人员落地智能体原型的门槛。 - **内生安全与可观测性**：内置强大的 DLP（数据防泄漏）层与全链路 Observability（可观测性）层，让企业能够自主掌控护栏规则与运行参数，而不是完全受制于底层大模型厂商自带的黑盒策略。 - **生态集成与部署自由**：广泛支持连接企业内部结构化与非结构化数据源，支持 MCP 协议及 Agent-to-Agent 协同，并允许客户完全自主决定应用的托管位置与数据推理环境。 --- ### [17:46](https://www.youtube.com/watch?v=jWWKhnPqCss&t=1066s) Pax8 选择 Airia 的核心考量：身份鉴权与员工创新民主化 Ben Eichholz 重点阐述了 Pax8 决定与 Airia 达成深度合作的两大核心原因。首要原因在于**身份识别与鉴权机制（Identity & Authentication）的架构愿景高度一致**。在 Pax8 看来，这一机制是智能体安全与风险控制的最底层支柱：当一个完全自主运行的 Agent 执行任务时，它必须拥有独立的实体身份并严格遵循最小特权原则（Least-privilege access）；而当员工使用对话式助手代表自己去跨系统操作时，智能体则应使用用户的“委托访问身份”（Delegated access identity）。Airia 在底层设计上与这一安全理念完全契合。 另一个同样关键甚至更为重要的原因在于**以人为本的“人类要素”（The human element）**。Airia 极大地简化了安全护栏的配置流程，并为全员创新提供了坚实底座。它让普通员工能够随时提交创新想法，并直接在平台内动手构建、共享自己的智能体和助手，真正实现了企业内部 AI 创新的民主化。Ben 强调，在功能丰富之外，“部署这些智能体和助手的极简性与便捷度，才是真正打动我们的关键所在。” --- ### [19:01](https://www.youtube.com/watch?v=jWWKhnPqCss&t=1141s) 行业分析机构认可与现场互动致谢 在分享尾声，Alex 幽默地调侃道：“感谢 Ben，我保证下台后绝对没有任何金钱交易，虽然我们身在拉斯维加斯（Vegas）。”他指出，如果大家希望了解客观中立的第三方评价，Airia 已经多次入选多家知名独立分析机构（Analyst firms）关于生成式 AI 开发与安全领域的重磅行业报告。 随着演讲进入尾声，Alex 热情邀请现场观众扫描大屏幕上的二维码或前往 420 号展位（Booth 420）进行深入交流。随后，Alex 与 Ben 共同向现场参会者的聆听表示感谢，并宣布团队成员将继续在展台驻留答疑，全场在轻松融洽的气氛中圆满收尾。 ## Notable quotes > [04:43](https://www.youtube.com/watch?v=jWWKhnPqCss&t=283s) 这戳中了我们的痛处 — Ben Eichholz > [02:14](https://www.youtube.com/watch?v=jWWKhnPqCss&t=134s) Square peg in a round hole — Alex Bermudez > [07:30](https://www.youtube.com/watch?v=jWWKhnPqCss&t=450s) 地毯卷走攻击 — Alex Bermudez > [11:03](https://www.youtube.com/watch?v=jWWKhnPqCss&t=663s) 让完美成为优秀的敌人 — Ben Eichholz > [11:03](https://www.youtube.com/watch?v=jWWKhnPqCss&t=663s) 垃圾进、垃圾出 — Ben Eichholz > [15:44](https://www.youtube.com/watch?v=jWWKhnPqCss&t=944s) 能自研就自研 — Ben Eichholz > [17:46](https://www.youtube.com/watch?v=jWWKhnPqCss&t=1066s) 部署这些智能体和助手的极简性与便捷度，才是真正打动我们的关键所在。 — Ben Eichholz > [19:01](https://www.youtube.com/watch?v=jWWKhnPqCss&t=1141s) 感谢 Ben，我保证下台后绝对没有任何金钱交易，虽然我们身在拉斯维加斯。 — Alex Bermudez ## Mentioned Alex Bermudez, Ben Eichholz, Airia, Pax8, Ai4, IDC, Gartner, Microsoft 365 Copilot, Model Context Protocol (MCP), ChatGPT, Google Gemini, Anthropic Claude, DeepSeek, Pax8 AI Playground, SharePoint ## Video description Alex Bermudez, Sales Leader, Americas, Airia Ben Eichholz, Sr. Director of AI Innovation, Pax8 Abstracts Enterprise AI success requires more than powerful models—it demands sophisticated orchestration. This session explores how AI orchestration platforms enable organizations to securely deploy, manage, and scale AI agents across their enterprise while maintaining strict security and compliance standards. Through real-world examples, discover how leading organizations use orchestration solutions to transform disconnected AI experiments into unified, secure AI ecosystems. Learn best practices for building a security-first orchestration architecture that accelerates AI adoption while mitigating risks. Key takeaways include orchestration fundamentals, security implementation strategies, compliance frameworks, and practical steps for scaling AI initiatives across your organization with confidence. [Source](https://www.youtube.com/watch?v=jWWKhnPqCss)

[4:43] Ben Eichholz: 这戳中了我们的痛处 [2:14] Alex Bermudez: Square peg in a round hole [7:30] Alex Bermudez: 地毯卷走攻击 [11:03] Ben Eichholz: 让完美成为优秀的敌人 [11:03] Ben Eichholz: 垃圾进、垃圾出 [15:44] Ben Eichholz: 能自研就自研 [17:46] Ben Eichholz: 部署这些智能体和助手的极简性与便捷度，才是真正打动我们的关键所在。 [19:01] Alex Bermudez: 感谢 Ben，我保证下台后绝对没有任何金钱交易，虽然我们身在拉斯维加斯。

[0:04] Alex Bermudez: Vegas suit [2:21] Alex Bermudez: square peg, round hole [6:25] Ben Eichholz: iron fist [6:25] Alex Bermudez: rug pull attacks [11:03] Ben Eichholz: perfect be the enemy of good [11:03] Ben Eichholz: garbage in, garbage out [15:44] Ben Eichholz: we believe we can do it best [17:01] Alex Bermudez: battleground scenarios [17:46] Ben Eichholz: When we have an autonomous agent, we want that to be operating as itself, its own identity with least privileged access and with assistants that are, you know, maybe conversational that I might use to go do take action for me. I want it to use my delegated access identity to go take those actions. [19:01] Alex Bermudez: there's no money changing hands after this, I do promise

[0:04] This is really cool. This uh event last [0:06] year was actually area's coming out [0:08] party. So, I'm thrilled to be here for [0:09] everybody. Um, thank you for taking the [0:11] time to hang out with us. As it says on [0:14] the tin, if you're familiar with that [0:16] term, we're going to be talking about [0:17] orchestrating secure agent deployments. [0:20] Um, one of the great things is hopefully [0:22] this is going to end up being a very [0:23] pragmatic session given that I have this [0:25] guy over here to my left, which uh I'll [0:27] have him introduce himself in just a [0:28] second, but if it doesn't end up being [0:29] the most pragmatic session you sit in, [0:31] at least you got to see me in my Vegas [0:32] suit, you know, so we're here. We're [0:34] here for that, too. Um, Ben, please. [0:37] >> Yeah, thanks Alex and great to be here. [0:39] Thank you everybody for coming to our [0:40] session. My name is Ben Ikez. I run the [0:43] AI transformation office at PAX 8. Pax [0:46] is a global technology marketplace. We [0:49] serve the channel uh which serves small [0:52] businesses, their technology and their [0:54] services. And we're about 1,800 [0:56] employees spread across the globe. Uh we [0:58] have offices in uh APAC, AMA and the [1:02] Americas. Um a little bit about me, my [1:05] background is actually in sales. So I [1:07] spent most of my career in sales. Um [1:10] including some time at PAX 8 in sales [1:12] leadership. That's where I met Alex [1:13] originally. And a few years ago, [1:16] actually uh number of years ago now, um [1:19] transitioned out of sales into our [1:21] corporate strategy and business [1:22] development and uh really as this [1:24] generative AI um technology was burst [1:28] onto the scene, I was like, "All right, [1:29] I saw I see this. I see where this is [1:31] going. I want to be a part of it." Um [1:33] and so earlier this year, I actually [1:35] transitioned into this role of heading [1:37] up the AI transformation office at PAX [1:39] 8. So excited to be here. [1:41] >> Don't let what Ben just told you about [1:42] his background fool you. and is actually [1:44] very technical and could blow your mind [1:46] with probably some good conversation [1:47] after this and encourage everybody to [1:49] come find us after. So, Alex Bermudez, I [1:52] have been in sales and sales leadership [1:54] and high growth technology industry for [1:56] uh almost 15 years now. Uh spent a [1:59] significant portion of that time at the [2:01] world's largest prior uh provider of SAS [2:04] for privacy, data governance, and [2:06] governance risk and compliance programs. [2:08] Um and cookie banners, too. if you've [2:10] had to reject cookies recently. Sorry, [2:12] we we had a big part to play in that. [2:14] So, let's kind of dive in. Um, again, [2:18] pragmatic, but everybody needs a stat [2:21] slide, right? And I think this one is [2:23] really poignant because whether you look [2:24] at IDC or you look at Gartner or you [2:26] look at another analyst firm or maybe [2:28] you're even just doing some analysis [2:30] with your in your own organization, [2:31] almost 90% of PC's at the enterprise [2:35] level with generative AI are failing [2:37] according to IDC. And if you look at [2:39] Gartner, they say 82 to 85. There's all [2:42] sorts of statistics you can go and find [2:44] out there, but the reality is we're not [2:46] getting past PC quickly enough or [2:48] successfully enough. [2:51] Now, you might be thinking about your [2:53] own internal organization and saying, [2:54] well, maybe there's one elephant in the [2:56] room or maybe you've got multiple AI [2:58] elephants in the room as to why this [3:00] actually isn't occurring, why you can't [3:02] lead with success, why you can't [3:04] innovate quickly enough. And every [3:06] organization has some nuance. Your [3:08] industry, your risk tolerance, lots of [3:11] different things that could create some [3:13] of these barriers. But if we try to [3:15] distill down the list from most of the [3:16] conversations that we have with [3:18] organizations today, most of your issues [3:21] fall into one or several of these [3:23] buckets. In a lot of instances, we know [3:25] the people here in this room today have [3:26] a lot of pressure from their leadership [3:28] teams to say they're using AI. Talk [3:31] about how they're working with AI. Maybe [3:33] you're a public company and you've got [3:34] to do board reporting on how AI is [3:36] making an impact in the business. Or [3:37] you've got a hyperaggressive CEO, CIO, [3:40] CTO that is saying we're not doing [3:42] enough with AI today. So there's a lot [3:43] of pressure there from a practitioner [3:45] standpoint. We understand that. Maybe [3:46] you've also got change management and [3:48] user adoption issues. I think this is a [3:50] pretty prevalent one that most [3:51] organizations are struggling with today [3:53] in some capacity because either the uh [3:56] the business hasn't invested enough in [3:57] training and AI literacy or perhaps [3:59] you're still trying to figure out how to [4:01] take AI into your users workflows in a [4:04] way that actually works and is natural [4:05] for them versus sort of square peg round [4:08] hauling AI into what they're doing. Um, [4:11] and the last thing really comes down to [4:14] something that Shandra was talking about [4:15] in the last session for anybody that was [4:16] here talking about that balance of risk [4:18] and control. Business has certain [4:20] tolerances. Security, privacy, [4:22] governance, compliance, all of those [4:23] other functions have their own [4:25] tolerances. And so that there's [4:26] literally sort of cataclysmic argument [4:29] that's happening within a lot of [4:30] organizations today, whether you call it [4:32] that or not. And irrespective of where [4:34] you are on the maturity journey, this is [4:36] a pretty distilled list of of how we [4:38] think about boiling this down to [4:40] adoption challenges that we really see. [4:41] Any comments, Ben? [4:43] >> Uh, I take that statistic personally um [4:46] because that hurts from experience. Um, [4:49] it's it's really challenging, you know, [4:52] um, talking a little bit about our [4:53] journey at PAX 8 with generative AI. Um, [4:57] a couple years ago we had uh we launched [5:00] co-pilot for Microsoft 365 and it it's [5:04] amazing for personal productivity. It uh [5:07] it summarizes my meetings, helps helps [5:09] everybody write emails, better emails, [5:12] kind of levels up the quality of uh our [5:15] content generation. Um, but we were [5:17] really struggling to actually prove the [5:20] value. uh we knew in theory our you know [5:23] the gut says yeah it's increasing our [5:26] productivity but we can't prove this how [5:28] how how are we actually impacting our [5:31] workflows um well it it remained at the [5:33] surface area surface level and never [5:36] made it into the the workflows and the [5:38] processes of business. Uh so actually [5:41] early last year uh we implement uh we [5:44] introduced a a new team uh focused on [5:46] custom development of AI agents and we [5:49] were going to we had all these proofs of [5:51] concepts um that we were we were [5:54] starting to develop on. We brought in [5:55] some partners to help us develop those [5:58] and lo and behold none of them ended up [6:01] making it to production. either it was [6:03] too slow, we didn't have the right [6:05] capabilities, the technology changed [6:07] before we before we finished the [6:09] project. Uh, and ultimately it didn't go [6:12] anywhere. Um, and you know, I think it [6:15] was really that was an overcorrection of [6:17] our challenge with with co-pilot 365 and [6:21] other chatbot AI. Um, and so we made the [6:25] decision earlier this year to go with [6:27] the platform strategy. Um, we are going [6:29] to lean on vendors and platforms like [6:32] area uh to to [6:36] stay close to all the technology, stay [6:38] on the bleeding edge of research and [6:40] ensure that we have the tools that we [6:42] need to deliver AI agents uh and [6:44] assistance. Uh, so we're really excited. [6:46] We um definitely have hit some bumps [6:48] along the way, but we we believe that [6:50] our strategy is is right now. And you [6:53] know, one thing that I really appreciate [6:54] about this community is, you know, we're [6:56] all working on this together and we're [6:58] all learning together. So that's where [7:00] we're at right now, [7:01] >> 100%. Um, and I think we, you know, as [7:04] we start going down this path, every [7:07] organization needs to come to grips with [7:08] a variety of different issues that [7:10] you're going to run into. And so now we [7:12] want to switch gears and get a little [7:13] bit more practical on those those [7:14] issues. And again, would encourage [7:16] everybody to think about your own [7:17] organization. This is a sort of a best [7:20] we can list that covers a lot of [7:21] different uh issues that we hear from [7:23] the market that you might be hearing [7:25] yourselves internally or even externally [7:27] with the peers that you speak with. Um [7:30] let's start here because I think this is [7:32] one of the least talked about issues in [7:36] generative AI today. But let's not just [7:39] kind of you know postulate a little bit. [7:41] Let's actually talk about some real [7:42] issues. [7:44] First off, when we start thinking about [7:47] how threat vectors have emerged, how our [7:50] threat landscape uh and overall surface [7:52] that we have to cover has changed, [7:54] there's a whole bunch of stuff that [7:56] happens just as we start thinking about [7:57] data collection. And anybody that had to [7:59] live through some iteration of GDPR or [8:02] CCPA or CP, insert any other acronym [8:06] here, you understand that starting uh [8:08] you know, close to 10 years ago now, our [8:09] data collection practices really had to [8:11] start changing from a regulatory [8:13] landscape. But you also had to be more [8:15] cautious not just about what were you [8:16] disclosing but you had to make sure data [8:18] wasn't getting modified. And now we talk [8:20] about data poisoning in different data [8:21] sets where perhaps uh things are being [8:23] inserted and we're not necessarily [8:25] covering the right profile. We also have [8:27] to start looking at untrusted inference [8:30] uh infrastructure and looking at what [8:32] are the the parameters that we can place [8:34] around how we're going through these [8:35] processes, how we are subjecting our [8:37] data to LLMs or even machine learning [8:40] models or some combination therein. And [8:42] then as we start getting into the real [8:44] deployment stages and the real practical [8:46] applications of scaling out, there are [8:48] other issues that are popping up. Who in [8:50] here is uh in some capacity or another [8:52] already working with MCP? [8:54] >> Great. Keep your hands up if you really [8:56] think you can trust MCP. [8:59] >> Okay, a few. [9:01] >> I appreciate that. Um, is anybody in the [9:03] room familiar with the concept of a [9:05] rugpole attack? [9:07] Okay, highly highly recommend everybody [9:09] to go look up what a rug rug pull attack [9:11] is and start thinking about how it could [9:12] be applied to MCP as one of the [9:14] potential threat vectors that is [9:16] starting to get deployed as we extend [9:17] out to all of these different solutions. [9:19] But especially with MCP and the open [9:21] source community to Ben's point uh can [9:23] also introduce a lot of risk that we [9:25] have to think about. [9:27] >> Yeah, 100%. And this is this is [9:29] something we're facing on the on the [9:30] day-to-day basis because we we see the [9:32] vision of model context protocol um and [9:35] this idea of modularizing your AI [9:38] capabilities um across your [9:40] organization. U it's an amazing future [9:43] but lots of challenges. One other uh one [9:45] I want to pull out here is the shadow AI [9:48] um risk. This was something we have [9:50] faced uh across organization constantly [9:54] for the last couple years. I'm sure you [9:55] all have as well. Um, we started with a [9:58] with a very much an iron fist approach. [10:00] No shadow AI, no chat GPT, no Gemini, no [10:04] Claude. Um, definitely no Deep Seek. [10:07] And, uh, you know, we what we discovered [10:09] is people were doing it anyways. Um, [10:12] because they didn't they weren't getting [10:14] what they needed from our AI stack. [10:17] That's what it boiled down to is we were [10:19] not delivering the technology that they [10:21] wanted to use in their day-to-day uh, [10:23] work. And so that was the very first [10:26] thing that my uh AI transformation [10:28] office did when we were formed is we [10:30] need to give the tools to our employees [10:32] that they want to use that they're [10:33] comfortable using. Uh so we launched [10:36] what we call an AI playground uh built [10:38] on the area platform where we have uh [10:41] access to all of the latest and greatest [10:42] flagship models. I actually pushed GBT5 [10:45] last night uh before this week. So [10:47] really excited to get all of those in [10:49] our employees hands and our shadow AI [10:51] risk went way down because people were [10:53] now getting what they wanted with PAX 8 [10:57] approved IT approved secure AI. [11:03] There's another aspect of this that um [11:05] maybe doesn't seem risky inherently but [11:09] um I used to live in the UK. I really [11:11] love like the idyllic English [11:12] countryside house and you see a lot of [11:14] stuff that looks like this. But the the [11:17] point behind this is on the surface, you [11:18] know, up top where everybody's actually [11:20] going to see it if they're walking past [11:21] the hedges is just what looks like this [11:23] nice structured old idyllic house, [11:26] right? It's beautiful. It's pristine. [11:28] You see, it's got the four walls. The [11:30] corners look good. And then you've got [11:32] this stuff down below that is kind of [11:36] structured, kind of not. You're not [11:38] really sure what's down there. You're [11:39] not really sure how it's going to hold [11:41] together as you go do this. And this is [11:43] another really practical challenge that [11:45] organizations are facing, but for two [11:47] different reasons that we'll talk about [11:48] very quickly. Number one, when you start [11:51] looking at a lot of unstructured data [11:52] sources, it's pretty difficult. It's [11:54] pretty challenging to have a relative [11:56] level of certainty of exactly what's in [11:58] there, right? Because you've got a mix [11:59] of plain text files, PDFs, Word [12:01] documents, maybe CSVs, uh, and it's all [12:04] kind of sitting in an enterprise [12:05] application like a SharePoint probably. [12:07] And even if you've gone through an info [12:09] governance uh exercise, stuff gets added [12:11] to these folders all the time. There's a [12:14] second challenge though. We cannot [12:16] expect that LLM of their own valition [12:19] are going to be able to work with [12:20] unstructured and structured data the [12:22] exact same way. So if you're trying to [12:25] fundamentally approach structured data [12:27] the same way that you do unstructured [12:28] data, you're going to run into issues. [12:30] And this architectural thinking that [12:32] organizations need to do as they [12:34] continue to push the boundaries of what [12:35] they do with generative AI really does [12:38] merit review of what the structure of [12:40] that data is. Now, one of the things [12:42] we've been talking about is this over [12:44] pivot to focusing on overperfect data, [12:46] which I know you want to comment on. [12:47] >> Yeah, definitely. It's it's this idea, [12:49] this concept of letting perfect be the [12:50] enemy of good. Um, we have had countless [12:54] conversations and uh discussions around [12:57] we need to get our data right. We need [12:58] to get our data right. we need to get [12:59] our data right because garbage in, [13:00] garbage out, right? We all know that. We [13:02] know we the context is what matters with [13:04] with generative AI. All that is true. [13:07] However, you can probably do a lot more [13:11] with your data today than you think you [13:13] can. Uh so we've actually pivoted to a a [13:16] use case approach. So each use case that [13:18] comes in uh for an agent or an AI [13:22] assistant, we look at the specific data [13:24] that that needs and we identify, you [13:27] know, the small chunk of the giant the [13:30] elephant, the data elephant, uh that we [13:32] can take a bite out of. Uh and if it if [13:35] we can't if it's too big, that use case [13:37] gets dep prioritized. So that has [13:39] actually become a key criteria in our [13:41] use case prioritization. [13:45] Let's get hyper hyper practical for just [13:48] a moment. Where a lot of organizations [13:50] are today, whether you're in PC or [13:52] you've started to see some production [13:54] deployment success is you're probably in [13:56] this mode where you started looking at [13:58] one or two models and then you said, [13:59] well, this works for that, maybe it [14:00] doesn't work for this. And so then we [14:02] started looking at how do we take a [14:03] multimodel approach maybe across [14:05] different applications or linearly one [14:07] model per application and let's go [14:09] hardcode that in uh let's maybe do a [14:12] little bit of refinement with it or then [14:13] maybe we have our developers upgrade to [14:15] another model when it comes out. Um and [14:18] then we got to figure out the nuance of [14:20] managing each of these models. Great. [14:22] This is actually not bad stuff. There [14:25] are still a lot of organizations who [14:26] haven't quite figured out how to get [14:27] here yet. This isn't what we want if [14:29] we're really honest with ourselves. [14:32] What do we actually want though? [14:36] We want this beautifully orchestrated [14:38] multi-agentic system that can read rag [14:41] through right into different data [14:44] sources that uh could be structured. [14:46] They could be unstructured. They could [14:47] be semi-structured. They could also be [14:49] more transactional data sources that are [14:51] really updating all the time based on [14:53] operational sales finance systems. or [14:56] they could be static that contain our [14:57] policy documents, lots of historic data [15:00] as well. But we don't just want to [15:02] understand that data and have LLM tell [15:04] us what's in there and help us to make [15:06] decisions on it. We then want that same [15:09] actionoriented approach to writing back [15:11] to those systems that real agentic [15:13] framework. But each of these systems, [15:15] each of these models comes with their [15:17] own complexity. You know, each of those [15:18] models that I showed earlier, the model [15:20] providers don't train in guardrails the [15:22] exact same way. So you have to be able [15:23] to look at how do we build these systems [15:25] in a way that is respectful of the [15:28] governance policies, the security [15:29] policies, the data use standards that we [15:31] have in place, but can still take this [15:33] actionoriented approach. This is where [15:34] most organizations really want to get to [15:37] over the next 6, 12, 18 months. But from [15:40] that first slide I showed to where you [15:42] get here, [15:44] there's a lot of work that goes into [15:45] that, which is part of the reason that [15:46] we wanted to talk about some of these [15:48] real challenges so that everybody can go [15:49] back and start thinking about how they [15:51] want to address those if you haven't yet [15:52] or think about how you address those in [15:54] a sort of a phase 2 type of approach if [15:56] you're ready to. Anything else to to add [15:59] there, Ben? [16:00] It's it's definitely what we want. It's [16:02] really really challenging to get there [16:04] and especially if you're building it all [16:06] yourself. Um, like I mentioned earlier, [16:08] PAX 8 has a bias for building it [16:10] ourselves because we believe we can do [16:11] it best. But, um, there's certain things [16:14] where we we've decided that we can't do [16:16] that best and that's where we we decided [16:18] to work with partners. [16:20] So, very quickly because you know our [16:23] names are on this and my name's on it. [16:25] So, I get to tell you just a second [16:26] about area. We have built a enterprise [16:30] governance security and orchestration [16:32] platform. Irrespective of use cases or [16:35] what industry you work in, we have the [16:37] ability to help you shepherd your [16:39] business into the agentic future and [16:41] support all those lines of business from [16:43] basic assistance and sort of that AI [16:44] playground uh capacity to offer the [16:48] average user like Ben talked about all [16:50] the way into these agentic multi-tiered [16:52] orchestrated systems that have the [16:54] ability to give you that action that you [16:57] need to take back to the business to [16:58] make meaningful impact and ROI happen. [17:01] We do that by offering a platform that [17:03] gives you rapid prototyping [17:04] capabilities, principally no code, and [17:06] also gives you the benefit of access to [17:07] over 150 foundational models directly [17:09] within the platform today that you can [17:11] test against each other, take [17:13] battleground scenarios, do evaluations [17:15] of multiple use cases together. We also [17:17] give you a robust DLP layer, an [17:20] observability layer as an integrated [17:22] component to the platform so that you [17:24] actually have the control of the guard [17:25] rails and the parameters that you place [17:27] and not leaving it just at the behest of [17:29] the model providers themselves. We also [17:31] give you the ability to connect to a [17:33] variety of different structured and [17:34] unstructured enterprise applications. [17:36] MCP support A to A and you ultimately [17:39] get to determine where you host the [17:41] application, where you do that [17:42] inferencing of data as well. Um [17:46] yeah, so I want to touch on two things. [17:47] The two reasons why PAX 8 is working [17:49] with with area. Um first off is the [17:53] their vision of the identity and [17:54] authentication challenge. Um this is at [17:57] the core of this issue of security and [18:00] risk uh with with these agents. Uh and [18:03] it is that when we have an autonomous [18:06] agent, we want that to be operating as [18:09] itself, its own identity with least [18:11] privileged access and with assistants [18:13] that are, you know, maybe conversational [18:15] that I might use to go do take action [18:17] for me. I want it to use my delegated ad [18:20] uh access identity uh to to go take [18:23] those actions. And that's that's a [18:26] common vision that area shares. Um so [18:27] that was reason number one. But reason [18:29] number two and and pro maybe even more [18:30] importantly is the the human element. Um [18:34] area makes it so easy to set up [18:36] guardrails. Uh they make it easy to [18:38] collect ideas and and uh democratize [18:42] innovation with all of your employees by [18:45] letting them submit and share their [18:46] ideas and build them right in the in the [18:49] uh area platform. Um and you know [18:51] there's so much more. I know we don't [18:52] have time to go into everything, but uh [18:54] the ease of and simplicity of deploying [18:57] these agents and these assistants is [18:59] really what makes the difference for us. [19:01] >> Thank you, Ben. There's no money [19:03] changing hands after this. I do promise. [19:05] I know we're in Vegas, but um [19:07] >> if you want to read what third parties [19:09] that are independent are saying about [19:10] us, there's a number of different [19:11] analyst firms that we've been featured [19:13] in reports and different cycles of what [19:14] they're going through on their own uh [19:16] aspects around generative AI, [19:17] development, security. Uh, last thing is [19:20] I know a bunch of you took pictures and [19:21] I'm so sorry that this was last, but if [19:23] you do uh want to grab the QR code or [19:26] come find our team, we're at booth 420. [19:28] Uh, Ben and I are happy to stick around. [19:29] Also, a number of other members of the [19:31] area team are going to be here as we're [19:33] wrapping up. But thanks very much for [19:34] participating. Uh, appreciate the time [19:36] from everybody. And Ben, thank you very [19:38] much. [19:38] >> Yeah, thank you. Thanks for having me.