Norstella Chief Says AI Pilots Fail In Life Sciences Without Data Links

Norstella负责人：缺乏数据关联导致生命科学AI试点失败 未关联的企业数据和未经核验的AI原型在重大制药决策中造成了根本性的信任鸿沟，从而推高了药物研发成本并延长了研发周期。 Kris Kaneta探讨了生命科学领域的“AI生产力悖论”，即尽管投入了巨额AI资金，药物研发的周期、成本与失败率却不降反升。他将这一现象归咎于无法规模化、缺乏信任的AI原型所导致的“试点困境”，并提出了四项“明确性准则”（Clarity Criteria：有依据/Cited、一致/Consistent、具语境/Contextual、具实质影响/Consequential），以在关键制药决策中建立信任。为说明解决方案，他重点介绍了Norstella的Atlas平台及其竞争情报智能体，展示了基于统一纵向数据并经领域训练的AI如何大幅超越通用LLM。最后，他敦促行业领导者专注于加快决策速度，而非为了做AI而做AI，从而推动疗法更快惠及患者。 Norstella 为制药领域的高风险决策提供核心数据基础，其情报为市场上超过 80%（五分之四）的救命疗法提供了决策支持。 (0:04) 生命科学行业正面临“AI 生产力悖论”：尽管 AI 投资大幅增加，但在过去五年中，药物研发成本翻倍、研发周期延长超过 10%，且监管审批成功率持续下降。 (2:19) 制药决策动辄耗资数亿至数十亿美元且无法撤销，大量 AI 原型因缺乏扩展性、情境理解以及与工作流的整合，陷入了难以规模化的“试点困境”。 (4:03) Kris Kaneta 提出了“清晰度标准”（Cited 可溯源、Consistent 一致性、Contextual 情境化、Consequential 具决策影响力），旨在将 AI 演示转化为值得信赖的决策工具。 (8:05) Norstella 推出智能体决策平台 Atlas，通过纵向打通管线、临床试验、支付方及真实世界数据等孤岛，帮助专业分析团队实现从“AI 演示”向真正业务决策的跨越。 (12:13)

Kris Kaneta, Chief Product and Innovation Officer at Norstella, argues that despite massive AI investment, pharmaceutical productivity has deteriorated over the past five years. He contends that high-stakes, multi-billion-dollar drug development decisions cannot rely on unverified AI prototypes, requiring instead decision-grade infrastructure built on unified, longitudinal data. The Life Sciences Productivity Paradox (2:19) According to Kaneta, life sciences is experiencing a stark contradiction between artificial intelligence investment and operational output. Despite widespread excitement and financial backing aimed at making pharmaceutical teams faster and better informed, the industry's broader productivity metrics have trended downward over the last five years. Drug development costs have doubled, overall development timelines have increased by more than 10%, and the probability of technical and regulatory success has systematically dropped. Kaneta notes that this downturn occurred in parallel with the generative AI boom, demonstrating that adopting AI tools has not automatically translated into industry-wide efficiency or faster therapeutic development. The Scaling Failure of Experimental Pilots (4:03) Pharmaceutical organizations typically fall into two camps: those targeting specific AI outcomes and those exploring open-ended possibilities. Kaneta observes that both groups suffer from a pervasive "pilot problem," characterized by an influx of prototypes that fail to scale technically, lack workflow context, or fail to integrate into core operations. In drug development, decisions carry massive downstream momentum—such as allocating $3 to $5 billion to a single drug candidate, selecting clinical investigators, or setting trial protocols. Because there is no "undo" function for irreversible, multi-billion-dollar pharmaceutical commitments, unvetted outputs and unscalable prototypes fail to earn the trust required for executive action. The Four Clarity Criteria for Trust (8:05) To evaluate whether an AI deployment can transition from an impressive demonstration into a reliable enterprise tool, Kaneta outlines a four-part rubric called the "Clarity Criteria." First, outputs must be Cited, tracing directly to a verified source of truth. Second, systems must be Consistent, returning the same verified truth across repeated inquiries. Third, tools must be Contextual, understanding specific industry workflows and the surrounding data ecosystem. Finally, results must be Consequential, reliably prompting the correct real-world action. Kaneta compares deploying enterprise AI to managing 1,000 human interns: organizations must establish explicit goals, supply authoritative foundational datasets, and demand structured outputs. He notes from Norstella's own development history that simple data lookups are insufficient; systems require data linked historically and longitudinally across time to maintain consistency. Unifying Enterprise Silos Beyond Generic LLMs (12:13) A critical barrier in pharmaceutical operations is functional fragmentation, where business development, clinical trial planning, and market access teams keep essential information isolated from one another. Kaneta argues that generic large language models relying on public domain repositories risk amplifying an AI echo chamber, frequently omitting approved therapies or misidentifying discontinued drugs in market landscape analyses. To overcome these limitations, Norstella introduced Atlas, an infrastructure platform that consolidates drug pipelines, clinical trials, investigator networks, payer coverage, and real-world evidence into a single architecture. Kaneta asserts that enterprise decision-making requires specialized agents trained directly by domain analysts who perform specific industry tasks, rather than generic models that lack operational depth. Kaneta concludes that pharmaceutical leaders must prioritize infrastructure that moves organizations from a "demo to a decision" rather than adopting AI for its own sake, emphasizing that resolving the productivity paradox is necessary to accelerate the delivery of life-changing therapies to waiting patients.

Norstella 首席产品与创新官 Kris Kaneta 指出，生物制药领域对人工智能的巨大投入并未带来研发效率的同步增长，行业反而陷入成本翻倍与研发周期拉长的“生产力悖论”。他主张企业必须走出无法扩展的算法原型困境，以真实业务决策为核心，构建具备溯源性、一致性与业务上下文的专业智能决策基础设施。 药企AI投资激增与研发效能恶化的矛盾 (2:19) 在生命科学领域，AI 技术的兴起伴随着巨额的资金投入与高度的行业期待，其最初承诺是帮助团队提高研发速度、认知深度与决策生产力。然而 Kaneta 指出，过去五年间整个制药行业的生产力关键指标却在持续反向恶化：开发一款新药的资金成本增加了一倍，研发周期拉长了超过10%，并且技术与监管审批的成功率也出现了系统性下降。AI 繁荣与实际产出之间的脱节，构成了当前行业严峻的生产力悖论。 根据 Norstella 的数据，该公司旗下汇聚了 Citeline、MMIT、Evaluate、Panalgo 以及 The Difference Group 等业务品牌，为全球超过3,000家客户提供数据与智能支持，其中包括全球前100家制药企业的全部成员；每年上市的开创性疗法中有超过80%受到其数据决策的指导。Kaneta 强调，行业面临的问题不在于数据和模型算法的缺乏，而在于现有的 AI 部署方式未能切实扭转研发效率下滑的客观趋势。 高风险决策环境下的原型困境与信任缺失 (4:03) Kaneta 分析认为，药企在应用 AI 时通常分为目标明确的推进者与探索技术可能性的观察者，但两类组织均深陷同质化的“原型困境”（Pilot Problem）。市场上充斥着大量被称为“AI 残渣”（AI slop）的算法原型与概念验证（POC），它们大多停留在演示阶段，因缺乏技术可扩展性、缺乏对具体问题业务场景的理解，以及未能融入实际决策流程而无法在企业内部全面铺开。 这种落地困难的根源在于制药决策极高的风险属性与对工具的信任缺失。新药研发涉及耗资30亿至50亿美元的单一分子管线推进、适应症拓展、临床试验方案设计及中心研究者筛选等重大决断。Kaneta 强调，制药决策具有连锁反应，一旦失误便没有“撤销键”（Ctrl+Z）可供挽回，这使得决策团队在面对缺乏信任保障与可解释性的独立 AI 原型时普遍保持审慎态度。 驱动业务转化的清晰度四原则与管理隐喻 (8:05) 基于 Norstella 在评估和部署智能体过程中的实践教训，Kaneta 提出了用于检验 AI 是否具备企业级决策能力的“清晰度四原则”（Clarity Criteria）：第一，可溯源（Cited），模型输出必须能够追溯至受信任的真实事实源；第二，一致性（Consistent），多次重复同一查询必须返回稳定且一致的经核实结果；第三，情境化（Contextual），系统需深入理解岗位工作流与企业专有数据生态；第四，实质影响（Consequential），输出结果必须能够切实指导并触发正确的现实行动。 Kaneta 借用“管理1,000名人类实习生”的比喻向行业建言：企业不应单纯为了追求 AI 技术而构建系统，而应为了达成业务决策而设计架构。面对实习生团队，管理者必须明确工作目标、提供权威基础数据并制定清晰的交付结构；对待 AI 亦是同理。他坦承 Norstella 早期研发也曾遭遇缺乏可溯源性与时间纵向关联的弯路，证明仅靠单点抓取孤立数据无法建立稳定的决策信任，必须在时间序列上对异构数据进行深层纵向串联。 消除数据孤岛与垂直领域智能体的优势 (12:13) 在传统药企业务中，部门间的职能孤岛严重阻碍了数据流动。例如，业务拓展（BD）团队掌握的信息往往不与市场准入团队同步，市场准入洞察也与管线预测人员脱节。为此，Norstella 推出了名为 Atlas 的智能体决策基础设施，将研发管线、临床试验、研究者网络、支付方覆盖政策、市场准入以及真实世界数据（RWD）等全链条语料打通，使专业 AI 智能体能够在统一的数据基底上跨职能协同。 以“评估多动症（ADHD）市场格局”为例，Kaneta 对比了通用大语言模型与经过垂直训练的竞争情报智能体（Atlas CI）的差异。依赖公共网络（如 PubMed 或 ClinicalTrials.gov）的通用模型容易受限于“AI 回音室”，不仅遗漏现有在售药物，还会错误计入已退市药物，且缺乏分析师业务常识；而由资深行业分析师联合训练的垂直智能体，能够精准识别药物全生命周期状态，直接输出契合分析师决策标准的专业结构化结果。 Kaneta 总结认为，药企推进人工智能战略的核心在于实现从“技术演示”向“实质决策”的跨越。只有通过满足可溯源、一致性、情境化与实质影响的评价标准，打破数据壁垒并构建由行业专家训练的专业智能基础设施，才能真正消除生产力悖论，加速研发决策落地，最终让创新药物更快惠及等待中的患者。

## Background 本内容由 YouTube 频道 Ai4 于 2026 年 8 月 19 日发布，主讲人为生命科学数据与情报机构 Norstella 的首席产品兼创新官 Kris Kaneta。视频介绍指出，针对新药研发周期冗长的问题，演讲探讨了阻碍行业提速的根本原因，并阐述了 Norstella 如何通过 AI 连接研发管线到患者的全流程数据，推动制药团队从问题高效迈向决策与行动。 ## TL;DR 在 AI 投资空前繁荣的背景下，制药行业却反常地遭遇了成本翻倍、周期拉长以及获批率下滑的“AI 生产力悖论”。演讲指出，由于多数 AI 项目脱离决策场景且存在信任危机，全行业普遍深陷无法规模化落地的“试点困境”。为此，Norstella 提出了以“可溯源（Cited）、一致性（Consistent）、具上下文（Contextual）、具影响力（Consequential）”为核心的清晰度准则，并推出贯通研发至上市全生命周期的 Atlas 决策基础设施。通过对比通用 LLM 与垂直专家 Agent 在商业竞争情报分析上的表现，Kris 强调企业必须打破部门数据孤岛，始终以“赋能高风险决策”为导向来构建 AI 基础设施，而非盲目为了 AI 技术本身而建设。 ## Atmosphere & style Kris Kaneta 的演讲风格幽默沉稳且直击行业痛点，开场以轻快的调侃迅速拉近与台下观众的距离。随着话题切入行业尴尬反常的“生产力悖论”，现场氛围转为严肃务实，充满对当下浮躁 Demo 与 AI 泡沫的冷峻反思。其将 AI 比喻为“一千名实习生”的形象比喻深入浅出，实测对比通用模型缺陷时更是犀利透彻；整场分享兼具商业战略的严密说服力与行业老兵的真诚呼吁，给听众带来强烈的情感共鸣与行动紧迫感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=M5K78HGlmvE&t=4s) Norstella 旗下整合了 Citeline、MMIT、Evaluate 等核心医药品牌，为全球排名前 100 的全部制药巨头提供从管线研发到患者准入的数据决策支持。 - [00:04](https://www.youtube.com/watch?v=M5K78HGlmvE&t=4s) 市场上每年有超过 80% 的创新疗法在研发或决策中使用了 Norstella 的数据，但制药行业正面临严重的“AI 生产力悖论”。 - [02:34](https://www.youtube.com/watch?v=M5K78HGlmvE&t=154s) 过去 5 年间，尽管 AI 投资暴增，新药研发成本却翻倍、开发周期延长逾 10%，且技术与监管获批率（PTRS）持续走低。 - [04:14](https://www.youtube.com/watch?v=M5K78HGlmvE&t=254s) 制药行业普遍陷入“试点困境”（Pilot Problem），原型验证和炫酷 Demo 泛滥，但因缺乏扩展性与业务上下文而难以规模化落地。 - [04:14](https://www.youtube.com/watch?v=M5K78HGlmvE&t=254s) 医药研发涉及动辄数十亿美元且无法撤回的高风险决策，缺乏可信度的 AI 输出导致业务团队面临严重的“信任危机”。 - [07:34](https://www.youtube.com/watch?v=M5K78HGlmvE&t=454s) Norstella 提出评估 AI 应用落地能力的“清晰度准则”（Clarity Criteria），包含可溯源、一致性、具上下文和具影响力四个维度。 - [10:02](https://www.youtube.com/watch?v=M5K78HGlmvE&t=602s) 最优秀的 AI 基础设施从来不是为了 AI 本身而建，一切技术构建必须以促成精准的业务决策为唯一导向。 - [10:02](https://www.youtube.com/watch?v=M5K78HGlmvE&t=602s) 管理 AI 应如同管理“一千名实习生”，必须清晰界定任务目标、权威事实来源以及优秀成果的交付标准。 - [12:42](https://www.youtube.com/watch?v=M5K78HGlmvE&t=762s) 针对部门间严重的数据孤岛，Norstella 推出决策基础设施 Atlas，跨越药物生命周期数据构建专业角色 Agent。 - [14:04](https://www.youtube.com/watch?v=M5K78HGlmvE&t=844s) 通用大模型因检索公共互联网容易陷入“AI 回音室效应”，常产生遗漏在售药或误算退市药等致命商业分析错误。 - [14:04](https://www.youtube.com/watch?v=M5K78HGlmvE&t=844s) Atlas 专用 Agent 依托纵向打通的全生命周期医药专有数据，并由一线资深分析专家调优，具备深度的商业情报逻辑与理解力。 - [15:53](https://www.youtube.com/watch?v=M5K78HGlmvE&t=953s) AI 实现从实验 Demo 跨越到真实业务决策的核心关键，在于构建服务于“决策”的基础设施，而非盲目堆砌 AI 模型。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=M5K78HGlmvE&t=4s) Norstella 的定位与“AI 生产力悖论” 演讲者 Kris Kaneta（Norstella 首席产品兼创新官）以轻松幽默的开场向现场观众致意。他调侃大家足够明智，早早排队领到了 Ai4 大会的胸卡，因为场外排起的长队看起来“简直疯狂”。随后，他借着核对演讲幻灯片的契机切入正题，为尚不熟悉 Norstella 的听众梳理了公司的业务轮廓。对于生命科学领域的从业者来说，Norstella 旗下的多个知名品牌早已深入日常工作，包括 Citeline、MMIT、Evaluate、Panalgo 以及 The Dedham Group 等。Norstella 的核心定位是为制药行业最重大、赌注最高的决策提供数据与智能支撑，业务覆盖从早期研发管线、投资组合策略、临床试验规划，到市场准入及商业化上市的完整链条（“从管线到患者”）。 Kris 强调，Norstella 拥有极高的行业渗透率和广泛的视野：在全球服务超过 3000 家客户，囊括了全球排名前 100 的全部制药巨头；在任何给定的年份里，市场上超过 80%（五分之四以上）改变患者生命的创新疗法，都在不同程度上受到 Norstella 数据与情报的塑造、启迪与指引。在技术实现上，这些智能成果转化为分析型产品、预测与处方模型、推荐引擎，并通过 API、MCP（模型上下文协议）、小语言模型（SLM）、大语言模型（LLM）以及 AI Agent 等多种形态落地。在此背景下，他抛出了本次分享的核心议题：在全行业即将被充斥整周的“AI”两个字母搞得审美疲劳之前，必须客观审视生命科学领域正在上演的“AI 生产力悖论”（AI Productivity Paradox）。 - **覆盖全球制药头部**：服务全球排名前 100 的全部制药企业，客户总量超过 3000 家。 - **极高的药物渗透率**：每年上市的新药中，80% 以上在决策与研发流程中应用了 Norstella 的数据与智能。 - **技术交付形态**：涵盖预测/处方推荐引擎、API、MCP、SLM、LLM 以及自主 AI Agent。 --- ### [02:34](https://www.youtube.com/watch?v=M5K78HGlmvE&t=154s) 医药研发的反常成绩单：投入加倍、周期变长、成功率下滑 尽管过去几年整个制药与生命科学领域对人工智能抱有无与伦比的热情，资金投入与技术狂热达到顶峰，但制药行业的实际成绩单却极其反常地“滑向了错误的方向”。作为领导者，每个人都在大力押注 AI，指望它能革新工作模式、让团队变得更敏捷、更聪明、更高效、更具洞察力。然而残酷的现实表明，在这场轰轰烈烈的 AI 复兴浪潮中，制药行业反而变得更缺乏生产力了。 ```mermaid flowchart LR accTitle: 制药行业在 AI 时代的生产力悖论 subgraph S1["巨大投入与技术期望"] A["资本狂热与 AI 大规模投资"] --> B["预期：加速研发 / 提升成功率"] end subgraph S2["反向演进的现实困局"] B -.-> C["研发资金成本翻倍"] B -.-> D["研发周期延长逾 10%"] B -.-> E["技术与监管获批率下滑"] end ``` Kris 用不容置疑的数据和沉痛的语气重复了这一行业现状：新药研发所需的投资直接翻倍；过去 5 年间，新药开发周期不降反升，延长了超过 10%；更糟糕的是，技术与监管获批概率（PTRS，Probability of Technical and Regulatory Success）反而持续走低。他向全场发问：“花得更多、耗时更长、成功率更低——在这个号称被 AI 彻底重塑的时代，我们究竟该如何解释这种巨大的断层？” - **研发成本激增**：开发一款新药所需的整体资金投入较以往翻了一番。 - **研发周期拉长**：过去 5 年内，药物研发整体耗时增加了 10% 以上。 - **成功概率下滑**：药物迈向最终监管审批的获批率指标不升反降。 --- ### [04:14](https://www.youtube.com/watch?v=M5K78HGlmvE&t=254s) 困扰全行业的“试点困境”与信任危机 审视生命科学领域各家机构在 AI 探索中的表现，大致可以划分为两个阵营：第一类阵营采取“一头扎入”的激进姿态，目标极其明确，有清晰的量化产出预期，清楚界定了 AI 投资在理想状态下的具体面貌；第二类阵营则坦承尚未看清全部技术影响和具体应用路径，但深知其中蕴藏着巨大机遇与“可能性的艺术”，致力于不断试探和摸索。然而，Kris 指出，无论是哪一个阵营，他们目前所撞上的核心痛点惊人一致——整个制药行业陷入了严重的“试点问题”（Pilot Problem）。 各种缺乏实际落地能力的“AI 垃圾”（AI slop）、原型设计、概念验证（POC）和试点项目正在组织内部泛滥成灾。这些展示往往听起来诱人、看起来极具说服力，但最终无一例外无法规模化。这种无法规模化并非单纯的技术瓶颈，而是源于三个维度的脱节：首先是技术本身的扩展性不足；其次是系统对所要解决的具体医药业务场景缺乏上下文理解；最后也是最致命的一点，它完全脱离了制药人做出实际决策的工作流程。这就直接导致了深层的“信任危机”。在制药领域，从业者承担的都是动辄数十亿美元、押上身家性命的高风险决策，比如是否为某个分子投入 30 亿至 50 亿美元、是否拓展全新的适应症、临床试验方案如何设计、选定的研究中心研究者是否能满足入组指标等。这些决策没有“撤销键”（No Control-Z），一旦失误便无可挽回，并且会对后续的系列行动产生巨大的连锁反应。当无法百分之百信任 AI 输出的内容时，绚丽的 Demo 就永远无法转化为实际的商业决策。 --- ### [07:34](https://www.youtube.com/watch?v=M5K78HGlmvE&t=454s) 建立信任的“清晰度准则”（Clarity Criteria） 在过去的十余年里，Norstella 自身也在经历同样的演进历程，从自然语言处理（NLP）、机器学习，一路探索到 SLM、LLM 以及部署在内部系统与商业产品中的数十个自主智能体。在不断摸索“如何从 Demo 跨越到真实决策”的过程中，Norstella 总结出了一套评估 AI 应用的核心评估框架——“清晰度准则”（Clarity Criteria）。Kris 呼吁台下的企业管理者们，在下周一回到办公室、面对各种天花乱坠的 Demo、插件和酷炫模型前，先用以下四个基准进行灵魂拷问： ```mermaid flowchart TD accTitle: 评估 AI 决策信任度的清晰度准则 subgraph Criteria["清晰度准则 (Clarity Criteria)"] C1["1. 可溯源 (Cited)<br/>输出是否能锚定到可信的真实事实源？"] C2["2. 一致性 (Consistent)<br/>重复相同任务 5 次，是否始终给出稳定结果？"] C3["3. 具上下文 (Contextual)<br/>是否深度契合业务目标并贯通纵向数据生态？"] C4["4. 具影响力 (Consequential)<br/>所驱动的决策与任务能否引发正确的业务行动？"] end C1 --> Trust["构建商业决策层面的真实信任"] C2 --> Trust C3 --> Trust C4 --> Trust ``` 1. **可溯源（Cited）**：AI 生成的每一个结论与数据点，是否都能明确回溯并引用到业务人员认可的底层“单一事实来源”（Source of Truth）？ 2. **一致性（Consistent）**：如果将同一个问题或推理任务重复执行 5 次，系统给出的结论和引用的事实源是否具有高确定性与一致性，从而让人产生信赖感？ 3. **具上下文（Contextual）**：系统是否真正理解使用者试图达成的业务目标？它是否接入并理解了支撑整个组织的复杂数据、情报及洞察生态系统？ 4. **具影响力（Consequential）**：基于该输出所促成的决策或触发的下游任务，是否真正具备业务价值，能否有效引导团队采取正确的行动并重塑组织行为？ --- ### [10:02](https://www.youtube.com/watch?v=M5K78HGlmvE&t=602s) 勿为 AI 而建：把 AI 当作一千名实习生，以决策为导向 Kris 向行业发出了恳切的呼吁：“最优秀的 AI 基础设施，从来都不是为 AI 而建的。不要为了 AI 而建 AI，一切建设必须以‘促成决策’为唯一导向。”他给在座的管理人员提供了一个极其形象的思维模型：假设下周上班时，身边没有 Claude，没有 Gemini，也没有 ChatGPT，取而代之的是“一千名实习生”突然被交到你手中。 面对一千名渴望产出却缺乏行业背景的实习生，管理者必须做到三件事：第一，必须极其明确地告知他们任务的具体目标；第二，必须清晰界定基础事实与知识来源，因为刚入职的实习生自身并不具备专业上下文；第三，必须极其清楚地定义“优秀的结果长什么样”，包括产出的结构形式、将要支撑哪些下游决策，以及将触发哪些具体的商业行动。AI 模型在本质上就是这一千名实习生，如果脱离了严密的决策支持框架，它产出的东西就毫无意义。 Kris 坦言，Norstella 在构建自身技术架构的过程中也经历过深刻反思与“吃自己的狗粮”（Eat our own dog food）的阵痛。早期发布的部分模型同样缺乏精准的可引证性，上下文各层之间彼此割裂。仅仅看到事实源是一回事，但要将历史维度与纵向时间轴上的多源事实严密串联、持续给出稳定可靠的答案，则是完全不同的工程挑战。正因如此，Norstella 研发推出了决策基础设施 Atlas，旨在从底层数据与决策链路入手，彻底击碎 AI 原型与临床研发决策之间的信任鸿沟。 ### [12:42](https://www.youtube.com/watch?v=M5K78HGlmvE&t=762s) 破除数据孤岛：Atlas 决策基础设施与专业 Agent 要让 AI 在制药行业真正发挥价值，底层情报必须在历史与时间纵向上深度关联。Kris Kaneta 打了个比方：如果要训练一千名实习生，或者打造一个顶级的 AI Agent 来完成一项专业任务，它必须能够跨越所有这些信息维度自由穿梭。Norstella 推出的 Atlas 正是这样一个智能体平台（Agentic Platform）。它构建在覆盖整个药物生命周期的完整数据语料库之上——从前期的药物研发管线（Drug Pipeline）、临床试验（Trial）、研究者（Investigator），到后期的支付方（Payer）、医保报销准入（Coverage and Market Access），直至真实世界数据（RWD）。基于这套全链路数据，系统能够灵活派生出多种角色的智能体，例如自豪地宣告：“你好，我是一名竞争情报（Competitive Intelligence）分析师”、“你好，我是一名可行性（Feasibility）分析师”，或者“你好，我是一名市场准入与洞察（Market Access and Insights）分析师”。 现实工作中的核心症结在于严重的信息孤岛（Data Silos）。无论身处哪家制药企业，数据往往都被牢牢困在各个部门内部无法流通：业务拓展（BD）团队手握数据却不愿向市场准入团队公开，市场准入团队积累的资料又对预测（Forecasting）团队秘而不宣。Kris 强调，企业必须打破这种部门壁垒，建立正确的“决策基础设施（Decision Infrastructure）”。这正是开发 Atlas 的初衷——在统一的底层数据上，构建高度专业、深谙业务逻辑的专用 Agent，直击日常工作中的真实业务痛点。 ```mermaid flowchart TD accTitle: Norstella Atlas 架构示意图 accDescription: 展示从底层贯通数据到构建专业角色 Agent 的决策基础设施架构 subgraph DataCorpus ["Norstella 纵向贯通数据语料库"] D1["药物管线 (Pipeline)"] D2["临床试验 (Clinical Trials)"] D3["支付方与准入 (Payer & Access)"] D4["真实世界数据 (RWD)"] end subgraph Platform ["Atlas 决策基础设施平台"] P1["Agentic 平台核心"] end subgraph SpecializedAgents ["垂直角色智能体 (Specialized Agents)"] A1["竞争情报分析师 (CI Analyst)"] A2["试验可行性分析师 (Feasibility Analyst)"] A3["市场准入与洞察分析师 (Market Access Analyst)"] end DataCorpus --> Platform Platform --> A1 Platform --> A2 Platform --> A3 ``` - **覆盖全生命周期的数据广度**：整合管线、试验、研究者、支付方、医保目录及真实世界数据。 - **部门壁垒痛点**：BD、市场准入、财务与销量预测部门之间的数据割裂，直接阻碍了关键决策的制定。 - **破局理念**：超越工具层面的修修补补，打造一套承载多角色 Agent 的统一决策基础设施。 --- ### [14:04](https://www.youtube.com/watch?v=M5K78HGlmvE&t=844s) 通用大模型与垂直专家 Agent 的实战对比 为了展现通用模型与专业 Agent 的本质差距，Kris 现场对比了一组测试用例。左侧是调用通用 LLM 并连接公共检索源（如 PubMed 和 ClinicalTrials.gov）的输出结果，查询指令是“请提供 ADHD（注意缺陷与多动障碍）的市场格局全景”。团队进行了一连串严格的测试，每一次通用 LLM 给出的回答都非常抓人眼球、听起来极具说服力。然而，只要由资深从业人员深入审视，致命缺陷就会暴露无遗：通用模型扫描的是公共互联网领域的内容，其中充斥着大量由 AI 自动生成的内容。这种“AI 生成内容再被 AI 抓取学习”的恶性循环，形成了一个严重缺乏真实行业上下文的“回音室（Echo Chamber）”。结果就是，该模型遗漏了市场上真正关键的在售药物，却把早已退市的药物计算在内；它根本不理解竞争情报分析师到底需要什么样的核心商业逻辑。 相比之下，右侧展示的 Atlas CI（Atlas 竞争情报 Agent）则完全不同。它的核心优势有两点：第一，它在 Norstella 独有的、全面贯通的完整行业数据集上完成了训练；第二，也是更为关键的一点，它是由“真正在行业一线从事这项工作的资深专家”亲自参与训练调优的。Kris 警醒听众，不能妄想把一千名实习生塞在一起就能自动具备资深分析师的眼界和业务深度。无论是做竞争情报、市场准入洞察、高度复杂的药物销量预测，还是设计严谨的临床试验方案，面对的都是极度微妙且精细的专业任务与商业决策。 | 对比维度 | 通用 LLM（挂载公开文献与试验库） | Atlas CI（垂直专用 Agent） | |:--- |:--- |:--- | | **数据源** | 公共互联网、PubMed、ClinicalTrials.gov、AI 生成内容 | 纵向打通的 Norstella 独家全生命周期医药数据 | | **质量与准确度** | 产生 AI 回音室效应；漏算在售药、误算退市药 | 数据准确核实，紧扣商业竞争现实，无盲区幻觉 | | **训练指导者** | 基础语言建模与泛化训练 | 由实际从事医药分析的资深一线专家联合训练调优 | | **业务理解深度** | 停留在表面文本汇总，缺乏商业情报视角 | 理解复杂商业决策场景（预测、准入、协议设计等） | --- ### [15:53](https://www.youtube.com/watch?v=M5K78HGlmvE&t=953s) 从 Demo 迈向落地：四大核心标准与总结展望 在演讲的尾声，Kris 抛出了留给现场所有人的核心思考题。当大家在下周一回到办公室，开始规划公司交代给自己的各项 AI 专项、着手设计大模型或搭建 Agent 时，务必用四个核心标准来审视自己的方案：引述溯源（Cited）、输出一致（Consistent）、贴合场景（Contextual）以及关乎结果（Consequential）。如果一个 AI 方案能够同时满足这四项准则，它就能打破玩具和摆设的魔咒，实现“从 Demo 演示跃迁到业务决策，从实验原型转化为可企业级规模化交付的真实生产力”。这是 Norstella 在深入市场观察并反复实践后所获得的宝贵真知。 最后，由于演讲时间有限，Kris 热情邀请对该领域感兴趣的参会者在会后继续深入交流，欢迎随时前往 Norstella 的 1034 号展位（Booth 1034）探讨探讨。在走下讲台前，Kris 留下了全场最具分量的一句结语：“请时刻记住：我们是在为‘决策’构建基础设施，而不是为了‘AI’而构建 AI。（We are building infrastructure for decisions. We are not building for AI.）” - **四 C 准则（AI 规模化落地的评判维度）**： - **Cited（来源可溯）**：所有输出与结论必须有坚实权威的源头数据支撑。 - **Consistent（结果一致）**：在关键业务指标与趋势推导上保持严密稳定，经得起反复核验。 - **Contextual（深嵌场景）**：具备深厚的制药专业上下文，绝非泛泛而谈的网文拼贴。 - **Consequential（关乎结果）**：能直接驱动实际行动与战略决策，带来可衡量的商业与临床成果。 - **展位互动与收尾**：Norstella 团队在现场 1034 号展台为观众提供面对面咨询与 Demo 体验。 ## Notable quotes > [02:34](https://www.youtube.com/watch?v=M5K78HGlmvE&t=154s) 花得更多、耗时更长、成功率更低——在这个号称被 AI 彻底重塑的时代，我们究竟该如何解释这种巨大的断层？ — Kris Kaneta > [04:14](https://www.youtube.com/watch?v=M5K78HGlmvE&t=254s) 在制药领域，从业者承担的都是动辄数十亿美元、押上身家性命的高风险决策，比如是否为某个分子投入 30 亿至 50 亿美元、是否拓展全新的适应症、临床试验方案如何设计、选定的研究中心研究者是否能满足入组指标等。这些决策没有“撤销键”（No Control-Z），一旦失误便无可挽回 — Kris Kaneta > [04:14](https://www.youtube.com/watch?v=M5K78HGlmvE&t=254s) 当无法百分之百信任 AI 输出的内容时，绚丽的 Demo 就永远无法转化为实际的商业决策。 — Kris Kaneta > [10:02](https://www.youtube.com/watch?v=M5K78HGlmvE&t=602s) 最优秀的 AI 基础设施，从来都不是为 AI 而建的。不要为了 AI 而建 AI，一切建设必须以‘促成决策’为唯一导向。 — Kris Kaneta > [10:02](https://www.youtube.com/watch?v=M5K78HGlmvE&t=602s) 面对一千名渴望产出却缺乏行业背景的实习生，管理者必须做到三件事：第一，必须极其明确地告知他们任务的具体目标；第二，必须清晰界定基础事实与知识来源，因为刚入职的实习生自身并不具备专业上下文；第三，必须极其清楚地定义‘优秀的结果长什么样’ — Kris Kaneta > [10:02](https://www.youtube.com/watch?v=M5K78HGlmvE&t=602s) AI 模型在本质上就是这一千名实习生，如果脱离了严密的决策支持框架，它产出的东西就毫无意义。 — Kris Kaneta > [12:42](https://www.youtube.com/watch?v=M5K78HGlmvE&t=762s) 必须打破这种部门壁垒，建立正确的‘决策基础设施（Decision Infrastructure）’。 — Kris Kaneta > [14:04](https://www.youtube.com/watch?v=M5K78HGlmvE&t=844s) 不能妄想把一千名实习生塞在一起就能自动具备资深分析师的眼界和业务深度。无论是做竞争情报、市场准入洞察、高度复杂的药物销量预测，还是设计严谨的临床试验方案，面对的都是极度微妙且精细的专业任务与商业决策。 — Kris Kaneta > [15:53](https://www.youtube.com/watch?v=M5K78HGlmvE&t=953s) 从 Demo 演示跃迁到业务决策，从实验原型转化为可企业级规模化交付的真实生产力 — Kris Kaneta > [15:53](https://www.youtube.com/watch?v=M5K78HGlmvE&t=953s) 请时刻记住：我们是在为‘决策’构建基础设施，而不是为了‘AI’而构建 AI。（We are building infrastructure for decisions. We are not building for AI.） — Kris Kaneta ## Mentioned Kris Kaneta, Norstella, Ai4, Citeline, MMIT, Evaluate, Panalgo, The Dedham Group, Atlas, Atlas CI, Claude, Gemini, ChatGPT, PubMed, ClinicalTrials.gov ## Video description Kris Kaneta - Chief Product Officer, Norstella Why does it still take so long to get life-changing treatments to patients? We know it’s not a lack of data or tools. In this session, I’ll talk about what’s actually slowing us down and why most solutions haven’t fixed it. I’ll share how, at Norstella, we’re using AI to connect data across the pipeline to patient journey, helping teams move from question to decision, and from decision to action. Come see how we’re reducing the work it takes for pharma teams to get to the next step and how we’re fundamentally changing how teams interact with their data. Because when that changes, everything moves faster. And therapies reach patients sooner. [Source](https://www.youtube.com/watch?v=M5K78HGlmvE)

[2:34] Kris Kaneta: 花得更多、耗时更长、成功率更低——在这个号称被 AI 彻底重塑的时代，我们究竟该如何解释这种巨大的断层？ [4:14] Kris Kaneta: 在制药领域，从业者承担的都是动辄数十亿美元、押上身家性命的高风险决策，比如是否为某个分子投入 30 亿至 50 亿美元、是否拓展全新的适应症、临床试验方案如何设计、选定的研究中心研究者是否能满足入组指标等。这些决策没有“撤销键”（No Control-Z），一旦失误便无可挽回 [4:14] Kris Kaneta: 当无法百分之百信任 AI 输出的内容时，绚丽的 Demo 就永远无法转化为实际的商业决策。 [10:02] Kris Kaneta: 最优秀的 AI 基础设施，从来都不是为 AI 而建的。不要为了 AI 而建 AI，一切建设必须以‘促成决策’为唯一导向。 [10:02] Kris Kaneta: 面对一千名渴望产出却缺乏行业背景的实习生，管理者必须做到三件事：第一，必须极其明确地告知他们任务的具体目标；第二，必须清晰界定基础事实与知识来源，因为刚入职的实习生自身并不具备专业上下文；第三，必须极其清楚地定义‘优秀的结果长什么样’ [10:02] Kris Kaneta: AI 模型在本质上就是这一千名实习生，如果脱离了严密的决策支持框架，它产出的东西就毫无意义。 [12:42] Kris Kaneta: 必须打破这种部门壁垒，建立正确的‘决策基础设施（Decision Infrastructure）’。 [14:04] Kris Kaneta: 不能妄想把一千名实习生塞在一起就能自动具备资深分析师的眼界和业务深度。无论是做竞争情报、市场准入洞察、高度复杂的药物销量预测，还是设计严谨的临床试验方案，面对的都是极度微妙且精细的专业任务与商业决策。 [15:53] Kris Kaneta: 从 Demo 演示跃迁到业务决策，从实验原型转化为可企业级规模化交付的真实生产力 [15:53] Kris Kaneta: 请时刻记住：我们是在为‘决策’构建基础设施，而不是为了‘AI’而构建 AI。（We are building infrastructure for decisions. We are not building for AI.）

[4:03] Kris Kaneta: there is no control Z [10:02] Kris Kaneta: Don't build for AI. Build to get to a decision. [10:02] Kris Kaneta: eating our own dog food [12:13] Kris Kaneta: give me the ADHD market landscape

[0:04] Good morning everybody. Thank you so much for  coming out and hopefully you guys are off to a [0:08] great start in AI4. At least you were smart enough  and got your badges early. That line outside looks [0:14] insane. So um anyway, my name is Chris Ketta and  I am the chief product and innovation officer for [0:22] Norstella. Some of you coming to us from life  sciences may be familiar with Norstella. Uh [0:28] many of you have probably engaged with some of our  brands out in the market. Brands like Sighteline, [0:33] brands like MMIT, brands like Evaluate, Panalgo,  the Denim Group to name a few. Uh and some of [0:38] you are learning about Norstella for the very  first time. So if you all will indulge me, um [0:47] oh no, is that the right Yeah. Okay. Just making  sure I'm looking at the right slides here. [0:51] If you all will indulge me, what I'd like to do is  just level set on what Nstella is. what do we do? [0:58] Uh provide a little bit of context as we  spend the next 15 or so minutes talking about [1:03] what we're calling this AI productivity paradox.  Okay. So, [clears throat] Norstella, at Norstella, [1:10] we provide the data and the intelligence backbone  to some of Pharma's biggest decisions, some of [1:17] their highest stakes decisions from early pipeline  to portfolio strategy to clinical trial planning [1:25] to market access to commercialization. And in that  work, right, from pipeline to patient, we support [1:34] all 100 of's top 100, we support over 3,000  clients across the globe. And in any given year, [1:41] better than four and five life-changing  therapies that come to market are in some way [1:48] shaped, informed, and guided by our data and  our intelligence. that intelligence manifests [1:55] itself in analytical products in predictive and  prescriptive models and recommendation engines, [2:04] APIs, MCPS, SLMs, LLMs, and of course agents. So  we have a a fairly wide aperture as we think about [2:14] what is happening across the  pharma and biotech landscape. [2:19] And all of you are here because obviously we're  here talking about AI. We're here talking about [2:24] artificial intelligence. And we will all be super  sick of those two letters by the end of the week. [2:30] So I'm glad we're early here and we're  getting to this session early. But [2:34] what I want to do is maybe take you  on a little bit of a journey as to [2:37] what we are seeing happening across the life  sciences landscape. Because the reality is [2:45] despite the enthusiasm, despite the energy,  despite massive exuberance and investment into AI [2:52] over the last few years, the scorecard for  pharma has actually gone in the wrong direction. [3:00] All right, it now takes us twice as much  investment in terms of new drug development. [3:08] Our drug development timelines have actually  gone up by more than 10% in the last 5 years [3:15] and our probability of technical and  regulatory success. Our likelihood [3:19] of getting to that approval has gone down.  I'll say that again. We are spending more. [3:27] Our timelines are getting longer and our  likelihood to succeed is getting worse. [3:33] How do we explain this? Right? How do we  explain this? All of you in your roles, [3:38] in your jobs as leaders, you're looking at  investments into AI and this promise that it's [3:44] going to transform the way we do work, right? It's  going to make us faster. It's going to make us [3:49] smarter. It's going to make us more productive.  It's going to make us more informed. And yet, [3:54] in the midst of this AI era, this renaissance,  we've actually become less productive. [4:03] So, what I want to do is take you into a little  bit of what we've learned over the last few years [4:07] as it relates to AI and our journey, how we're  seeing it get applied across life sciences. [4:14] Because in broad strokes, there's really  two camps here that we're observing. [4:19] The first camp are those organizations that  are going headirst. They're diving in. They've [4:26] got a commitment. They know what they want to  accomplish. They've got a defined set of outcomes. [4:32] They can tell you what good looks like in terms  of what they want out of their AI investments. [4:38] And there's a second camp that's saying,  you know what, we're not exactly sure what [4:43] these implications are. We're not exactly sure  how we're going to apply it, but we know that [4:47] there's an opportunity here. We know that there's  an art of the possible. We just need to find it [4:53] and uncover it and be able to articulate it. And  in both camps, the problem statement is actually [5:00] the same. Okay? And that is inside of pharma,  inside of what we do. We got a pilot problem. [5:10] All of you guys know, you know what I'm talking  about, right? The AI slop. It permeates everything [5:18] we are seeing across our organizations,  right? We've got pilots and prototypes [5:24] and proofs of concepts. And the problem that we  are seeing and observing is that they don't scale. [5:31] They don't scale because I can't technically  scale it. They don't scale because [5:36] I don't have the right contextual understanding  of the problem I'm trying to solve. [5:41] They don't scale because it doesn't  fit and apply to the workflow [5:46] in which I make decisions. Right? And so  fundamentally we have a problem of trust. [5:55] Because these prototypes exist and they sound and  look very compelling. They sound very convincing. [6:02] But at the end of the day, what's really going  on is that all of you in pharma like like us, [6:09] we are charged with making really big decisions.  Really big high stakes decisions. Decisions like [6:19] do I spend three, four, five  billion dollars on this molecule? [6:23] decisions like do I expand into this new  indication? Decisions like how do I design [6:29] the protocol for my clinical trial? Decisions  like will this set of investigators help me [6:34] meet my enrollment goals in my clinical trial.  Right? These are all big decisions for which [6:42] there's no control Z, right? There is no  undo if we get this wrong. And so all of [6:48] you understand that the decisions we are being  tasked with have not only massive implications [6:54] in terms of the investment themselves, but they  have a massive knock-on effect in terms of the [6:59] subsequent decisions that they will drive, the  subsequent actions that they will trigger. And [7:07] what are we left with? We're left with not  able to trust the output that we're seeing. [7:12] And so we've experienced this ourselves, right? So  for many years we've been providing this data and [7:18] intelligence background for more than a decade at  Norstella. We've been applying AI, right? Many of [7:24] you are probably on the similar journey. Natural  language, machine learning, small language models, [7:29] large language models, you guys are doing  it all, right? And at the end of the day, [7:34] what we are left with is a challenge of how do  I take the next step? How do I go from a demo [7:41] to a decision? And so at Nordstella, as we've as  we've wrestled with some of these same challenges, [7:49] we've built our own, you know,  rubric as in terms of how we [7:52] evaluate the dozens of agents that we  have operating inside of our business. [7:57] Uh the agents that we've got deployed in our  products, the agents that we've got in our [8:01] software development, and [clears throat]  they all have their own unique rubric. [8:05] But at the end of the day, it really comes down  to four things. Four things that create trust. [8:12] And this is where really where we want to get into  what I hope will provoke you and something that [8:17] you guys can take back as you think about, you  know, what am I going to do differently on Monday? [8:23] For us, we call it the clarity criteria. So before  we fall in love with this incredible prototype, [8:31] with this incredible model, with this  incredible plug-in, ask yourself whether [8:36] it solves these four criteria. Number  one is the output that is being generated [8:45] cited. Can I cite it back to a source of truth  that I believe in? Number two, is it consistent? [8:55] When it cites me that source of truth, if I ask  that same question, that same task five times, [9:01] am I getting that same consistent source of  truth such that I can believe in it? Right. [9:07] Number three, is it contextual? Does it actually  represent the work that I am trying to do, [9:16] the job that I've been tasked with accomplishing?  Does it have context into the ecosystem of data [9:23] and information and insights that drives my  organization? And lastly, is it consequential? [9:30] Is the output that I have just created, is  the decision that I have just gotten to, [9:35] is the task that I'm triggering, is it going to  drive the right set of actions and behaviors? [9:43] That's it. Right now, obviously, there's going to  be a much more sophisticated rubric as you guys [9:48] think about evaluating your own AI journey and  the models that you are building. But is it cited? [9:54] Is it consistent? Is it contextual? And is it  consequential? Because at the end of the day, [10:02] the best AI infrastructure was never  built for AI. Don't build for AI. [10:12] This is my plea to all of you as you think about  what you are creating. Don't build for AI. Build [10:19] to get to a decision. All right? Take pretend  for a minute that when you went back to work, [10:28] you had no Claude, you had no Gemini,  you had no Chad GPT, and instead I said, [10:36] "Here is 1,000 interns." That's  really what we're talking about here. [10:41] If I gave you a thousand interns, and for those of  you that have had interns and work with interns, [10:45] you know where I'm going with this. I have to  be super explicit with what I want to achieve [10:52] with an intern. Number one. Number two, I have to  be very clear as to where that source of truth, [10:58] where that foundational understanding is going  to come from because that intern doesn't have it [11:04] right. And number three, I've got to be  really clear in terms of what good looks like. [11:09] What is the output I'm expecting to see? How do I  want it structured? What decisions is that going [11:15] to feed into? what actions is that going to  create for me? Right? So, don't build for AI. [11:22] Build to get to that decision and pretend  that you didn't have those models, [11:27] but instead had a thousand hopefully capable  interns. You could still get to that decision. [11:34] That's what we've had to wrestle with over the  last few years as we've built our own models [11:39] and we've had to in many respects eat our  own dog food, right? Some of the models [11:44] we came out with first didn't have that  citability. They didn't have that context, [11:51] weren't linked together. Right? As  we think about that context layer, [11:55] being able to see sources of truth is one thing.  Being able to see sources of truth that are linked [12:02] historically and longitudinally such that I trust  that it's going to give me the consistent right [12:07] answer, that's entirely different. So, we've had  to go through some of these learnings ourselves [12:13] and what we what we've done is we've released  Atlas and and I apologize guys, this is not meant [12:18] to sound like a product pitch. I'm by the way very  happy to talk to you about what we're building [12:23] in Atlas, but I'm using this for a context. If  you want to learn more though, please come by [12:26] and talk to us um on the show floor. But [snorts]  what we did with Atlas was we said you know what [12:34] there is a need for this decisionra  infrastructure. There is a need for this data [12:42] and this intelligence to be linked historically  and longitudinally such that if I were to train [12:49] a thousand interns or one great agent to solve  a task to complete a job, it would be able to [12:56] cut across that entire set of information. Right?  So what Atlas is is an agentic platform that sits [13:03] on top our entire our entire corpus of data from  drug pipeline to trial to investigator to payer [13:12] to coverage and market access to real world  data to be able to cut across that and say hello [13:19] I am a competitive intelligence analyst. Hello I'm  a feasibility analyst. Hello I'm a market access [13:25] and insights analyst. whatever it might be. The  problem is is that in our day jobs, doesn't matter [13:31] where we are, that information is siloed. That  information is stuck, right? And your business [13:38] development folks have data that they don't  want to share with their market access folks. [13:42] Your market access folks have things that they  don't want to share with their forecasting folks. [13:46] And you've got to break free of that. You've got  to create the right decision infrastructure. And [13:50] that's what Atlas set out to do. And then  we said, we're going to build very specific, [13:56] highly informed agents on top that solve  a real use in what we do every day. [14:04] So to make it real, on your left, you will see  a query that is just to a standard a standard [14:12] generic LLM connected to PubMedct.gov. We asked  it, "Hey, give me the ADHD market landscape." [14:23] And we asked it many times, by the way, right  through a series and battery of queries. [14:27] Every time we got a super compelling answer,  super credible sounding, but when you dig into it, [14:35] what was missing? Well, it was scanning public  domain. It was scanning, you know, all sorts of [14:42] AI generated content. By the way, that's the other  thing, right? The more content we generate by AI, [14:48] the more that content feeds back into itself  and we get this echo chamber that lacks context, [14:53] right? But it was missing drugs that were in  the market. It was counting drugs that were [14:58] no longer in the market and it ultimately did not  understand the role of a competitive intelligence [15:05] agent in this case versus on the right hand side  this is Atlas CI or Atlas competitive intelligence [15:12] that has been trained against our entire set of  data number one but number two more importantly [15:18] was train was trained by experts who actually  do the job and this is the last thing we [15:25] really need to take way. We cannot expect a  thousand interns to have the sophistication [15:32] of a competitive intelligence analyst, of a  market access analyst, of somebody that is [15:37] doing a very complex forecast, as somebody that is  designing a protocol. These are extremely nuanced, [15:45] highly specific tasks and decisions that we  are trying to get to every single day. So [15:53] what I really want you guys to take away with  is as you go back on Monday as you guys are [15:56] thinking about hey what are the AI what are  the AI initiatives that I've been tasked with [16:01] how am I going to create this model how am I  going to build this agent again think back to is [16:07] what I'm created cited consistent contextual and  consequential if you can answer those four things [16:19] you will go from a demo to a decision. You will  go from a prototype to something that is real, [16:25] that you can scale, that you can  deliver across your organization. [16:29] This is at least based on what we've learned  and what we've observed happening in the market. [16:34] I realize that we've got a limited amount of  time here. So, I'll simply close by saying [16:39] I would love to talk more about this. Anybody in  our on our team that's here would love to talk [16:43] to you more about this. You can come by on uh our  booth and uh give us a give us a hello booth 1034. [16:52] But just remember we are building infrastructure  for decisions. We are not building for AI. [16:59] All right. Thank you guys so much for being with  me here. Hopefully you enjoyed it. Thank you.