Airia Finds Enterprise AI Deployments Exceed Internal Audits by 600%

Airia发现企业实际AI部署超出内部审计600% 除非企业部署自动化实时控制平面，否则不受监控的智能体蔓延与供应商模型弃用将引发严重的成本超支和安全漏洞。 Kevin Kiley详细阐述了快速且分散的AI采用如何在企业中引发普遍蔓延，并指出组织通常会将内部AI使用量低估400%至600%。他概述了关键运营风险，包括不可预测的Token成本、供应商服务中断、模型突然弃用、监管碎片化以及API密钥劫持等安全漏洞。为了重新获得可见性与控制权，Kiley介绍了Airia围绕发现、安全防护栏、治理和成本优化构建的集中式控制平面。他强调了自动化变更管理和动态预算阈值，这些机制可实现零停机时间的模型故障转移，并防止严重的财务超支。 企业AI已迅速从分散的试点采用和生产力排行榜，转向严重的运营成本冲击与不受监控的Token消耗。 (0:04) 针对确定性软件设计的传统治理审查委员会，正被非确定性的智能体行为、突发的供应商模型弃用以及基础设施中断所压垮。 (2:32) 类似欧盟《人工智能法案》（EU AI Act）和州级法律等全球监管规定，加上主权AI边界与活跃的零日漏洞利用，使缺乏管理的AI成为直接的运营隐患。 (4:09) 在影子AI工具、未经通知的SaaS供应商更新以及孤立API密钥的共同作用下，深度自动化发现持续表明，组织内的实际AI规模比领导层意识到的多出400%至600%。 (6:02) 全面的AI管理需要四大持续运营支柱：持续发现、主动运行时安全网关、自动化合规治理以及成本优化。 (8:44)

Rapid enterprise AI adoption has shifted from an era of uncritical experimentation into severe cost overruns, operational vulnerabilities, and regulatory exposure. AI governance firm Airia argues that organizations can no longer rely on manual oversight, requiring automated control planes to discover hidden deployments, enforce active guardrails, and orchestrate model traffic. Unmonitored Sprawl and the Enterprise Visibility Gap (0:04) Enterprise AI adoption progressed rapidly from basic chat interfaces and Microsoft Copilot licensing into embedded third-party software features, custom internal applications, and employee shadow AI. According to Airia, organizations attempting comprehensive audits consistently underestimate their actual AI footprint by 400% to 600%. This disparity stems from three primary blind spots: unsanctioned browser tools used by staff, incumbent enterprise software vendors quietly enabling generative capabilities within existing contracts, and siloed engineering teams running foundation model workloads without central IT oversight. This unmanaged proliferation has turned initial executive enthusiasm into acute financial shock. While companies recently incentivized model usage through internal leaderboards and promotional metrics, unmonitored token consumption and escalating infrastructure expenses have transformed tokenomics into an urgent budget liability that standard IT mechanisms fail to capture. Operational Instability and Accelerating Regulatory Pressures (2:32) Traditional AI Centers of Excellence and governance boards are failing because their evaluation frameworks were designed for deterministic software architectures. They struggle to assess probabilistic systems prone to model drift, dynamic learning, and unpredictable agent behaviors. This friction is compounded by vendor volatility. In one cited incident, Fable 5 users received only a 90-minute sunset notice at 5:00 p.m., forcing emergency migrations. Routine deprecations and peak-hour outages—such as disruptions exceeding 90 minutes at major providers like OpenAI and Anthropic—threaten production continuity for unhedged systems. Simultaneously, compliance requirements are shifting from voluntary guidance to legal mandates. Organizations face binding frameworks including the European Union AI Act, Singapore's governance guidelines, and five enacted US state laws. Cross-border operations must also navigate emerging sovereign AI mandates and jurisdictional restrictions against specific foreign models, alongside technical exploits disclosed by security researchers that bypass conventional perimeter defenses. Automated Change Management and Multi-Model Resilience (11:50) Relying on a single foundation model provider creates substantial vendor lock-in and eliminates pricing leverage during contract renewals. However, managing multi-model environments introduces severe technical instability. Subtle changes to prompts, integrations, or underlying model versions can alter output accuracy, response latency, temperature, and tool-calling execution. To manage this variability without incurring downtime, Airia outlines an automated five-step change management pipeline: offline evaluation, automated red teaming via adversarial agent swarms, formal human approval, shadow traffic monitoring, and risk analysis. Running continuous background "what-if" simulations against mirrored production traffic enables enterprises to validate emerging models against existing agent workloads, allowing automated failover routing when primary vendor systems fail, deprecate, or suffer security breaches. Financial Governance and Safeguards Against Runaway Costs (14:56) Financial waste frequently occurs when employees deploy trillion-parameter flagship models on routine tasks, such as generating slide decks. Managing consumption requires granular token tracking across business units, departments, projects, and individual users, backed by policy-driven thresholds. Airia illustrates a tiered escalation system: reaching 70% of an allocated budget triggers an automated user notification, an 80% threshold alerts the managing supervisor, and a 90% cap prompts automated throttling or dynamic rerouting to pre-validated, lower-cost model alternatives. Without automated guardrails, unmonitored systems remain vulnerable to catastrophic financial anomalies. Airia highlights a recent case where an enterprise client's Anthropic API key was compromised at 3:00 a.m. on a Saturday. Attackers executed massive unauthorized workloads, burning through $40,000 in unbudgeted spend before detection. Continuous runtime gateways are required to spot anomalous usage patterns, cut unauthorized access immediately, and prevent sudden financial losses. Airia, an enterprise AI governance platform that scaled to over 500 customers within 22 months of exiting stealth, asserts that continuous discovery and active orchestration are essential as systemic complexity compounds. The firm notes its inclusion in Gartner's inaugural AI Governance Magic Quadrant, but notes that organizations must ultimately balance the rigid security required for regulatory compliance with the operational agility that business units demand.

根据企业人工智能管理平台 Airia 的观察，企业在经历前期的无序采纳后，正面临影子 AI 泛滥、合规监管趋紧、基础设施中断以及 Token 成本失控等多重冲击。该机构认为，企业必须部署集资产发现、安全网关、持续治理与动态优化于一体的控制平面，以防范财务损失并维持业务连续性。 无序采纳引发成本剧增与认知偏差 (0:04) 企业引入人工智能的路径通常始于简单的内部对话实验，随后快速扩散至微软 Copilot 批量采购、传统 SaaS 供应商静默内嵌自主功能、定制化应用开发以及员工自行使用的影子工具。随着数月前高管团队展示使用率排行榜、埃森哲公开鼓励员工通过多用 AI 争取晋升的热潮退去，未受监控的 Token 消耗和基础设施开销导致运营预算严重超支。企业在缺乏集中治理的情况下，迅速从采纳狂热转向了“Token 经济学”带来的财务冲击。 影子资产泛滥加剧风险敞口 (6:02) Airia 指出，当管理层认为企业内部 AI 部署处于受控状态并提供已批准清单时，深度自动化发现往往揭示出实际使用规模比预期高出 400% 至 600%。这一巨大差距源于三个方面：未被登记的员工影子应用、第三方企业软件在原有合同中静默开启的生成式功能，以及工程部门内部孤立运行的基座模型工作负载。为了建立风险基准，必须持续测绘所有数据流向、基座模型、API 密钥、Model Context Protocol（MCP）服务器以及终端用户访问路径。 模型迭代与云端中断威胁业务连续性 (2:32) 传统面向确定性软件的治理委员会和卓越中心在应对概率性系统、模型漂移与自主智能体时往往疲于奔命。与此同时，外部基础设施的脆弱性进一步放大了运营风险。例如 Fable 5 曾在下午 5 点发出仅 90 分钟的下线通知，迫使团队紧急回滚；Anthropic 和 OpenAI 也在工作高峰期发生过超过一个半小时的服务中断。此外，许多模型供应商在核心业务上线前夕突然宣布弃用旧版本，迫使企业在数天或数周内完成迁移，严重威胁业务连续性。 主权合规与安全漏洞的双重挑战 (4:09) 在监管层面，全球合规要求正从理论准则转为法律义务。美国已有五个州出台具体 AI 法规，新加坡发布了治理指南，欧盟 AI 法案也提出了严格的披露要求，同时多个国家要求实行主权数据边界或限制特定境外模型。在技术防御方面，安全研究人员在 Black Hat 等会议上频繁披露能绕过传统边界防护的零日漏洞和注入技术。这些压力要求企业具备根据地域司法管辖区自动路由模型、实施运行时行为护栏与自动化红队对抗测试的能力。 多层预算控制与密钥失窃防护 (14:56) Token 成本膨胀往往源于任务与模型能力的不匹配，例如员工使用万亿参数的旗舰模型撰写基础演示文稿。Airia 建议实施动态多级预算策略，例如达到 70% 触发用户通知，80% 向主管报警，90% 则自动限制速率或将流量切换至低成本备选模型。缺乏此类监控可能引发灾难性后果，一家企业曾在一个周六凌晨 3 点遭黑客盗用 Anthropic API 密钥，短时间内被消耗 40,000 美元未授权工作负载，突显了自动化异常检测与即时断流机制的必要性。 自动化变更管理实现零停机故障转移 (11:50) 过度绑定单一供应商会削弱企业在续约谈判中的议价能力，但维护多模型环境又面临提示词变化、输出温度漂移和工具调用不稳定的挑战。通过在后台对生产流量进行离线“假设模拟”与影子流量分析，企业可以在不影响前端业务的情况下评估备选模型。一旦主供应商服务宕机、曝出安全后门或强制弃用模型，系统可以利用预先通过验证的模型执行自动化故障转移，实现零停机业务切换。 该报告由成立约 22 个月、入选 Gartner 首份 AI 治理魔力象限的 Airia 呈现。其核心论点在于将 AI 治理视为涵盖发现、安全、合规与成本优化的持续运营控制系统，但对于跨模型自动迁移中业务逻辑的实际损失率，以及小型企业部署该类治理架构的成本门槛，简报并未展开深入讨论。

## Background 该内容由 Ai4 于 2026 年 8 月 14 日在 YouTube 平台发布，由 Airia 首席执行官 Kevin Kiley 担任主讲。视频简介指出，随着 AI 和自主代理（Agent）的爆发式增长，传统安全手段难以应对新型风险，本次演讲旨在为 CIO 等企业管理者提供治理“AI 蔓延”（AI Sprawl）、防范新兴漏洞以及建立运营韧性的实战策略。 ## TL;DR 演讲系统梳理了企业 AI 从早期单点测试走向影子 AI 泛滥与 Token 成本失控的演变轨迹，指出传统确定性 IT 审查面对自主漂移的 Agent 已不堪重负。通过实际扫描，企业内部通常潜藏着 400% 到 600% 未被察觉的 AI 资产。Airia 提出了覆盖“资产发现、安全防护、审计治理与弹性优化”的四大管理支柱，并通过自动化“What-If”沙盒变更模拟与阶梯式预算智能降级路由，帮助企业在保障业务连续性的同时彻底防范突发巨额账单与安全合规风险。 ## Atmosphere & style 整场演讲节奏明快、务实且极具警示意味，展现出深厚的技术底蕴与企业级实战洞察。演讲者不谈空洞的理论假设，而是通过诸如凌晨 3 点 API 密钥被盗损失 4 万美元、模型突发下线仅留 90 分钟迁移窗口等真实商业案例，直击现场管理者的焦虑痛点。现场互动热烈、信息密度极高，呈现出一种在技术狂潮中为企业重塑控制秩序的专业与笃定感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=4s) 企业 AI 应用正经历从单点实验、Copilot 铺开，迅速演变至传统软件静默嵌入 AI 及员工自发使用影子 AI 的全方位蔓延阶段。 - [00:04](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=4s) 企业态度从早期的全员激进化使用转向对 Token 经济学与失控预算账单的焦虑，成本控制成为首要痛点。 - [02:32](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=152s) 针对确定性软件设立的传统 IT 审查与 AI 卓越中心面对自主学习、行为持续漂移（Drifting）的 Agent 已不堪重负。 - [02:32](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=152s) 模型突发下线、核心供应商服务宕机、强制版本废弃以及全球合规法规（如 EU AI Act 与主权 AI 限制）对企业业务连续性构成直接威胁。 - [06:02](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=362s) 企业实际存在的 AI 资产远超预期，深度扫描通常会发现比原本掌握的应用多出 400% 至 600% 的未报备隐藏资产。 - [08:44](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=524s) AI 全生命周期治理依赖于四大核心支柱：资产发现（Discovery）、安全防护（Security）、审计治理（Governance）与弹性优化（Optimization）。 - [08:44](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=524s) 安全防御需要结合 Agent 安全网关、自动化红队蜂群持续渗透测试以及关键决策节点的人类在环（Human-in-the-loop）审批机制。 - [11:50](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=710s) 通过“What-If”沙盒情景模拟测试进行变更管理，企业可在不影响生产的前提下验证新模型并实现即时容灾切换，避免供应商锁定。 - [14:56](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=896s) 引入阶梯式预算风控策略（如 70% 预警、80% 升级、90% 自动降级路由至低成本平替模型），可从底层杜绝预算被击穿的风险。 - [15:58](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=958s) 部署实时的 API 异常监控与熔断保护机制至关重要，能有效拦截因密钥泄露或异常调用导致的瞬时巨额财务损失。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=4s) AI 蔓延的演变：从实验试点到成本失控 AI 正在以不可逆转的态势融入各类企业的日常运营，但随之而来的“AI 蔓延”（AI Sprawl）已成为管理、安全与治理层面的紧迫难题。回顾过去几年的发展轨迹，多数企业的 AI 之旅始于几位先行者搭建的简易聊天应用，用于向管理层展示该技术赋能业务的潜力。随后，微软等巨头开始向大中型企业分发 Copilot 许可证，AI 进入局部测试阶段。紧接着，许多已在生产环境中运行多年的传统软件供应商，悄然在自身产品中嵌入了 AI 功能与自主代理（Agent）；由于这些软件早已掌握企业敏感数据且此前已通过采购审查，这一转变几乎完全绕过了传统的 IT 评估流程。 随后，AI 的应用迅速走向分散化与自主化：企业或是投入自研，或是聘请外部咨询公司构建定制化 AI 应用，各个业务部门也争相采购针对特定场景的点对点解决方案。与此同时，员工出于“希望工作更高效、交付更迅速”的初衷，纷纷将未经批准的外部 AI 工具带入工作环境（影子 AI）。这种全面开花的状态带来了严峻的治理与数据外泄隐患，而最直接的冲击则是成本的失控暴增。 在 Tokenomics（Token 经济学）成为焦点的当下，企业的态度发生了 180 度大转弯。就在半年到一年前，企业管理层还热衷于展示员工 AI 使用率的排行榜，埃森哲（Accenture）甚至公开宣布“不使用 AI 就无法获得晋升”；而如今，越来越多的企业却在为预算被迅速打穿、账单爆表而焦头烂额。 - **六阶段演进路径**：单点聊天测试 $\rightarrow$ 企业级 Copilot 铺开 $\rightarrow$ 传统供应商静默嵌入 AI $\rightarrow$ 定制化 AI 应用开发 $\rightarrow$ 部门级点解决方案采购 $\rightarrow$ 员工自主引入影子 AI。 - **企业心态剧变**：从激进鼓励全员调用、将其与职业晋升挂钩，迅速转变为面对失控的 Token 账单与预算超支时的焦虑与收紧。 --- ### [02:32](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=152s) 传统治理体系的崩溃与多维外部风险 当最初的试点扩展至全公司范围，企业原有的治理体系已被推向崩溃边缘。许多企业设立的“AI 卓越中心”（Center of Excellence）每隔数周开会审查一次应用评估，面对当前爆炸式的 AI 落地速度早已不堪重负。更根本的矛盾在于，传统的 IT 审查流程是为“确定性软件”设计的，面对具有自主学习能力、行为不断演化漂移（Drifting）的 AI 模型与 Agent，传统方法根本无法评估其安全性与鲁棒性。 与此同时，基础设施与外部环境的不确定性进一步加剧了运营风险。以模型可用性为例，此前曾发生过某模型（如 Fable 5）在下午 5 点突发通知、仅给用户留出 90 分钟迁移窗口便彻底下线的极端情况；供应商频繁推出新版本并强制废弃旧版本，让企业耗费数月完成的验证测试瞬间失效。此外，主权 AI（Sovereign AI）的合规壁垒要求跨国企业必须按司法管辖区实施模型路由；头部模型供应商（如 Anthropic 发生过约 1.5 小时的服务中断，OpenAI 亦屡见故障）的频繁宕机更对核心业务连续性构成威胁。 在安全与法律层面，各类新型对抗攻击与 Agent 漏洞不断绕过传统防御体系，而监管环境也在急剧收紧。仅在美国就已有 5 个州出台了不同程度的 AI 法规，另有数个州正在立法审议中；叠加欧盟《人工智能法案》（EU AI Act）与新加坡治理框架等国际法规，合规压力成倍上升。Kevin Kiley 强调："这些风险绝非学术界的理论探讨，而是切切实实正在发生的商业现实。" ```mermaid flowchart TD accTitle: AI 蔓延下传统企业治理面临的多维风险挑战 subgraph 内部治理困境 A["传统确定性审查流程"] --> B["AI 卓越中心不堪重负"] B --> C["模型行为漂移与不可控 Agent"] end subgraph 外部运营与供应链风险 D["模型突发下线与强制版本弃用"] E["供应商核心服务宕机<br>(如 Anthropic / OpenAI)"] F["主权 AI 限制与跨国路由要求"] end subgraph 安全与合规风暴 G["黑客新型漏洞攻击<br>(Black Hat 安全威胁)"] H["全球监管收紧<br>(EU AI Act / 各州法案)"] end C --> I["企业核心业务连续性受阻"] D --> I E --> I F --> I G --> I H --> I ``` --- ### [06:02](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=362s) 全面 AI 管理平台与资产发现：揭开 400% 到 600% 的隐藏蔓延 面对高度复杂的 AI 资产组合（涵盖供应商、模型、数据集成和合规法规），企业必须建立一套集中式的控制平面。Kevin 强调：“无论你现在处于什么阶段，你今天的 AI 架构都是它未来最简单的一刻，后续只会呈指数级变复杂。”这正是 Airia 致力于打造独立 AI 管理与治理层的原因——它不是一次性的咨询项目，而是一个常态化运行的集中控制平台。 治理的第一步是资产发现（Discovery）。当 Airia 与企业管理层初次沟通时，客户通常会提供一份自认为记录详尽的已批准应用清单，并坚信企业内部环境处于受控状态。然而，在实际部署扫描后： > "平均而言，我们会在企业内部多发现 400% 到 600% 之前未被察觉的 AI 资产，这是一个非常真实的统计数据。" 这些隐藏资产主要来自三大板块：员工未报备使用的影子 AI、现有 SaaS 供应商未做显式通知便开启的嵌入式 AI 能力，以及分散在各业务线的原生模型调用。Airia 能够在数天内建立起覆盖全公司数据、模型、供应商、服务商和使用人员的完整映射图谱，为后续的安全介入奠定基础。 --- ### [08:44](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=524s) Airia 的四大支柱：发现、安全、治理与优化 Airia 的核心产品体系由四大支柱构成，覆盖 AI 生命周期的全方位管理： ```mermaid flowchart LR accTitle: Airia 平台四大核心支柱架构 subgraph Pillar1["1. 资产发现 Discovery"] A1["9 大向量深度扫描"] --> A2["全量资产清单与合规暴露评估"] end subgraph Pillar2["2. 安全防护 Security"] B1["Agent 行为约束 & 网关"] --> B2["自动化红队 Agent 蜂群渗透"] B2 --> B3["人类在环 (Human-in-the-loop) 审批"] end subgraph Pillar3["3. 审计治理 Governance"] C1["持续合规报告生成"] --> C2["全链路审计追踪 & 可观测性"] end subgraph Pillar4["4. 弹性优化 Optimization"] D1["故障自动切换 & 成本路由"] --> D2["Token 消耗追踪 & 内部计费"] end Pillar1 --> Pillar2 --> Pillar3 --> Pillar4 ``` 1. **资产发现（Discovery）**：通过跨越 9 个不同向量的深度扫描，识别企业内所有被批准或未被批准的 AI 使用行为，生成完整资产清单，精准评估监管暴露风险并指出亟待修复的短板。 2. **安全防护（Security）**：约束 Agent 行为与权限变更，防止其越权演化；在各类 AI 调用、集成及 MCP（Model Context Protocol）服务器间部署安全网关与护栏；利用自动化红队（Red Teaming）Agent 蜂群进行持续对抗性测试，在漏洞被恶意利用前主动发现；并在关键决策节点引入人类在环（Human-in-the-loop）监督审批。 3. **审计治理（Governance）**：为处于强监管行业的企业自动生成审计报告与合规证据链，满足全球监管机构的信息披露要求，提供全生命周期的可观测性与审计注册表。 4. **弹性优化（Optimization）**：保障系统韧性，在供应商服务中断时实现毫秒级自动切换；同时根据任务复杂度进行智能成本路由，防止员工“使用万亿参数的旗舰大模型去完成写幻灯片等简单任务”，并提供部门级调用追踪（Tracing）与内部计费（Chargeback）。 --- ### [11:50](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=710s) 变更管理与“What-If”沙盒模拟测试 许多企业在 AI 规划中忽略了“变更管理”（Change Management）。若企业将所有业务孤注一掷地绑定在单一模型供应商上，虽然短期内架构简单，但在商业续约谈判时将彻底丧失议价权。此外，模型频繁升级、Prompt 提示词迭代、后端数据连接变化都会对现有系统的输出质量与稳定性产生连锁反应。 为了解决这一痛点，Airia 在后台提供不影响生产环境的“What-If”自动化情景模拟功能。系统能够在后台对新模型进行多维度对比分析，评估其温度参数、响应精准度、行为变化以及潜在的额外 Token 成本。对于过去需要耗费数周甚至数月进行人工验证的企业而言，这种自动化的测试流水线避免了“刚完成漫长测试，旧模型即遭废弃”的尴尬局面。 - **五步变更保护程序**：评估（Evaluation） $\rightarrow$ 自动化红队对抗测试（Red Teaming） $\rightarrow$ 影子流量监控（Shadow Traffic Monitoring） $\rightarrow$ 审批上线（Approving） $\rightarrow$ 风险持续监控。 - **即时容灾切换**：当遭遇模型宕机、主权限制或突发后门漏洞时，企业能够基于预先完成的验证数据，将 Agent 与生产工作负载自信且无缝地切换至替代模型。 --- ### [14:56](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=896s) Tokenomics 与基于预算梯度的自动化智能路由 针对企业面临的成本失控问题，平台在项目、团队、业务单元乃至个人维度建立了多层级的财务控制与 Token 消耗监控体系。管理员可以根据预算消耗进度设置阶梯式的响应规则，实现精细化风控： - **70% 预算阈值**：触发系统预警通知，提示当前项目或用户 Token 消耗速率过快，存在超支风险。 - **80% 预算阈值**：告警自动升级，系统将告警邮件或通知推送给对应的直属业务主管介入关注。 - **90% 预算阈值（终极自动处置）**：系统触发自动化降级路由。如果该员工或 Agent 此前一直在调用昂贵的顶级旗舰模型，系统将在触碰该阈值后，自动将其请求无缝切换至预先验证过的免费或极低成本模型。 由于平台在前期已通过基准测试锁定了各场景下的平替模型，降级后的模型在确保任务准确率与性能基本一致的前提下，能够从底层彻底阻断预算击穿的风险。 ### [15:58](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=958s) 构建财务防护与模型智能路由机制 在 AI 管理中，设立财务安全防护机制至关重要，以防止企业遭遇意料之外的巨额账单。通过在系统中引入财务保护策略，即使某些任务可能会带来略高延迟、处理速度稍慢，或者需要转为夜间批处理，也能确保企业不会陷入被迫向 CFO 解释突发暴增开支的尴尬境地。演讲者分享了一个发生在刚过去的周六凌晨 3:00 的真实案例：某客户的 API 密钥不幸被窃取（团队仍在调查泄露源头），黑客利用该密钥通过 Anthropic 许可证短时间内跑了极高负荷的工作负载，直接导致 40,000 美元在瞬间蒸发。如果部署了完善的保护机制，平台不仅能够实时发出预警并自动关停异常调用，还能将请求智能路由至其他备用模型，从而彻底杜绝此类账单死角的发生。 ```mermaid flowchart TD accTitle: 财务防护与模型路由流程图 subgraph Trigger["异常检测与触发"] A["监测到 API 密钥异常调用"] --> B{"预警与阈值判断"} end subgraph Mitigation["自动化拦截与路由"] B -- 超出预算阈值 --> C["自动切断 API 调用并预警"] B -- 成本与性能平衡 --> D["智能路由至替代模型 / 批处理"] end C --> E["规避突发高额账单风险"] D --> E ``` - 案例细节：周六凌晨 3:00 密钥被盗，客户因 Anthropic API 密钥被盗用而在短时间内遭遇了 40,000 美元的非预期损失。 - 应对方案：通过平台策略实时监控、自动熔断停机，或将流量动态路由至低成本模型及夜间批处理模式。 --- ### [16:55](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=1015s) Airia 公司背景与行业权威认可 Airia 作为专注于 AI 管理领域的全球化平台，自结束隐形模式（Out of stealth）仅 22 个月以来，呈现出极其惊人的增长势头。公司目前在全球范围内已拥有超过 500 家客户以及近 200 名员工。在 Gartner 首次发布的 AI 治理魔力象限（Magic Quadrant for AI Governance）报告中，Airia 不仅入围并最终通过遴选，更被评为拥有同类厂商中最全面平台（most comprehensive platform）的企业。演讲者指出，Gartner 推出该魔力象限本身就验证了市场对 AI 治理和管理平台的强劲需求，而 Airia 的卓越表现正是源于其团队在网络安全、合规性及 IT 基础设施管理领域的深刻积累。 - 团队履历：演讲者及其合伙人拥有深厚的技术 background，此前曾共同参与打造了 AirWatch 和 OneTrust 两家知名企业。 - 行业报告：Airia 已获得该 Gartner AI 治理魔力象限报告的官方授权，并可免费向有需求的参会者提供完整研究副本。 --- ### [18:22](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=1102s) 落地资源与现场互动 为了帮助企业顺利推进 AI 治理与管理项目，Airia 提供了丰富的免费资源与服务，其中包括针对企业内部 AI 使用现状的“发现评估服务”（Discovery Engagements），帮助团队精准识别全公司范围内已存在的 AI sprawl（AI 蔓延）及未受控应用。在演讲尾声，面对现场众多拍照记录的听众，演讲者表达了诚挚的感谢，并邀请大家扫描屏幕上的二维码获取完整 PPT 演讲幻灯片，或在会后前往展厅现场的 Airia 展位进行深入交流和产品功能演示。 - 参会互动：听众可通过扫描现场 QR 码或留下邮箱获取演示 Slides。 - 展会活动：Airia 团队在展厅下方设有专门展位，现场提供系统 Demonstration，帮助企业规划可持续发展的 AI 管理路线图。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=4s) 不使用 AI 就无法获得晋升 — 埃森哲（Accenture） > [02:32](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=152s) 传统的 IT 审查流程是为“确定性软件”设计的，面对具有自主学习能力、行为不断演化漂移（Drifting）的 AI 模型与 Agent，传统方法根本无法评估其安全性与鲁棒性。 — Kevin Kiley > [02:32](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=152s) 这些风险绝非学术界的理论探讨，而是切切实实正在发生的商业现实。 — Kevin Kiley > [06:02](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=362s) 无论你现在处于什么阶段，你今天的 AI 架构都是它未来最简单的一刻，后续只会呈指数级变复杂。 — Kevin Kiley > [06:02](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=362s) 平均而言，我们会在企业内部多发现 400% 到 600% 之前未被察觉的 AI 资产，这是一个非常真实的统计数据。 — Kevin Kiley > [08:44](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=524s) 使用万亿参数的旗舰大模型去完成写幻灯片等简单任务 — Kevin Kiley > [11:50](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=710s) 若企业将所有业务孤注一掷地绑定在单一模型供应商上，虽然短期内架构简单，但在商业续约谈判时将彻底丧失议价权。 — Kevin Kiley > [15:58](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=958s) 某客户的 API 密钥不幸被窃取（团队仍在调查泄露源头），黑客利用该密钥通过 Anthropic 许可证短时间内跑了极高负荷的工作负载，直接导致 40,000 美元在瞬间蒸发。 — Kevin Kiley > [16:55](https://www.youtube.com/watch?v=pI8N9ZEhvn8&t=1015s) Gartner 推出该魔力象限本身就验证了市场对 AI 治理和管理平台的强劲需求，而 Airia 的卓越表现正是源于其团队在网络安全、合规性及 IT 基础设施管理领域的深刻积累。 — Kevin Kiley ## Mentioned Kevin Kiley, Airia, Ai4, Microsoft, Copilot, Accenture, Fable 5, Anthropic, OpenAI, EU AI Act, MCP (Model Context Protocol), Gartner, Magic Quadrant for AI Governance, AirWatch, OneTrust ## Video description Kevin Kiley - CEO, Airia The rapid explosion of AI and autonomous agents introduces risks that traditional security measures weren't built to address, placing newdemands on CIOs to govern the AI sprawl across the organization. In this session, we’ll arm you with practical strategies to identify and manage these emerging vulnerabilities and discuss effective strategies for maintaining control over AI agents. Attendees will learn actionable approaches for risk mitigation, governance, and operational resilience that defines successful AI transformation. [Source](https://www.youtube.com/watch?v=pI8N9ZEhvn8)

[0:04] 埃森哲（Accenture）: 不使用 AI 就无法获得晋升 [2:32] Kevin Kiley: 传统的 IT 审查流程是为“确定性软件”设计的，面对具有自主学习能力、行为不断演化漂移（Drifting）的 AI 模型与 Agent，传统方法根本无法评估其安全性与鲁棒性。 [2:32] Kevin Kiley: 这些风险绝非学术界的理论探讨，而是切切实实正在发生的商业现实。 [6:02] Kevin Kiley: 无论你现在处于什么阶段，你今天的 AI 架构都是它未来最简单的一刻，后续只会呈指数级变复杂。 [6:02] Kevin Kiley: 平均而言，我们会在企业内部多发现 400% 到 600% 之前未被察觉的 AI 资产，这是一个非常真实的统计数据。 [8:44] Kevin Kiley: 使用万亿参数的旗舰大模型去完成写幻灯片等简单任务 [11:50] Kevin Kiley: 若企业将所有业务孤注一掷地绑定在单一模型供应商上，虽然短期内架构简单，但在商业续约谈判时将彻底丧失议价权。 [15:58] Kevin Kiley: 某客户的 API 密钥不幸被窃取（团队仍在调查泄露源头），黑客利用该密钥通过 Anthropic 许可证短时间内跑了极高负荷的工作负载，直接导致 40,000 美元在瞬间蒸发。 [16:55] Kevin Kiley: Gartner 推出该魔力象限本身就验证了市场对 AI 治理和管理平台的强劲需求，而 Airia 的卓越表现正是源于其团队在网络安全、合规性及 IT 基础设施管理领域的深刻积累。

[2:32] Kevin Kiley: Guess what? You're about to go live. We've got the new version. You should really look at that one. [6:02] Kevin Kiley: as simple as it'll ever be [6:02] Kevin Kiley: somewhere between 4 to 600% more AI across the organization. [15:58] Kevin Kiley: in the CFO's office explaining why overnight we hit a $40,000 expenditure that nobody budgeted for.

[0:04] Very excited to speak to you about a topic that I  think is urgent and one that's probably growing in [0:08] importance for all of you in this room and that  is AI sprawl. As your organizations become more [0:14] and more dependent on AI in the work that each of  you are doing inside of your organization, this [0:20] problem is proliferating. How do we manage it? How  do we secure it? How do we keep up with the pace [0:26] of evolution that is only accelerating? And I'll  begin by describing that problem a little bit. [0:30] If we talk about what most organizations started  with a couple of years ago, it was perhaps just a [0:36] simple chat application that some of you or others  in your organization were sort of pioneering, [0:42] showing the power of how this could be applied  to your business and what you could do with it. [0:46] From there, you probably got some Copilot  licenses, Microsoft offering those to you if [0:50] you're a larger enterprise or really of almost any  size, you probably got some group to test with. [0:57] We then saw a lot of organizations their or  vendors embedding AI in tools that were already [1:03] in production. A lot of these were going right  past any sort of review process because you'd [1:07] already deployed them and they had your sensitive  data after years of using them. Now they've got [1:12] some level of agency or autonomy and and the risks  and management concerns that could come with that. [1:19] A lot of organizations then expanded into  building their own or bringing in maybe an outside [1:23] consultancy to help them create bespoke AI-powered  apps for what they were doing. We saw a lot of [1:28] organizations then bringing in point solutions  that maybe were very important for one department. [1:34] Again, this proliferation of everybody finding  their own use of AI and some really powerful [1:39] solutions that added a lot of value. Uh from  there, employees bringing in their own tools, [1:44] again not acting out of malice, but these are  employees that probably want to move faster and [1:48] they're looking to harness the latest and greatest  technologies to be more effective in their roles. [1:53] Again, all of this just creating more management  and concerns for security and governance of where [1:59] is the data going? Uh at the same time, uh each  of your organizations undoubtedly has seen that [2:05] the cost is going up. We've heard a lot of talk  recently about tokenomics, something I'll focus on [2:09] today as well. How do you manage this increasing  spend? And that's quite a flip from where we were [2:14] a year ago, even 6 months ago, where so many  organizations I talked to were proud to show [2:18] me their leaderboards of which employees were  using the most AI. Accenture very publicly said [2:23] that you wouldn't get promoted if you didn't use  more AI. To now, we're hearing about companies [2:28] blowing through their budget and  running into all sorts of challenges. [2:32] Uh that co-pilot started 2 or 3 years ago,  well that's probably now expanded out across [2:36] multiple teams, departments, perhaps it's gone  company-wide. You've got new risks coming from [2:42] other governance processes that are strained to  keep up with with all of this change. Uh you're [2:47] perhaps there's an AI center of excellence  that reviews assessments every few weeks, [2:51] looks at what should be approved, etc. They're  really feeling inundated right now by all of this [2:56] and how quickly it's happening. At the same time,  that process was probably built for reviewing [3:00] software or other business processes that were  more deterministic. How do you evaluate the safety [3:06] and security and robustness of some of these  measures models are evolving and drifting? You've [3:11] got the agents that are learning and they're also  doing things that weren't anticipated. It's very [3:15] very difficult. So this process is probably  at a breaking point. Uh couple that with other [3:20] regulatory concerns and availability. We saw  the Fable 5 example just about a month ago [3:26] where at 5:00 p.m. if you had moved quickly  to deploy Fable 5, you got 90 minutes notice [3:31] that it was going to be taken offline and you were  going to have to quickly migrate back to another [3:35] model that you might have used before or find  something that would be compatible. Uh we've got [3:41] different deprecation concerns. Uh if you've built  a core business process on one of these models, [3:47] you may have seen this already that after going  through all the testing and validation and make [3:50] sure it was going to work the way you needed it  to, the vendors now come back and said, "Guess [3:54] what? You're about to go live. We've got the new  version. You should really look at that one." And [3:58] maybe it's not even the just that it's the latest  and greatest, they may be deprecating the old one, [4:03] so you have no choice but to move off of it in  a matter of weeks or even days in some cases. [4:09] We've seen a lot of other political risks as  countries are now saying that they want to use [4:13] sovereign AI models or they don't want an AI  model from a particular country being deployed. [4:19] This is yet another risk that you'll have to be  prepared for. How do you route between models [4:23] when you work as a global company? If you're a  large organization, getting ready for all those [4:27] jurisdictional demands is going to be an important  concern. Outages continue to happen. Last week we [4:32] saw a big one, I believe Anthropic was down  for about an hour and a half at least in our [4:35] experience. Uh we've seen other outages recently  from OpenAI and other companies. Again, as popular [4:41] as these services are, they're straining to  keep up with the demand, and that means that [4:44] you are going to face outages. Are you prepared  for that core system to go down right when your [4:49] employees are all logging into it? Do you have  a way to route to another model will be a bigger [4:53] and bigger question. And all this began with uh  the explosion of agents and just proliferation [4:59] of these systems getting larger and greater, new  risks and security concerns emerging all the time. [5:04] I'm also here this week at Black Hat down the  street, with the major security conference where [5:08] already today several big announcements  have come out with new security threats, [5:13] vulnerabilities that have been found that that  are circumventing a lot of security systems. [5:18] Lastly, we'll just point to the increasing risk  of regulations. More of these coming all the time. [5:24] Right now in the US here, we have five states with  different levels of AI regulation, and several [5:28] more that are proposing or working through  their committees to release new regulations. [5:33] That's in addition to the EU AI Act, the  Singapore framework, and many others around [5:37] the world that you and your organization may have  to contend with. The point of all of this is that [5:43] AI is moving so fast. The headlines that we see  every week across the world. These risks are not [5:49] academic, they are not potential, they are real.  And you and your programs will have to be prepared [5:56] to contend with them as this proliferation only  expands and accelerates across the organization. [6:02] I would sum this up by saying that your program  is a critical portfolio. Vendors, models, data [6:09] integration, regulations, all of this will have  to be managed in one sort of comprehensive way. [6:15] And the fact of the matter is wherever you are,  whatever work you're doing, your AI program today, [6:21] it's as simple as it'll ever be. It's only  going to get much more sophisticated and much [6:25] more complicated. And that is why the crux of my  talk today is around the need for a comprehensive [6:32] AI management tool. My company Area offers such  a a product and there are several others on [6:39] the market. Our approach though is a little  bit unique. We're going to start by doing a [6:43] full discovery. Identify where all of that AI is  across your organization. When I first meet with [6:49] an organization, they often will give me a list of  published apps that they know about and they tell [6:53] me that they're pretty locked down, that they're  confident that this is it. On average, we're going [6:57] to find somewhere between 4 to 600% more AI across  the organization and that is a very real stat. [7:05] This is again all of those different groups  of AI users. It's the employees shadow AI, [7:10] it's your existing vendors that have turned on  AI capabilities inside their products and they [7:14] didn't tell you about it. And then of course all  the AI native or foundation models or AI work that [7:20] you know about and you're probably paying for. The  point is all of this is growing and it's becoming [7:25] a very, very complicated picture to manage if  you can even see it. So we'll start with that [7:29] discovery and within a matter of days produce  a comprehensive mapping of all of the data, the [7:34] models, the vendors, the providers, the users and  give you insight into where it's all happening. [7:39] From there we usually move into security and  governance from locking down. Let's make sure that [7:44] we're not introducing new risks, that we've got  good protections and guardrails in place to manage [7:48] the way that our users are employing AI, so that  we can trust it, so that we can move faster and [7:54] faster, and that we've got good process. The most  critical phase for all of you in this room is the [7:59] fourth, which is uh really around the maintenance  and optimization of it. Let's make sure our AI use [8:05] is efficient. It's appropriate. It's robust and  resilient enough that if something goes wrong or [8:10] something changes, we're prepared to address those  changes in a way that is evergreen and systematic. [8:16] Our products are divided in such a  way that we can continue to do this [8:19] on an ongoing basis. This is not a one-time  engagement. We are not a consulting provider, [8:24] but rather as an independent governance layer  that helps you manage and maintain all of this, [8:29] regardless of which vendors or tools or  technologies, open source, private providers, [8:34] whatever you might be deploying, we're going  to give you one centralized control plane [8:38] to all of it and make sure that it continues to  function the way that your organization needs. [8:44] Our offerings are split into four pillars, and  the first is all around that discovery capability, [8:48] which is again looking across nine different  vectors to find AI that your other vendors [8:53] maybe aren't seeing or aren't telling you about  or you're just not aware of. We have the most [8:57] comprehensive set of capabilities for this, I  believe, in the market today to identify AI use [9:02] anywhere across your organization, sanctioned  or unsanctioned. From there, we'll prepare [9:06] a full inventory of all of it and give you an  understanding of what your regulatory exposure is, [9:11] based on the way that the AI is being used. We'll  help you judge where you may have concerns that [9:16] need to be remediated and where you're doing  well, so that you can invest in the right areas. [9:21] The next is that security phase again for your  security teams, where we'll give you all of the [9:25] controls to constrain agent behavior, make sure  it's being used in the ways that it was expected, [9:29] and that it can't change its permissions or  evolve in ways that might introduce risk. [9:33] We employ safety guardrails, as well as a set of  gateways across all your AI, your integrations, [9:38] your MCP servers, so that you can manage and again  make sure that they're working the way that you [9:42] intended them to and not introducing other risk.  We also have a red teaming capability here, which [9:48] is adversarial testing. This is a set of agents  that we're going to deploy, it's a swarm of them, [9:53] that will run across your systems and help  you identify where there might be concerns or [9:57] something that might be introduced a vulnerability  that an a bad actor could manipulate. Very, [10:02] very important, we do this on an automated way  and then bring those findings back to you so that [10:06] you can get ahead of them before someone exploits  it. And then lastly, the full oversight. So, being [10:12] able to make sure that we have human oversight  where it's needed at the right steps that someone [10:16] can approve something when we see a real risk  that needs that additional level of attention. [10:21] That buckles right into our governance  components here, which again is [10:24] all of the continuous reporting for compliance  purposes. If you're a regulated organization, [10:29] we're going to produce all the artifacts, the  reporting to validate that you're doing the [10:32] right things and that if you had to evidence that  to a regulator anywhere in the world, you can show [10:37] how your program is working, you've got good  disclosures and transparency around all of it. [10:42] We also have our own concepts here  of audit trail and observability [10:46] to give you that full registry of all of the  governance practices across the organization. [10:50] Fourth and final pillar of our offering here  is all around that optimization I mentioned, [10:55] giving you the ability to really ensure that your  AI systems are resilient, that if something goes [11:01] down, if one of your providers is offline, that  we can immediately run Oops, if uh if they if [11:07] it's a matter of financial concerns, being able  to route to the right model for the right use [11:12] case. So often your employees are going to want  to reach for that big 1 trillion parameter model [11:17] when they need a simple task like building a  slide deck or something like along those lines. [11:21] We'll help them get to the right model that is the  perfect fit for their needs, but also the right [11:26] fit from a cost perspective that we can manage all  of that in addition to providing the chargeback, [11:31] full tracing so that we know which teams are  running hot. If someone's over budget, we can [11:35] stop them before that happens or simply notify  them. A lot of controls around that. And that's [11:40] in addition to all the other integration and  management capabilities around ensuring that your [11:46] AI deployment continues to evolve  the right way, the change management. [11:50] I'll speak to that just very briefly because I  think change management is a really big concern [11:54] that a lot of organizations I've spoken to that  this week are missing. Are you prepared to shift [11:59] between models and vendors? If you're all in  on one vendor, that's great. It might add some [12:03] simplicity right now, but you're offering up a lot  of leverage to one particular vendor in that case. [12:10] Time to renegotiate, they're going to exercise  that leverage. So, being able to embrace a [12:14] lot of different models and even within the  particular family as they release new models, [12:19] are you prepared to move through those changes?  They're going to continue to change and your [12:23] agents will as well. We have new models that  are coming out, your prompts are changing, [12:27] the data integrations. Each of these changes  is going to have implications for the way that [12:32] your systems are working. Being able to test  that and validate it before it happens in an [12:37] automated autonomous way will become a critical  capability as your deployment accelerates. [12:43] We'll also look at this from uh different  dimensions. Uh I've mentioned token spend and [12:48] how cost is changing. How is it going to interact  and what additional costs would we incur when we [12:54] move to that new model? How will it play with our  tools? What's going to change around the behavior, [12:59] the temperature, the accuracy of something? Do  you today have a good way of testing all of this [13:04] in a quick, efficient way? That is exactly what  we offer here. Being able to run multiple what-if [13:10] scenarios to see how different models would  perform for your agents for different use cases. [13:15] And we do this again in a highly efficient,  automated way that will allow your systems to move [13:21] and embrace all the innovation as it comes out.  We'll analyze each of those potential changes [13:27] without impacting any of your production systems.  All of this happens in the background so that [13:32] you are continuously being ready for that next  evolution of a model, of a change in your systems, [13:37] or the agents. We can have it all queued up, and  in the event something does go offline, if there's [13:42] an outage, or if we understand that maybe there's  a security concern with a particular model, [13:47] we find that there's a backdoor that a government  might be exploiting, we can immediately shift [13:52] any of those agents, any of those workloads over  to another model, and you can do it confidently [13:56] because you know it's been tested, and you've  got all of the analysis to back up the validation [14:00] that already happened. We meet sometimes with  several large organizations that have told us [14:04] that this process can take weeks or months. And  again, the worst-case scenario for them is that [14:09] when they get through all of that testing to make  sure it's all going to work the way it needs to, [14:13] they find out that the vendor is releasing a new  model, and that the old one is being deprecated [14:17] and pushed offline. That's a very scary position,  and that's one that I think is going to become [14:21] more common for a lot of organizations if they  haven't prepared some sort of management program. [14:27] Now, we do this as part of our governed governance  offering as well, so that you can see that we step [14:32] through all the right assessments to analyze the  risk, the validation, the red teaming I mentioned, [14:38] and ensure that you're really ready to  move to that new model as it comes out. [14:41] It's a five-step process going from evaluation  to red teaming to approving and then manage [14:45] monitoring all the shadow traffic and any risk  that might come with it, all as part of this [14:49] change management protection program that we offer  in our tool. That buckles right into the financial [14:56] management that I mentioned before of being able  to monitor your token consumption. We can do this [15:00] today at any level that might be appropriate for  your organization. Being able to look at the the [15:05] project, the team, the business unit, even in an  individual level, set budgets and thresholds for [15:12] certain actions. I might suggest that at 70% of  my budget, I'm going to get a notification warning [15:18] me that it's running too fast or that we're in  danger of hitting those limits. Perhaps 80% my [15:23] manager gets a notification alerting them to the  situation. And then maybe as a ultimate escalation [15:29] at 90% perhaps we automatically route this user or  this agent or this project to a different model. [15:36] So maybe I've been enjoying using the latest  and greatest most expensive model, but at 90% [15:42] my company is going to route this project over  to a different model that is free or very very [15:47] low cost. Uh we will have already identified uh  alternative models for a given project or workflow [15:53] so that we can say using this other model is going  to be just as accurate, just as performative, but [15:58] perhaps the latency is a little higher or this is  going to be a little slower for me to use or maybe [16:02] it has to be done overnight in batches. The point  is we build in these financial protections to [16:07] ensure that you don't find yourself in the CFO's  office explaining why overnight we hit a $40,000 [16:13] expenditure that nobody budgeted for. Another real  example of this, uh we had a client this weekend [16:20] uh Saturday morning who at about 3:00 a.m. uh  had a key hijacked. Uh they're still working [16:26] to determine who stole it, but the fact is $40,000  was disappeared in no time at all. Uh whoever took [16:33] the key ran an enormous amount of work through  their Anthropic uh license uh through their [16:38] Anthropic key and uh that customer is now trying  to deal with the fallout of the situation. With [16:44] these sort of protections in place, not only could  we notify, we could shut it down or as I mentioned [16:49] we could route to other models to ensure that you  don't have a big bill blindside you just that way. [16:55] Now, I've said a lot over the last 18 minutes or  so uh just a quick introduction to my firm. As [17:01] mentioned, Area is a AI management platform.  We are a global company. We are only about [17:07] 22 months out of stealth, so relatively  new if you hadn't heard of us before, [17:10] I I certainly wouldn't hold it against you.  Uh but already today with over 500 customers, [17:15] close to 200 employees around the world,  and we are growing very, very quickly. [17:20] If any of you are Gartner clients, you may  have seen the recent Magic Quadrant. They just [17:24] launched an AI governance Magic Quadrant.  It's the first time they've ever done one. [17:28] And we were very proud to to see it because again,  it validates the space that there is this need [17:33] in the market that they're hearing enough analyst  inquiry on. Well, our company was not only invited [17:38] to participate, we made the down selection, and  we were judged as having the most comprehensive [17:43] platform of any provider. Uh if you're interested  in the research, we've licensed it and I can give [17:48] you a copy of it afterwards today. Um but uh a  good good selling point again as you think about [17:53] your own programs, this this research really steps  through some of the key concerns. We are today uh [18:00] working on a very comprehensive platform as you've  seen, and I think that comes from our background. [18:04] I worked with a lot of your organizations,  myself and my business partner uh I've built [18:08] two other companies before this, AirWatch and  OneTrust. If you were ever a client of either, [18:12] uh we were part of building both of those, and uh  again, deep background here in both cybersecurity, [18:17] but also compliance and infrastructure management.  So, all themes that are very relevant as you [18:22] consider your your AI program. I mentioned that  Gartner research, there's a whole lot of other [18:27] resources that I would point you to. If any of you  are subscribers, you can find all these or again, [18:32] we can help you get you copies of them as you  build out your own programs. We would love to [18:37] help in any way we can with either free research  or a number of free resources that we provide. [18:41] I mentioned our discovery engagements to help you  identify where AI use is across your organization. [18:48] Lastly, uh I've seen a lot of phones going  up today, and if the content was helpful, [18:52] I'm happy to share my slides with you. You can  scan the QR code on the screen or come see me [18:57] afterwards, and I can email them to you. We'll  also be on the show floor downstairs. If you have [19:02] time to stop by, we'll be doing demonstrations,  talking about some of the specific capabilities, [19:06] and would love to discuss your program and  how we might be able to support as part of [19:10] your go-forward plan to keep your AI program  running smoothly. I want to thank you very [19:15] much for the time today. I hope some of this  was informative and again would welcome any [19:18] questions or discussions if anyone wants to  talk afterwards today. Thank you very much.