YASH Technologies Unveils NUEAC Platform to Curb Shadow AI Risks

YASH Technologies推出NUEAC平台以遏制影子AI风险 除非尽早加装统一治理层，否则不受约束的自主AI代理将通过失控的Token成本和数据破坏威胁企业稳定性。 Nitin Gupta探讨了缺乏治理的AI如何引发严重的企业风险，并将其与个人电脑、互联网及云计算的历史普及浪潮进行了类比。他详细列举了现实中的失败案例——包括预算超支、未受监控的代理导致的数据清除以及知识产权泄露——凸显了对全面支出控制、风险缓解以及明确ROI框架的迫切需求。为解决这些盲区，Gupta提出建立统一治理层，并介绍了YASH Technologies的内部平台NUEAC，该平台提供集中式模型管理、安全测试和实时优化功能。 从个人电脑、互联网到云计算，企业在以往技术浪潮中往往耗费十余年才建立起完善治理，如今 AI 部署正面临重蹈覆辙的严峻风险。 (0:04) 缺乏管控的 AI 引发了严重事故：有企业在 4 个月内耗尽全年预算与 token 配额，某项目因缺乏 ROI 指标而直接计提 6200 万美元损失并关停，还有超 5 家大型企业因数据外泄提起诉讼，更有未受控智能体误删了 3 个月的客户交易记录。 (0:38) 尽管以“vibe coding”为代表的快速 AI 辅助编码日渐风靡，但研究显示超 40% 的此类解决方案无法达到企业级的安全与可靠性标准。 (0:38) 统一的 AI 治理框架必须超越单纯的合规要求，主动检测影子 AI（Shadow AI），并覆盖责任 AI、实时 FinOps 成本监控、模型可解释性以及零数据复制架构。 (0:24) 针对业务与技术团队的差异化需求，YASH Technologies 打造了 NUEAC 内部管控平台，整合基于角色的权限控制（RBAC）、模型上下文协议（MCP）及元智能体优化层，以实现全流程的安全性与投入产出比追踪。 (0:19)

Nitin Gupta argues that enterprise AI deployment without centralized governance replicates historical technology blunders, creating severe financial, operational, and intellectual property risks. He contends that organizations must establish a unified governance and optimization framework before scaling AI initiatives. Historical Precedents of Lagging Technology Governance (0:04) Gupta contextualizes modern AI adoption within earlier enterprise technology cycles, arguing that organizations have repeatedly suffered from retrofitting oversight years after deployment. During the personal computing wave, it took over a decade to institute effective controls as unmonitored floppy disks enabled uncontrolled data movement. Similar patterns emerged during the rise of the commercial internet, which exposed enterprises to security and privacy vulnerabilities, and cloud computing, where decentralized citizen developers on platforms like SharePoint triggered ballooning infrastructure costs. According to Gupta, modern enterprises face the same choice with AI: establish proactive guardrails early or repeat the cycle of slow, costly retrofits. A successful approach must address governance challenges at the outset, encompassing corporate policies, regulatory standards, FinOps, and Responsible AI principles. Material Risks of Ungoverned AI Deployments (0:38) Gupta details several industry case studies illustrating the severe consequences of ungoverned AI. In one instance, an enterprise exhausted its entire annual token and AI budget within four months due to an absence of spend visibility. In another, an organization wrote off a $62 million AI initiative after failing to institute KPI tracking or ROI governance. Intellectual property leakage has also prompted lawsuits from more than five major corporations after proprietary data entered the public domain through unmonitored AI tools. Operational vulnerabilities are further exacerbated by autonomous agents and informal development practices. Gupta highlights a case where an ungoverned agent wiped out three months of customer transactions and database records, requiring manual restoration. Furthermore, studies cited by Gupta indicate that over 40% of rapid 'vibe-coded' solutions fail to satisfy enterprise-grade security and reliability standards, heightening risks as unmonitored departmental tools and vendor agents proliferate. Designing Governance Across Executive and Technical Personas (0:24) To counter shadow AI, Gupta advocates for a unified governance layer that extends beyond basic compliance checklists. The proposed framework proactively identifies authorized versus unauthorized tools, issuing real-time alerts to security operations centers (SOC) and infrastructure teams. It also manages FinOps by tracking granular token use, per-seat licensing costs, explainability, toxicity, and hallucinations. Gupta stresses that governance architecture must balance divergent organizational priorities. Executive leaders (CEOs, CFOs, and COOs) require measurable ROI, financial optimization, and clear business cases before sanctioning projects. Conversely, technology and compliance leaders (CIOs, CISOs, CTOs, and CDOs) prioritize operational security, zero-data-copy architectures to prevent data duplication, platform interoperability, and enforceable ethical standards. Technical Execution via the NUEAC Architecture (0:19) Drawing from an internal initiative at YASH Technologies—prompted when aggressive AI adoption caused duplicate projects and soaring expenses—Gupta presents the NUEAC (Next-Generation Enterprise Unified Platform for AI Agent Control) framework. The architecture enforces role-based access control (RBAC) across both human users and agent-to-agent interactions, connects an 'LLM Zoo' to enterprise backends like SAP and Salesforce via Model Context Protocol (MCP), and mandates pre-onboarding vulnerability testing across text, voice, and video. To safeguard sensitive proprietary workflows, the system implements continuous compliance monitoring (covering HIPAA, SOC, and GDPR) alongside self-hosted sovereign models for confidential bidding and internal algorithms. NUEAC also integrates an intelligent optimization layer: a meta-agent that monitors system activity to suggest prompt adjustments, code remediations, and cost-saving model downgrades, such as shifting tasks from large language models to small language models when advanced reasoning is unnecessary. Gupta concludes that organizations must not allow governance blind spots to undermine their AI ambitions. In his assessment, implementing upfront controls and continuous spend optimization provides the necessary foundation for scalable, secure enterprise adoption.

Nitin Gupta指出，如果企业未能吸取PC、互联网及云计算时代治理滞后的历史教训，放任未受监管的人工智能与自主智能体在内部蔓延，将直接面临巨额预算超支、系统数据灭失以及专有知识产权泄露等严重风险；企业必须在规模化之前建立统一的治理与优化架构。 历史技术浪潮与治理滞后的教训 (0:04) Gupta回顾了企业级IT技术的演进历程，指出过往的技术普及往往伴随着长达数年甚至十余年的治理空白。第一波个人电脑（PC）浪潮耗费了十多年才建立起规范控制，在软盘时代，企业无法有效监控进出数据，设备极易被滥用于未经监管的私人或办公场景；第二波互联网浪潮在初期同样缺乏管控，员工能够直接访问网络工具与电子邮件，迅速引发了严峻的系统安全与隐私担忧；第三波云计算与平民开发者（Citizen Developer）在SharePoint等平台上自行构建本地应用，更导致云支出急剧飙升，迫使企业在事后才被动补救治理。 Gupta警告称，当前生成式AI的普及正在重演这一周期，各业务部门跳过IT部门独立采购和使用“影子AI（Shadow AI）”的现象极为普遍。企业领导者当前面临关键抉择：是应当吸取历史教训并在AI部署初期尽早建立前置治理机制，还是继续重蹈覆辙，放任技术野蛮生长并在未来付出沉重代价。 缺乏管控的AI风险与业务灾难 (0:38) Gupta通过多起行业真实案例说明了未受治理AI所带来的破坏性后果。在财务与投资回报方面，一家企业由于缺乏支出控制，在短短四个月内就彻底耗尽了全年的AI预算与Token额度；另一家企业则因未设定明确的业务目标、KPI及投资回报率（ROI）评估机制，最终被迫叫停整个AI项目并直接注销了高达6200万美元的投资。在合规与法律层面，已有超过五家顶级企业因员工使用未受监控的AI工具导致专有数据流入公共领域，进而引发侵权诉讼。 在底层系统运行与开发层面，自主智能体（Agent）的失控造成了灾难性故障：某企业部署的一项未受信任的智能体在执行任务时做出荒诞决策，彻底清空了企业数据库中整整三个月的客户交易记录，迫使团队进行极其繁重的手动恢复。此外，虽然借助AI快速生成代码的“氛围编码（vibe coding）”迅速普及，但研究表明超过40%的相关方案仅适用于基础的概念验证，根本无法达到企业级的安全性与可靠性标准。随着各类OEM智能体与内部自研智能体交织，多智能体网络正变得不可控。 协同业务与技术的统一治理框架 (0:24) 为了消除盲区，Gupta提出了一套超越单纯合规审计的统一治理层架构。该体系必须具备主动监测能力，能够实时区分授权与未授权工具，并在检测到影子AI活动时向安全运营中心（SOC）和基础设施团队即时告警；同时，治理层需深度整合负责任AI（伦理与信任）、幻觉与毒性内容规避、模型可解释性，以及细粒度监控Token消耗、软件许可和ROI的FinOps财务运营机制。 该治理体系还必须精准协同组织内不同职能角色的诉求。以首席执行官（CEO）、首席财务官（CFO）和首席运营官（COO）为代表的业务高管，重点聚焦于实际价值转化、项目优先级的战略对齐以及具体用例的ROI实现；而首席信息官（CIO）、首席信息安全官（CISO）及合规官等技术与风控团队，则严格专注于防止数据泛滥复制的“零数据复制（zero-data-copy）”架构、多平台互操作性、伦理标准以及覆盖全企业的安全底线。 YASH的NUEAC平台治理架构实践 (0:19) Gupta以YASH Technologies自身的治理实践为例说明解决方案。在公司高层要求各部门积极引入AI后，企业内部同样迅速遭遇了成本暴增、影子AI涌现以及重复造轮子的问题。为此，YASH自主研发了名为NUEAC（下一代企业AI智能体统一控制平台）的管控架构，在确保员工敏捷开发自由度的同时建立严格约束。 NUEAC平台覆盖了八大治理维度：首先，通过基于角色的访问控制（RBAC）管理人类与智能体之间以及智能体相互交互的权限；其次，构建整合大语言模型（LLM）、小语言模型（SLM）及行业模型的“LLM动物园”，并通过模型上下文协议（MCP）打通SAP与Salesforce等核心业务系统；第三，在智能体上线前开展涵盖文本、语音和视频的漏洞渗透测试（VAPT）与隐私保护审查；第四，实施满足HIPAA、SOC和GDPR的持续合规监测，并部署自托管的主权AI以保护竞标与算法专有性；第五，深度追踪智能体推理链、开发者Token成本归属与成功KPI；最后，将治理延伸至物理与边缘硬件，并引入全局元智能体（Meta-agent）优化层，动态提供代码提示词纠偏或将高成本大模型自动降级为小模型的降本建议。 Gupta总结认为，企业决不能让治理盲区扼杀自身的AI雄心。实现AI安全扩展与业务落地的根本路径，在于优先确立兼顾多方利益相关者的统一治理机制，并通过全生命周期的持续优化来约束算力与模型支出。

## Background 该视频由行业频道 Ai4 于 2026 年 8 月 17 日发布在 YouTube 平台，主讲人为 YASH Technologies 的 AI 与数字化负责人 Nitin Gupta。视频简介指出，自主 AI 智能体的爆发式增长与缺乏管控的“影子 AI”正带来成本失控、合规失败及严重数据漏洞等企业盲区风险，本专题旨在分享如何通过集中式治理与实时可观测性构建安全可信的智能体运行体系。 ## TL;DR 演讲首先回顾了 PC、互联网与云计算三次技术浪潮中“先扩张、后治理”的历史教训，警示失控的 AI 部署已导致行业内出现预算击穿、6200 万美元投资核销及智能体误删核心数据等重大惨剧。面对多模型并存与智能体孤岛，传统基于人类角色的权限管理已无法应对，企业必须构建覆盖“影子 AI”监测、负责任 AI 及数据零复制的统一治理层。为此，讲者提出了涵盖智能体身份管理（Agent IAM）、多模型矩阵（LLM Zoo）及主权 AI 资产隔离的 NUEAC 治理架构，通过多模态渗透测试、细粒度 FinOps 与清晰的 ROI 准入机制消除企业盲区。最后，方案通过顶层部署自主“元智能体”提供代码与提示词级降本调优建议，倡导企业严格践行“健全治理、深度优化、规模采纳”的演进路径。 ## Atmosphere & style 整场演讲风格沉稳干练且极具警示意味，讲者以一针见血的企业级实战视角，将千万级项目核销与数据覆灭案例抽丝剥茧，现场氛围严肃而充满技术紧迫感。叙事节奏层层递进，从技术浪潮的历史周期律反思自然过渡到硬核的系统拓扑与架构蓝图推演，兼具宏观战略高度与细致到 Token 成本控制的工程务实感。聆听过程如同直面一场针对企业高管与架构师的“防灾演练”，既令人对放任自流的隐形风险心生敬畏，又展现出用体系化工程重塑技术秩序的清晰底气。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=4s) 企业在经历 PC、互联网与云计算浪潮后往往陷入“先扩张、后治理”的被动周期，面对生成式与自主 AI 时代，企业必须在早期主动介入治理体系建设。 - [03:09](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=189s) 缺乏治理的失控 AI 会带来灾难性盲区，包括 Token 耗尽击穿预算、核心知识产权泄露、“影子 AI”暗中蔓延以及自主智能体误删关键业务数据。 - [03:09](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=189s) AI 辅助生成的“氛围编程”（Vibe Coding）虽能加速原型开发，但超过 40% 的方案仅能停留在 PoC 阶段，因安全漏洞与逻辑缺陷难以满足企业级生产标准。 - [05:52](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=352s) 各厂商争推原生智能体导致工具碎片化与孤岛化，带来成本难以追踪、责任主体模糊、部门重复建设以及整体投资回报（ROI）归零等系统性风险。 - [07:35](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=455s) 企业需要建立覆盖影子 AI 监测、负责任 AI、FinOps 成本工程、可解释性验证、ROI 价值管理与零数据复制架构的“统一治理层”。 - [09:45](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=585s) 治理体系必须同时兼顾业务管理层对 ROI 价值变现与成本控制的诉求，以及技术治理层对系统互操作性、零数据复制架构与安全底线的保障。 - [11:47](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=707s) 在智能体协同协作场景下，传统针对人类用户的 RBAC 权限体系已不足够，必须引入细粒度的“智能体身份管理”（Agent IAM）与 MCP 连接器生态。 - [11:47](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=707s) 构建汇聚 LLM、SLM 与行业垂直模型的“LLM Zoo”，按需合理匹配算力与任务场景，从源头避免大模型算力浪费。 - [13:10](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=790s) 智能体上线前需进行覆盖文本、语音及视频的多模态 VAPT 渗透测试与合规监控，并通过私有部署的“主权 AI”隔离保护企业核心机密与专有算法。 - [14:38](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=878s) 实现智能体决策推理链路的全程可解释性，推行开发者级的 Token 成本精细化核算，并在立项初期绑定严格量化的 KPI 指标以追踪真实业务回报。 - [16:27](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=987s) NUEAC 平台通过顶层“元智能体”提供 24 小时主动诊断，智能提出 SLM 降级替代建议，并自动纠错系统提示词与架构代码设计。 - [16:27](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=987s) 企业 AI 规模化落地必须遵循“健全治理（Governance） → 深度优化（Optimization） → 规模化扩展（Scale & Adoption）”的三步演进法则。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=4s) 企业技术浪潮的历史镜像与 AI 治理的紧迫性 企业在拥抱新兴技术时，往往会重复陷入“先扩张、后治理”的历史周期。回顾数字化进程的第一波浪潮，个人电脑（PC）等计算终端刚进入企业生态时，业务灵活性与数字化进程大幅提升，但建立正确的企业级治理体系足足耗费了十多年时间。当时，计算资源不再局限于少数专业 IT 人员，而是普及到了每位员工的办公桌；随之而来的软盘、早期移动存储设备让数据进出完全失控，企业根本无法掌握员工在工位上究竟是在处理公事还是私事，公司资源被大量滥用。 随后的第二波互联网浪潮同样重演了这一幕。员工在第一天就获得了访问外网和电子邮件的权限，虽然这带来了业务上的便利，但在个人行为边界、访问内容控制以及数据防泄漏方面同样缺乏初期治理。到了第三波云计算与“全民开发者”（Citizen Developer）兴起阶段，业务人员开始依托 SharePoint 等低代码或云端工具自行构建应用，导致云成本激增，治理机制依旧是在混乱产生后才匆忙补救。 如今步入 AI 时代，企业领导者面临着核心抉择：是吸取前三次技术浪潮的教训在早期介入治理，还是继续拖延并在失控后再做弥补？现场针对企业治理核心挑战的实时互动调研显示，政策合规、FinOps 成本监控以及负责任 AI（Responsible AI）是当前企业领导者最广泛关注但尚未解决的痛点。 ```mermaid flowchart LR accTitle: 企业技术演进中的治理滞后周期 subgraph Waves["企业技术演化浪潮"] W1["第一波：个人电脑终端<br/>（软盘/数据进出失控）"] W2["第二波：互联网普及<br/>（网络访问与外发缺乏管控）"] W3["第三波：云与平民开发者<br/>（成本激增/自建系统泛滥）"] W4["第四波：生成式与自主 AI<br/>（当下阶段）"] end W1 --> W2 --> W3 --> W4 W4 -.-> Risk["历史治理教训：<br/>滞后治理带来巨大企业风险"] ``` --- ### [03:09](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=189s) 失控 AI 带来的企业盲区与行业惨痛教训 当前行业中由于缺乏 AI 治理而导致的灾难性案例屡见不鲜： - **预算失控与 Token 耗尽**：某企业由于缺乏支出可视性与管控机制，在短短四个月内就耗尽了全部预定的 Token 额度，彻底击穿了全年的 AI 预算。 - **6200 万美元投资直接核销**：另一家企业在累计投入 6200 万美元后，因无法衡量投资回报率（ROI）、缺乏明确业务 KPI 与交付成效验证，最终不得不将整个 AI 项目整体注销并直接关停。 - **核心知识产权（IP）泄露**：已有超过五家顶尖跨国企业遭遇严重的数据安全事故，专有知识产权被员工随意上传至公共大模型工具中，导致核心机密暴露于公共领域并引发法律诉讼。 - **“影子 AI”（Shadow AI）暗中蔓延**：类似于当年的影子 IT，海量新兴 AI 工具层出不穷，员工正通过各种未经批准的途径私下访问使用。如今的工具选型权从集中的 IT 部门下沉至业务部门负责人，导致组织内部充斥着完全不可视的“隐形 AI 采纳”。 - **自主智能体决策失误与数据覆灭**：某上线自主 AI Agent（智能体）的企业遭遇严重事故，智能体执行了荒谬的自动化指令，彻底删除了系统内长达三个月的交易记录与客户数据，最后企业不得不耗费巨大的人工成本进行灾难性手动恢复。 - **“氛围编程”（Vibe Coding）的安全隐患**：AI 辅助生成的代码虽然能极大加快开发速度，但统计表明超过 40% 的此类代码方案仅能停留在概念验证（PoC）阶段，其潜藏的逻辑漏洞与安全隐患根本达不到企业级生产环境的标准。 --- ### [05:52](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=352s) 碎片化 AI 孤岛与六大系统性企业风险 现代企业的技术架构极其复杂，没有任何单一厂商的工具或单一的大语言模型（LLM）能满足全部需求。各类独立软件开发商（OEM）正争先恐后推出各自的原生智能体，逼迫企业在此基础上自建定制化 AI 智能体来填补业务缝隙。 这种“智能体大爆炸”直接放大了企业的盲区风险： - **成本难以追踪与预算超支**：缺乏细粒度的用量监控与限额机制。 - **不可信 AI 与问责机制缺失**：模型产生幻觉、输出有害内容，而在出现差错时，究竟由业务部门还是 IT 部门担责界限模糊。 - **重复建设与协同失效**：各业务部门各自为政，在彼此不沟通的情况下，分别利用不同工具构建功能完全相同的生成式 AI 应用。 - **采纳度低下**：企业寄希望于所有普通员工自动成为合格的“提示词工程师（Prompt Engineer）”，这在现实中脱节严重，缺乏科学的用户采纳度衡量体系。 - **整体投资回报（ROI）归零风险**：缺乏对真实业务价值转化的追踪，巨额 AI 资金投入面临沉没风险。 --- ### [07:35](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=455s) 构建统一治理层（Unified Governance Layer）的核心支柱 面对来自市场竞争压力、董事会考核指标以及 AI 战略雄心的驱动，企业不应因噎废食，而是必须建立全方位的“统一治理层”。治理的内涵远超单纯的法规合规性，它需要覆盖六大关键维度： ```mermaid flowchart TD accTitle: 统一 AI 治理层的六大核心能力 Gov["统一 AI 治理层<br/>(Unified Governance Layer)"] Gov --> S1["影子 AI 监测<br/>(Shadow AI: 识别未授权工具/异常访问)"] Gov --> S2["负责任与合规 AI<br/>(Responsible AI: 伦理/无毒性/反幻觉)"] Gov --> S3["AI FinOps 成本工程<br/>(个人及部门 Token/许可配额能见度)"] Gov --> S4["可解释性与信任机制<br/>(Explainability: 高准确度输出验证)"] Gov --> S5["ROI 价值管理体系<br/>(追踪业务 KPI 与真实价值转化)"] Gov --> S6["零数据复制架构<br/>(Zero Data Copy & 安全互操作性)"] ``` - **影子 AI 防护（Shadow AI Detection）**：系统必须能够清晰识别组织内授权与未授权的工具，检测是否存在员工未经授权跨权访问授权工具的行为，并在新工具涌现时实时向安全运营中心（SOC）和基础设施团队告警。 - **负责任 AI（Responsible AI）**：建立道德与信任策略框架，从机制上杜绝输出毒性与模型幻觉，确保生成内容的真实准确。 - **AI 财务工程（FinOps）**：穿透许可证与 Token 实际消耗的黑盒。不能只看购买了多少软件 License，而要实时追踪特定团队或个人是否真正消耗了与其成本对等的 Token 算力。 - **ROI 与可解释性闭环**：验证 AI 是否真正达成了立项时的业务目标，让每一次决策链路具备透明的可解释度。 --- ### [09:45](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=585s) 多角色视角的治理诉求与企业内部实践 一个合格的 AI 治理平台必须同时兼顾企业内不同关键角色（Personas）的诉求差异： | 角色分类 | 关注核心与关键指标 (KPIs) | |:--- |:--- | | **业务与经营管理层**<br/>(CEO, CFO, COO, 业务部门主管) | 价值变现与最终 ROI、财务支出优化与成本控制、运营卓越性、基于企业整体战略的 AI 场景优先级排序。 | | **技术与架构治理层**<br/>(CIO, CTO, CISO, CDO, 合规官) | 践行“零数据复制架构”（Zero Data Copy Architecture）以避免数据碎片化、跨系统互操作性（Interoperability）、伦理规范与企业级统一安全防线。 | 在演讲尾声，讲者分享了其所在企业 YASH Technologies 内部的真实历程：起初，CEO 大力倡导拥抱 AI，要求所有业务及职能支持部门的主管自主推进落地；然而完全放开的后果与行业痛点如出一辙——云端账单失控、影子 AI 遍地开花、功能组件严重重复。企业随即意识到放任自流的危害，进而启动自上而下的系统性治理架构设计。 ### [11:47](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=707s) 智能体身份管理与多模型治理架构 在生成式 AI 向智能体 AI（Agentic AI）演进的过程中，企业必须在赋予员工灵活构建 AI 应用自由的同时，建立起全方位的防护与治理体系。传统的基于角色的访问控制（RBAC）仅能管理人类用户的权限，而当不同智能体之间频繁通信与协作时，系统必须具备专属的“智能体身份管理”（Agent Identity Management），实现 Agent 级别的精细权限控制。 为了应对多样化的业务场景，平台并不对所有任务一律采用大语言模型（LLM），而是构建了一个汇聚大模型、小语言模型（SLM）和领域专用模型的“LLM Zoo”，并结合模型上下文协议（MCP）连接器，无缝打通 SAP、Salesforce 等核心企业生态系统。从源头上解决“影子 AI”（Shadow AI）问题，确保任何进入企业生态的模型与智能体都能被第一时间识别和纳管。 ```mermaid flowchart TD accTitle: NUEAC AI Agent Governance Architecture subgraph Control ["身份与接入控制 (Identity & Access)"] A["企业用户 / 开发者"] --> C["RBAC 权限管理"] B["自主智能体 (AI Agents)"] --> D["智能体身份管理 (Agent IAM)"] end subgraph Core ["统一平台核心 (NUEAC Platform)"] C & D --> E["治理与合规网关"] E --> F["LLM Zoo (LLM / SLM / 领域模型)"] E --> G["MCP 连接器 (SAP / Salesforce)"] E --> H["主权 AI (Sovereign AI 私有资产)"] end subgraph Intelligence ["可观测与智能优化 (Observability & Optimization)"] E --> I["可解释性与安全审计 (VAPT/PII)"] E --> J["AI FinOps 成本与 Token 监控"] I & J --> K["智能元智能体 (代码/提示词/降本建议)"] end ``` - **智能体级 RBAC**：不仅管理人访问系统，还对智能体之间的跨系统调用实行身份识别与行为鉴权。 - **多模型协同矩阵（LLM Zoo）**：按需匹配 LLM、SLM 与行业垂直模型，避免大模型算力浪费。 - **MCP 生态集成**：通过标准化连接器安全对接 SAP、Salesforce 等底层业务系统。 --- ### [13:10](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=790s) 安全渗透测试、合规监控与主权 AI 保护 在智能体正式上线前，必须经历严格的安全准入评测。这类似于早期 Web 时代的漏洞扫描与渗透测试（VAPT），但在智能体时代，重点转向针对智能体特性的安全加固——全面检测文本、语音、视频等多模态交互中的个人身份信息（PII）泄露隐患，评估提示词注入风险，并验证输出的负责任程度。 上线后，平台会提供针对 HIPAA、SOC 2、GDPR 等行业法规与标准的持续合规监控，实时拦截有害内容，抑制幻觉并保证公平性。对于企业的核心知识产权（IP），如专有算法、客户投标策略与核心销售流程，平台通过部署自托管的主权 AI（Sovereign AI）进行隔离保护，确保机密数据绝不外流至公共模型。 - **多模态 VAPT 渗透测试**：覆盖文本、语音及视频全维度的 PII 敏感信息拦截与智能体漏洞扫描。 - **动态合规与负责任 AI**：全流程实时监测毒性（Toxicity）、偏见、幻觉与跨国合规（GDPR/HIPAA/SOC 2）。 - **主权 AI 资产隔离**：私有化托管核心提案模型与算法，严防企业核心商业机密外泄。 --- ### [14:38](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=878s) 可解释性、FinOps 成本精细化核算与 ROI 追踪 随着自主智能体在复杂流程中自主决策，企业必须具备清晰的“可解释性”（Explainability），深入洞察智能体做出每一步操作背后的推理逻辑。平台将监控粒度细化至单个开发者与具体用例级别，全流程追踪 Token 消耗量与实时成本，构建完整的 AI FinOps 治理闭环。 针对 AI 落地的实际成效，平台设立了量化的采纳度（Adoption）指标，对比商业授权软件与自研定制方案的使用活跃度，及时发现技能短板并跟进培训。在商业价值层面，从识别并立项每一个智能体的第一天起，业务团队就必须明确回答“为什么需要”以及“成功的具体 KPI 是什么”，杜绝“主观感觉效率提升了 30%”的模糊评估。此外，该治理架构不仅适用于云端数字 AI，也全面延伸支持边缘计算设备与物理 AI（Physical AI）。 - **决策链条可解释性**：完整回溯智能体自主推理路径与操作动因。 - **开发者级 FinOps**：精准定位各个项目、模型调用的 Token 消耗与费用归属。 - **严格的 KPI 准入机制**：智能体上线前必须绑定明确的 ROI 衡量标准，全周期跟踪落地收益。 - **物理与边缘 AI 覆盖**：统一纳管云端数字智能体与边缘端物理硬件智能体。 --- ### [16:27](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=987s) NUEAC 平台与智能元治理优化 为了使整个治理体系具备自适应与持续优化的能力，平台在顶层部署了一个智能层（Intelligent Layer），由专门的 AI 元智能体对上述维度进行 24 小时不间断的自主监控与主动诊断。当监测到某项任务成本过高且未充分利用大模型能力时，系统会主动建议降级切换至成本更低的小语言模型（SLM），甚至深入代码级与提示词级，自动提供架构设计调优和系统提示词（System Prompt）纠错建议。 这项完整的治理解决方案被命名为 **NUEAC**（Next-Generation Enterprise Unified Platform for AI Agent Control，下一代企业统一 AI 智能体控制平台）。演讲在最后向企业发出呼吁：“切勿让 AI 盲区阻碍了你的 AI 雄心。”企业唯有先建立正确的治理基石，落实深度的成本与架构优化，才能真正实现 AI 规模化采纳与长期业务增长。 - **元智能体主动推荐**：自动捕捉成本异常，建议模型选型替换（如 LLM 切换为 SLM）。 - **代码与提示词级诊断**：针对智能体架构模式、调用链以及系统提示词提出纠错与优化建议。 - **发展路径法则**：严格遵循“健全治理（Governance） $\rightarrow$ 深度优化（Optimization） $\rightarrow$ 规模化扩展（Scale & Adoption）”的三步演进路径。 - **现场交流互动**：主讲人 Nitin Gupta 邀请与会嘉宾移步 1051 号展台，观看 NUEAC 平台的实时现场演示并展开深入讨论。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=4s) 企业在拥抱新兴技术时，往往会重复陷入“先扩张、后治理”的历史周期。 — Nitin Gupta > [00:04](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=4s) 建立正确的企业级治理体系足足耗费了十多年时间。 — Nitin Gupta > [03:09](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=189s) 某企业由于缺乏支出可视性与管控机制，在短短四个月内就耗尽了全部预定的 Token 额度，彻底击穿了全年的 AI 预算。 — Nitin Gupta > [03:09](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=189s) 另一家企业在累计投入 6200 万美元后，因无法衡量投资回报率（ROI）、缺乏明确业务 KPI 与交付成效验证，最终不得不将整个 AI 项目整体注销并直接关停。 — Nitin Gupta > [03:09](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=189s) 已有超过五家顶尖跨国企业遭遇严重的数据安全事故，专有知识产权被员工随意上传至公共大模型工具中，导致核心机密暴露于公共领域并引发法律诉讼。 — Nitin Gupta > [03:09](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=189s) 某上线自主 AI Agent（智能体）的企业遭遇严重事故，智能体执行了荒谬的自动化指令，彻底删除了系统内长达三个月的交易记录与客户数据。 — Nitin Gupta > [05:52](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=352s) 现代企业的技术架构极其复杂，没有任何单一厂商的工具或单一的大语言模型（LLM）能满足全部需求。 — Nitin Gupta > [07:35](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=455s) 治理的内涵远超单纯的法规合规性，它需要覆盖六大关键维度。 — Nitin Gupta > [07:35](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=455s) 不能只看购买了多少软件 License，而要实时追踪特定团队或个人是否真正消耗了与其成本对等的 Token 算力。 — Nitin Gupta > [11:47](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=707s) 传统的基于角色的访问控制（RBAC）仅能管理人类用户的权限，而当不同智能体之间频繁通信与协作时，系统必须具备专属的“智能体身份管理”。 — Nitin Gupta > [14:38](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=878s) 从识别并立项每一个智能体的第一天起，业务团队就必须明确回答“为什么需要”以及“成功的具体 KPI 是什么”，杜绝“主观感觉效率提升了 30%”的模糊评估。 — Nitin Gupta > [16:27](https://www.youtube.com/watch?v=MAlpQ5cjqFE&t=987s) 切勿让 AI 盲区阻碍了你的 AI 雄心。 — Nitin Gupta ## Mentioned Nitin Gupta, YASH Technologies, Ai4, SharePoint, SAP, Salesforce, NUEAC, Model Context Protocol (MCP), GDPR, HIPAA, SOC 2 ## Video description Nitin Gupta - Head of AI & Digital, YASH Technologies The rise of autonomous AI agents is transforming business, but unexplained, disconnected deployments are creating chaos. These AI blindspots lead to runaway costs, compliance failures, and severe data vulnerabilities. This session breaks down the hidden dangers of "shadow AI" and provides a roadmap to fix it. Discover how to deploy centralized governance and real-time observability to ensure your AI agents operate securely, transparently, and completely aligned with enterprise standards. [Source](https://www.youtube.com/watch?v=MAlpQ5cjqFE)

[0:04] Nitin Gupta: 企业在拥抱新兴技术时，往往会重复陷入“先扩张、后治理”的历史周期。 [0:04] Nitin Gupta: 建立正确的企业级治理体系足足耗费了十多年时间。 [3:09] Nitin Gupta: 某企业由于缺乏支出可视性与管控机制，在短短四个月内就耗尽了全部预定的 Token 额度，彻底击穿了全年的 AI 预算。 [3:09] Nitin Gupta: 另一家企业在累计投入 6200 万美元后，因无法衡量投资回报率（ROI）、缺乏明确业务 KPI 与交付成效验证，最终不得不将整个 AI 项目整体注销并直接关停。 [3:09] Nitin Gupta: 已有超过五家顶尖跨国企业遭遇严重的数据安全事故，专有知识产权被员工随意上传至公共大模型工具中，导致核心机密暴露于公共领域并引发法律诉讼。 [3:09] Nitin Gupta: 某上线自主 AI Agent（智能体）的企业遭遇严重事故，智能体执行了荒谬的自动化指令，彻底删除了系统内长达三个月的交易记录与客户数据。 [5:52] Nitin Gupta: 现代企业的技术架构极其复杂，没有任何单一厂商的工具或单一的大语言模型（LLM）能满足全部需求。 [7:35] Nitin Gupta: 治理的内涵远超单纯的法规合规性，它需要覆盖六大关键维度。 [7:35] Nitin Gupta: 不能只看购买了多少软件 License，而要实时追踪特定团队或个人是否真正消耗了与其成本对等的 Token 算力。 [11:47] Nitin Gupta: 传统的基于角色的访问控制（RBAC）仅能管理人类用户的权限，而当不同智能体之间频繁通信与协作时，系统必须具备专属的“智能体身份管理”。 [14:38] Nitin Gupta: 从识别并立项每一个智能体的第一天起，业务团队就必须明确回答“为什么需要”以及“成功的具体 KPI 是什么”，杜绝“主观感觉效率提升了 30%”的模糊评估。 [16:27] Nitin Gupta: 切勿让 AI 盲区阻碍了你的 AI 雄心。

[0:06] Nitin Gupta: don't let the AI blind spot derail your AI ambition.

[0:04] So um today's topic I will cover is like  why ungoverned AI is becoming a more of an [0:09] enterprise risk. We'll also talk about um what are  the these blind spot which are actually impacting [0:15] the organizations to slow down on the AI journey  and most importantly how do we mitigate it right [0:22] so I'll take you uh to the journey where  like it all began right there was a first [0:27] wave I I'm sure like you know these terminals  which we used to call as personal computers [0:32] when these came into the enterprise that was the  first wave where uh things started becoming more [0:38] dynamic when digitization to start happening and  the challenge what we faced during that era was [0:44] it took almost more than a decade to bring the  right governance into the enterprise ecosystem [0:50] because it was not restricted to only the IT folks  but also to the all the employees where each and [0:55] every employee got desktops uh at their computer  at their workplace and the problem was governance [1:02] what data was coming in what data was going out  there was a era of floppy disk uh when drives came [1:08] later on but there was no governance in terms  of what employees are doing on a day-to-day [1:13] basis whether doing it personal work official  work but company resources were getting used [1:19] moving on the second wave came when the internet  came right same thing there was no governance [1:25] on day one every employee got the access to  online emails everything where they were using [1:32] of course for the official purpose but there was  no govern in terms of what they are doing on the [1:36] personal front what they're accessing saying what  the controls are and then same thing happened [1:41] in the third uh stage where the [snorts] citizen  developer program got got rolled out when the [1:47] cloud came in the cloud cost went up everybody  was kind of building their own applications being [1:52] at SharePoint or being at any other tools and  the same thing happened governance came later [1:58] question here is now we're in  the living uh in the age of AI [2:03] are we going to learn from our earlier  mistake and bring the governness early on [2:07] or we can still go slow and then bring  the governance later on. Now before I [2:13] deep dive into this, I have a quick uh poll which  if you can just scan the QR code from your mobile [2:24] and the question is uh already on the screen.  So it has a two-step process where you just have [2:29] to enter your name, email id and this u mobile  number. So it has six or seven answers. You can [2:36] have a multiple choice this thing where you can  see that what are the key challenges as you as an [2:42] enterprise leaders are facing on the governance  front and I can see the results coming up. [2:51] Of course the governance is high on  the policies, regulations, PHOPS, [2:58] responsible AI. So pretty much yeah it's all  over right like so everybody feels like one or [3:04] the other option is important. Now let's  look at what's happening in the industry [3:09] and I'm I'll maybe present some of the case  studies or stories which you might have heard. [3:14] So this is one of the stories where like one  of the companies announced that they ran out [3:18] ran out of tokens within the first four  months. They blew up the complete AI budget [3:23] and if you look at it the reasoning there  was a blind spot where they do not have a [3:28] control on AI spend and they ran out of tokens and  budgets. Second story after spending $62 million [3:40] company just did a write off on that  AI project and shut it down. Why? [3:45] Because there was lack of governance on the ROI.  there was no clarity in terms of what they are [3:50] trying to achieve. There was no measures in terms  of whether they are achieving those KPIs or not. [3:57] Third one where again like the IP got leaked into  public domain and there are more than five plus [4:06] top tier companies who reported and filed the  case and again why same the lack of governance [4:13] where there were security leaks on  the IP and the proprietary information [4:20] few more examples where shadow AI  uh we'll cover that as well where [4:26] every now and then keep hearing that  like a lot of AI tools are coming [4:30] and employees are accessing it through one or the  other mechanism which is also becoming a problem [4:35] which also means the shadow AI is popping up  in the enterprises how shadow it used to come [4:41] and also the the decision-m is happening  at now at business level department head [4:45] level are deciding which tool to use what not and  compared the enterprise IT deciding the AI tools [4:51] and that is resulting into invisible  AI adoption in the organization [4:55] maybe a couple of more examples where um again if  you look at it the autonomous agents which are now [5:02] coming up they wiped out 3 months of uh customer  data completely taking some uh uh absurd decisions [5:11] and they deleted the whole database of 3 months  transactions customer data everything and then [5:15] company revived that whole thing by doing a manual  effort why again uh unt un untrusted AI decisions. [5:26] Lastly again like wipe coding we know it has been  becoming very popular but there are questions [5:32] and concerns that can we rely on wipe coding  does it bring in AI or security risk in the [5:38] company environment uh can we trust those white  coding solutions and we saw that like more than [5:43] 40% of those white coded solutions are good  for pilots but may not be enterprise grade [5:52] on the other side if you look at it um every  enterprise is a complex landscape where [5:58] one tool does not suffice enterprise needs. There  are lot of OEMs who are also coming into the race [6:04] of AI and kind of offering their own agents.  So and on top of it we know that one tool and [6:11] one LLM is not going to solve the problem. Hence  enterprise are also building custom AI agents on [6:17] top of it and trying to bridge the gap and build  the use cases which they are aiming for. Now when [6:24] we look at it one side like a lot of agents are  coming we already saw some of blind spot what is [6:30] what is it that doing to the enterprises it's  creating the whole lot of risk in the blind spot [6:36] which is actually giving the right directly impact  on the business side being it a budget overrun [6:42] being it untractable cost being it a shadow AI  low trust low trust in terms of can we trust um [6:52] the outcome of AI in terms of accuracy the  responsible AI um who owns the decision who owns [6:59] the accountability is it business who is deciding  is it IT who's deciding so lack of accountability [7:04] right duplication [snorts] of solutions  one department doing the same solution of [7:09] generative AI and somebody else building the same  thing with different nobody talking to each other [7:13] so resulting into duplication poor adoption we are  expecting each and every employee to become a good [7:19] prompt engineer which is not going to happens.  How do you kind of ensure the adoption metrics [7:23] are also measured and eventually when companies  are investing so high are they seeing the overall [7:29] returns on investment and if it's not coming there  is a complete risk on the investment as well. [7:35] The question is yeah these are all great now [7:39] what's the uh what's the ask right so on one  side we know why companies are going for AI [7:46] there's a competition pressure there's a board  mandate um and there is an AI ambition we want [7:50] to do something and what we want to achieve is  of course the competitive edge we want to get ROI [7:56] but these are the blind spots which are  holding us back right now how do we mitigate it [8:03] this is our philosophy where Like we believe that  we need a unified governance layer and when by [8:09] governance we just don't mean the compliance  side of it. It is much more bigger aspect [8:14] and if you have the right governance layer being  it a shadow AI it starts from shadow AI where like [8:19] what you are using uh in in the whole  enterprise what are the authorized [8:24] tools what are the unauthorized tool who are  unauthorizedly accessing the authorized tools [8:28] all those thing getting covered into shadow  AISPs where you are getting direct alerts to [8:34] your sock team and your infrastructure team  to say that these these new tools are coming [8:38] because we can't keep up to the pace with  the technology is changing responsible AI. [8:43] Yes, we need compliances. We need policies  to kind of govern AI, how to use ethically, [8:48] how to use it trustworthy. We need to also bring  in the PHOPS part of it. How much are we spending? [8:55] Do we even have the visibility in terms of how  much budget each individuals are consuming on AI? [9:01] Yes, we have given the licenses, but do we know  whether that license cost is actually getting used [9:06] to whatever X number of dollars and employees are  actually consuming those many tokens? Do we have [9:12] that kind of visibility? Do we know that once  we invest X let's say X number of dollars in a [9:17] particular team or a department are we getting  enough ROI? So each of these components which [9:22] becomes important. Do we have the explanability?  Do we have the complete ROI governance and trusted [9:28] ecosystem where we can rely blindly on saying yeah  AI is doing good and I can confidently say that [9:34] the accuracy is high there's no hallucination  there's no toxicity in built into those AI. [9:39] So the governance overall all  these dimensions becomes important. [9:45] So when we have talk about governance  any good governance platform [9:49] there is always different personas in the  organization they will expect certain things [9:54] whether it's CEO uh CFO COO or even the business  leaders from the good governance platform they [10:01] will always aim for looking for certain KPIs  which suits their need right whether it's terms [10:06] of value realization whether it comes down to  the financial control optimization of AI spend [10:13] whether it comes comes down to operational  excellence or ROI from a specific use case [10:17] even to a prioritization of how do I  prioritize as a business leader which [10:21] AI use case is relevant I need to align with my  business strategies before even I consider that [10:29] and when these questions comes there is a next  layer of people as well which are important [10:33] from the technology front of your CIOS your  CTO's your uh CISOs your compliance officer [10:40] and of course CIO and CDO in is my data getting  duplicated everywhere. So how do I kind of ensure [10:46] there is a zero data copy architecture from the  infrastructure point or the architecture point of [10:52] do I have interoperability in place in  my solutioning? How do I govern that? [10:57] Do I have a right ethical and responsible AI  implemented? Do I have the right security measures [11:02] at across the enterprises which governs each and  every AI initiative which we are doing. So each [11:09] personas uh are expecting certain things from the  government. So any good governance platform needs [11:13] to cover all these things and I like to kind of  cover maybe a small case study which we at Yashuh [11:20] did internally. So again like any other  organization our CEO came and said like [11:24] yeah let's adopt AI let's do something and he  asked all the business leaders and uh support [11:29] function leaders to kind of just start adopting  AI and eventually the same thing happened [11:36] cost went out shadow AI started popping up  duplication this thing and then one point we [11:41] realized that this is not going to be uh helpful  so we need to do something and we started looking [11:47] at the options and we started uh building a  solution around that. So the the core essence [11:53] is like we wanted the flexibility for employees to  build AI application. The generative AI finished [12:00] and then agentic AI was coming in. We wanted  ability to kind of have agentic AI also coming [12:06] into the ecosystem. Now when AI applications  and agentic AI applications are coming up, [12:10] we also wanted the role based access control which  is not limited to only users who is accessing what [12:17] but also access [snorts] control at the  agent level. when we talk about agent exper [12:21] so one agent when they are talking do we have the  agent uh identity management built into the system [12:27] so we build that and then knowing that the market  is going to throw more MCPS and more models [12:32] not every model is required to have a large  language model every use case needs to have [12:38] large language model we will have a requirement  of small language models domain language models [12:42] all coming together so we created LLM zoo which is  more relevant for our organization we also created [12:48] MCP where we can have different connectors to  SAP and uh Salesforce and different ecosystem [12:55] but predominantly when when the we talk about  these foundational factors we were trying to [13:00] kind of build something around which covers all  the earlier pointer right shadow AI we wanted [13:05] right governance in terms of whatever comes in the  organization are we able to catch that early on [13:10] right once you find that like there are new  agents coming up how are we ensuring that there [13:15] is a complete uh in the earlier days when website  was there we called as VAP test right pent test [13:21] how are we doing that tests for agents are we  securing them how are they how secured they are [13:27] how what kind of quality they are are they  responsibly designed are the PI information cap [13:32] captured properly do they have PI information  captured at the voice level video level or is [13:39] it only text so we do all these test before we  onboard once we onboard again there's a continuous [13:45] monitoring on the policy front if you are HIPPA  compliant or so compl client all those policies [13:50] GDPR is built into the platform where right from  uh uh the rollout of any agents starts governing [13:57] those agents on the responsible the trust the  quality factor in terms of the toxicity um [14:06] fairness hallucination all those things which  are required for any good platform to build in [14:12] sovereign um again like we as an  organization we have our own IP [14:18] where we do not want to kind of share it with the  rest of the world. For example, how we bit to our [14:21] customers or sales and other processes. So  we build our own sovereign AI uh models and [14:26] library as well which only kind of considers our  algorithms and proposals and this thing which we [14:33] don't want to share it out. So we have our own  hosted sovereign AI as well which was required. [14:38] Now when we talk about all these agents autonomous  agents we also thought like this will be another [14:44] element which you want explanability we want to  learn about why agents are taking certain actions [14:51] we want to understand the reasoning behind it  so we built the explanability in part of its [14:57] we went to an extent where like at a developer  level as individual level we can kind of pinpoint [15:02] what they are doing how much cost they are taking  what tokens how many tokens are spending what use [15:07] cases they are working on right from the arbback  to the complete governance of the phops for AI [15:15] adoption as I said like until unless the AI is  adopted well you can't say it's a successful AI [15:22] so we have the adoption metric so even if you have  multiple tools and multiple agents when it's given [15:26] to the users how are they doing adoption on the  licensed software versus customuilt solutions so [15:31] giving the complete visibility what's relevant  what's working out what's not working out do we [15:35] need more training etc And lastly the ROI putting  all this together ROI is the key thing where like [15:43] right from identification of the first agent  itself that we want the agent the first question [15:48] is why what are the success KPI which this agent  will bring or this AI solution will bring we ask [15:54] that up front in the defining and onboarding of  the agent itself what would make it successful [15:59] what's success criteria for you and then we track  it so that we are also not just saying that yeah [16:04] I got 30% productivity just because I feel so  right but there is a complete mechanism and most [16:10] importantly we we thought like yeah digital side  AI it's important but the days are not far that [16:17] we will have the same requirement for physical AI  so build the complete platform for digital side [16:22] and the edge AI or the physical devices as well  where like the complete platform or the governance [16:27] is there across the and lastly we also built a  intelligent layer on top of it where all these [16:34] eight dimensions are continuously be monitored by  another AI agents which kind of start recommending [16:39] you hey you know what like the cost is going high  you may want to consider changing and switching to [16:44] the new LLM which are less costly because you  are not using really the power of LLM the SLM [16:50] might suit you right there are different  architecture design uh the agent patterns [16:55] and prompting techniques it starts going into  the code level kind of recommendations as well [16:59] if the system prompts are not right it will start  recommending you the system prompt correctness as [17:06] Um we we called it as newac next generation  enterprise unified platform for AI agent control [17:14] and uh before I leave uh my this thing so  we say like don't let the AI blind spot [17:20] derail your AI ambition first we need  to have the right governance in place [17:25] which follows by the optimization moment you  have the right governance and you optimize your [17:31] complete uh AI spend then only you can scale  with right adoption and uh the scalability [17:38] with that um let's take a quick check um based  on what we presented what what comes to your [17:46] mind like what are your next steps towards AI  governance can we just scan one more time and just [17:51] see if uh there's a question out  there so yeah we are at booth 1051 [17:58] uh we you can have a live demo of newac if you're  interested um and learn more about and exchange uh [18:05] thoughts. Uh I know it's pretty much on the time,  but um if you have any questions and comments, [18:10] I'll be at the booth and then see you  around. Thank you very much. [applause]