Claritev Eyes 3.5x AI Returns With Move to Agentic Health Workflows

The AI Compass: Navigating the Landscape of AI and Generative AI With Claritev — Ai4

Ai4

Organizations realize an average $3.50 return per dollar invested within 14 months only when AI shifts from side-desk experimentation to fully funded, product-managed architectures.

AI-assisted summary of The AI Compass: Navigating the Landscape of AI and Generative AI With Claritev — original source. Check the original for full context.

Source published: 2025-08-24

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 24, 2025, this presentation features Tamara Sipes, Vice President of Artificial Intelligence at Claritev. According to the video description, the talk cuts through confusing technical jargon surrounding machine learning, deep learning, and generative AI to provide business leaders with foundational concepts and practical steps to navigate enterprise AI adoption.

## TL;DR Tamara Sipes makes the business case for artificial intelligence, highlighting its projected multitrillion-dollar macroeconomic impact and an average 3.5x return on investment. She demystifies technical terminology by tracing the discipline's lineage from 1950s symbolic AI to deep learning, generative transformers, and autonomous multi-agent systems. To achieve sustained value, Sipes introduces three non-negotiable operational pillars: building foundational readiness, securing corporate enablement, and managing AI as an engineered product. She illustrates this roadmap using Claritev's Ben Insights platform as a case study, concluding with a nautical metaphor that urges organizations to experiment boldly and embrace failing fast.

## Atmosphere & style The presentation carries a calm, authoritative, and demystifying tone, avoiding breathless hype in favor of clear-eyed pragmatic realism. Drawing upon over twenty years of applied AI experience and her personal passion for boating in San Diego, Sipes communicates with the steady reassurance of a seasoned captain guiding a crew through turbulent waters. Watching the talk feels structured, pedagogical, and enterprise-focused, translating complex technical architectures into actionable, boardroom-ready strategy.

## Key takeaways - [01:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=68s) Artificial intelligence digital transformation is no longer optional, with macroeconomic forecasts projecting a $15T to $16.5T contribution to the global economy by 2030. - [01:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=68s) While 92% of enterprise AI implementations take up to a year and average 14 months to realize value, organizations achieve an average return of $3.50 for every $1.00 invested. - [02:38](https://www.youtube.com/watch?v=JhNZ13SzAog&t=158s) Artificial intelligence is an overarching superset founded in the 1950s that encompasses non-learning deterministic algorithms (like A* search), machine learning, deep learning, and generative AI. - [02:38](https://www.youtube.com/watch?v=JhNZ13SzAog&t=158s) The advent of the transformer architecture in 2017 catalyzed rapid generative AI progress, evolving from early GPT models to current autonomous agentic systems. - [06:25](https://www.youtube.com/watch?v=JhNZ13SzAog&t=385s) An AI agent fundamentally represents 'AI with autonomy' executing discrete tasks, whereas agentic AI coordinates tools, plans actions, and directs workflows with higher levels of autonomy. - [06:25](https://www.youtube.com/watch?v=JhNZ13SzAog&t=385s) Multi-agent systems enable peer-to-peer collaboration and iterative evaluation between specialized agents, rather than relying solely on centralized orchestration. - [10:14](https://www.youtube.com/watch?v=JhNZ13SzAog&t=614s) Successful enterprise AI adoption requires three core pillars: AI Readiness (talent, data, platform), AI Enablement (funding, executive backing, governance CoE), and Treating AI as a Product. - [10:14](https://www.youtube.com/watch?v=JhNZ13SzAog&t=614s) Treating AI as a product demands continuous drift monitoring to trigger retraining, building reusable modular components, and early alignment across engineering, security, and legal teams. - [13:37](https://www.youtube.com/watch?v=JhNZ13SzAog&t=817s) Claritev evolved its Ben Insights healthcare platform from static reporting in 2015 to predictive ML and now generative/agentic AI POCs like natural language queries and smart trend cards. - [17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) Organizations should embrace rapid experimentation and 'capsize fast and forward' on POCs, capturing learnings from failed tests to bring high-value solutions to production.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=JhNZ13SzAog&t=4s) Introduction & Agenda: Navigating the AI Seas

Tamara Sipes, Vice President of Artificial Intelligence at Claritev, opens the session by welcoming attendees and outlining a roadmap designed to demystify artificial intelligence for organizational leaders. Acknowledging that many professionals feel bewildered by an unrelenting wave of technical terminology, Sipes aims to establish AI as an accessible, high-leverage tool in the enterprise toolkit. She structures the presentation around four central topics: unpacking the buzzwords from classical algorithms to modern agentic systems, defining the core organizational pillars required for successful adoption, walking through the real-world evolution of Claritev’s Ben Insights product, and concluding with strategic takeaways drawn from her nautical framing of enterprise navigation.

[01:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=68s) The Economic Urgency and ROI of AI Adoption

Sipes establishes that digital transformation driven by artificial intelligence is no longer optional for modern enterprises: "Really, there is no opting out of AI." Citing projections from Forbes and other macroeconomic forecasters, she notes that AI is expected to contribute between $15 trillion and $16.5 trillion to the global economy by 2030. Enterprise leadership is responding in kind, with 92% of executives expecting continued or increased focus on AI initiatives over the coming year. Across the market, 76% of companies are actively leveraging AI in some capacity, and 69% of those users have already integrated generative AI into their workflows.

Addressing execution realities, Sipes emphasizes that AI implementation cannot be shoehorned into a conventional "two-week sprint." While roughly 92% of enterprise AI deployments take up to a full year to implement, organizations realize tangible business value within an average of 14 months. The capital justification is compelling: enterprise benchmarks demonstrate an average return of $3.50 for every $1.00 invested. This 3.5x ROI explains why executive suites across industries are accelerating investments in production AI capabilities.

- **Projected Global Impact**: $15T to $16.5T added to the global economy by 2030. - **Executive Sentiment**: 92% planning sustained or increased AI investment. - **Adoption Rates**: 76% of enterprises utilize AI; 69% of these deploy generative AI. - **Time to Value**: Average of 14 months to realize measurable business value. - **Capital Efficiency**: $3.50 return on investment for every $1.00 spent.

[02:38](https://www.youtube.com/watch?v=JhNZ13SzAog&t=158s) Unpacking the Buzzwords: From Dartmouth 1956 to Transformers

Sipes dispels the popular misconception that artificial intelligence originated three years ago with ChatGPT, pointing out that AI was founded in the 1950s and represents a field in which she has worked for more than two decades following classical training. She visualizes AI as an all-inclusive superset encompassing all subdisciplines: "If you're doing machine learning you're doing AI. If you're doing generative AI, you're doing AI." Within this hierarchy, machine learning focuses on learning directly from data experience through methods such as decision trees, regression trees, and neural networks. To illustrate AI outside of machine learning, Sipes points to standard smartphone mapping applications, which rely on deterministic graph search techniques like the A* pathfinding algorithm rather than trained statistical models.

```mermaid flowchart TD accTitle: Artificial Intelligence Hierarchy accDescription: Nested hierarchy showing AI as the superset containing Machine Learning, Deep Learning, Generative AI, and Agentic AI.

subgraph AI["Artificial Intelligence (Superset founded in 1950s)"] direction TB ASTAR["Symbolic AI & Pathfinding (e.g., A* Algorithm)"]

subgraph ML["Machine Learning (Data-Driven Learning)"] CLASSIC["Decision Trees, Regression, Fraud Detection"]

subgraph DL["Deep Learning (Multi-Layer Neural Networks ~2010)"] PERCEPTION["Face ID, Speech, Vision, NLU"]

subgraph GEN["Generative AI (Transformers ~2017)"] GPT["GPT-1 through GPT-4, LLM Generation"]

subgraph AGENT["Agentic AI (~2024+)"] AUTO["Autonomous Reasoning, Tool Calling & Workflows"] end end end end end ```

The evolutionary timeline accelerated during the late 2000s and early 2010s with deep learning, where multi-layered neural networks proved capable of automated feature extraction from unstructured data. Deep learning unlocked consumer-grade facial recognition—such as Apple's Face ID—alongside advanced natural language understanding. Generative AI arrived around 2017 with the invention of the transformer architecture—the structural "T" in Generative Pre-trained Transformers (GPT). Sipes charts the rapid iteration cycle from GPT-1 in 2018, GPT-2 in 2019, and GPT-3 in 2020, to the watershed launch of ChatGPT on GPT-3.5 in late 2022, culminating in the current frontier of agentic AI.

[06:25](https://www.youtube.com/watch?v=JhNZ13SzAog&t=385s) Defining AI Agents, Agentic AI, and Multi-Agent Workflows

Turning to the latest industry frontier, Sipes defines an AI agent at its most fundamental level as "AI with autonomy." While market definitions vary widely, she demarcates individual components along an autonomy spectrum ranging from level-one task processors and routers up to level-five fully autonomous systems (which she notes have yet to materialize in enterprise production). A basic AI agent executes discrete tasks—such as summarizing a call center transcript or translating a document—without needing complex reasoning.

In contrast, agentic AI operates with higher intelligence and autonomy, possessing the ability to perceive operational environments, formulate plans, call external tools or APIs, learn from outcomes, and orchestrate specialized agents. Sipes differentiates agentic workflows (automated, predetermined execution paths replacing legacy manual handoffs) from multi-agent systems. While agentic AI emphasizes centralized orchestration and tool execution—much like a soccer coach directing specialized players across the pitch—multi-agent systems focus on peer-to-peer collaboration:

> "A multi-agent system on the other hand can be something like: 'Okay, let's extract insight from PDFs,' and that's one agent, and then another agent evaluates the first agent and they go back and forth. 'Hey, I extracted this from this PDF.' And the evaluator says, 'Ah, I don't see it. Go back, do it again.' And so they can converse and work together collaboratively to achieve a goal."

[10:14](https://www.youtube.com/watch?v=JhNZ13SzAog&t=614s) The Three Pillars of Enterprise AI Success

Reflecting on why some organizations achieve outsized returns while others stall, Sipes lays out three non-negotiable operational pillars: AI Readiness, AI Enablement, and Treating AI as a Product. Readiness forms the foundational tier, demanding specialized in-house talent, clean data infrastructure, and an enterprise environment capable of model training, evaluation, and production deployment. Enablement supplies the institutional backing: "Remember, you have to invest $1 to get to the 3.5. You cannot do it with zero dollars. Somebody just tinkering on the side of their desk is not going to achieve the same results." Enablement includes executive sponsorship, documented roadmaps, and an AI Center of Excellence enforcing responsible, fair, and ethical governance.

```mermaid flowchart LR accTitle: Three Pillars of Enterprise AI Success accDescription: Flow showing AI Readiness leading to AI Enablement, resulting in AI as a Product.

subgraph P1["1. AI Readiness"] T["Specialized Talent"] D["Data Infrastructure"] M["Model Platform"] end

subgraph P2["2. AI Enablement"] F["Strategic Funding"] E["Executive Support"] G["Governance & CoE"] end

subgraph P3["3. AI as a Product"] L["Lifecycle Delivery"] O["Drift Monitoring"] R["Component Reuse"] end

P1 --> P2 --> P3 ```

The final pillar requires treating artificial intelligence not as an isolated research experiment, but as an engineered product. Product discipline means understanding target personas (external clients, internal developers, customer service agents), tracking production metrics, and actively monitoring data drift to trigger early model retraining before scheduled maintenance cycles. Furthermore, modern architectures must prioritize reusable modular components—such as an agent built to parse complex PDFs repurposed across disparate workflows. Sipes advises enterprises to assess "build, buy, or partner" at every layer while seating legal, enterprise information security, and engineering teams at the same table early via a RACI framework.

[13:37](https://www.youtube.com/watch?v=JhNZ13SzAog&t=817s) Claritev's Ben Insights: A Practical AI Journey

Sipes grounds these frameworks by detailing Claritev’s operational footprint and product evolution. As a health-tech organization focused on healthcare affordability, transparency, and clinical quality, Claritev supports over 700 payer organizations, employers, third-party administrators (TPAs), and brokers, maintaining a provider network of 1.4 million clinicians serving more than 60 million plan members. Within Claritev's Data and Decision Science Services unit, the Ben Insights platform has operated since 2015, ingesting over 200 distinct healthcare data sources to help self-insured employers and brokers optimize benefit design.

When launched in 2015, Ben Insights relied exclusively on automated, deterministic reporting without machine learning. Five years ago, Claritev embarked on an AI modernization program mapped directly against the three pillars: 1. **AI Readiness Phase**: Capital investment in dedicated machine learning teams, unified data pipelines, and scalable modeling infrastructure. 2. **Predictive & Prescriptive Modeling**: Deployment of traditional machine learning algorithms to model financial risk, clinical utilization, and plan recommendations. 3. **AI Enablement Phase**: Formal executive commitment, corporate AI roadmaps, and an AI Center of Excellence establishing governance protocols. 4. **Product Coupling & Lifecycle Discipline**: Close integration with product managers, automated drift monitoring, structured retraining cadences, and modular component libraries. 5. **Agentic AI & GenAI Readiness**: Provisioning of an enterprise-grade, secure generative AI platform permitting safe LLM consumption. Claritev is currently delivering two production proof-of-concept (POC) applications for Ben Insights: conversational natural-language querying over complex healthcare datasets, and dynamic "smart cards" that flag emerging clinical and market trends—such as the sudden utilization spikes driven by GLP-1 weight-loss medications—far faster than legacy static reporting.

[17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) Nautical Takeaways: Setting Sail Toward an AI-First Enterprise

Drawing on her passion for boating in San Diego, Sipes summarizes her guidance through a maritime metaphor for navigating the enterprise "seascape" of artificial intelligence:

- **Choose Your Vessel Wisely (AI Readiness)**: You get what you pay for. Secure high-caliber talent, clean data assets, and robust evaluation platforms before attempting deployment. - **Set Sail (AI Enablement)**: Back technical capabilities with dedicated corporate capital, executive alignment, clear strategic direction, and an active governance Center of Excellence. - **Reach Your Destination (AI as a Product)**: Build for production reliability by maintaining tight integration with end users, monitoring model drift, and institutionalizing component reuse. - **Capsize Fast and Forward**: Proof-of-concept experiments are meant to test uncharted waters; many will fail. When an experiment capsizes, fail fast, capture the empirical learnings, and iterate forward. Only a small, battle-tested percentage of POCs need to transition into production to generate outsized organizational value.

## Notable quotes > Have you ever felt lost in a sea of AI terms like machine learning, deep learning, or Gen AI? Planning to bring AI as a tool in your toolkit? This talk is for you! — Tamara Sipes

> It's time to stop feeling overwhelmed and start exploring AI with confidence! — Tamara Sipes

> [01:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=68s) Really, there is no opting out of AI. — Tamara Sipes

> [02:38](https://www.youtube.com/watch?v=JhNZ13SzAog&t=158s) If you're doing machine learning you're doing AI. If you're doing generative AI, you're doing AI. — Tamara Sipes

> [06:25](https://www.youtube.com/watch?v=JhNZ13SzAog&t=385s) AI with autonomy. — Tamara Sipes

> [06:25](https://www.youtube.com/watch?v=JhNZ13SzAog&t=385s) A multi-agent system on the other hand can be something like: 'Okay, let's extract insight from PDFs,' and that's one agent, and then another agent evaluates the first agent and they go back and forth. 'Hey, I extracted this from this PDF.' And the evaluator says, 'Ah, I don't see it. Go back, do it again.' And so they can converse and work together collaboratively to achieve a goal. — Tamara Sipes

> [10:14](https://www.youtube.com/watch?v=JhNZ13SzAog&t=614s) Remember, you have to invest $1 to get to the 3.5. You cannot do it with zero dollars. Somebody just tinkering on the side of their desk is not going to achieve the same results. — Tamara Sipes

> [17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) Choose Your Vessel Wisely — Tamara Sipes

> [17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) Capsize Fast and Forward — Tamara Sipes

## Mentioned Tamara Sipes, Claritev, Ai4, Ben Insights, Forbes, A* algorithm, Apple Face ID, ChatGPT, GPT-1, GPT-2, GPT-3, GPT-3.5, GPT-4, GLP-1

## Video description Tamara Sipes, Vice President of Artificial Intelligence, Claritev

Abstracts

Have you ever felt lost in a sea of AI terms like machine learning, deep learning, or Gen AI? Planning to bring AI as a tool in your toolkit? This talk is for you! I will cut through the jargon and explain what these buzzwords actually mean. More importantly, you'll walk away with a solid understanding of fundamental AI concepts and practical, actionable steps you can take to begin your own journey into the exciting world of artificial intelligence. It's time to stop feeling overwhelmed and start exploring AI with confidence!

[Source](https://www.youtube.com/watch?v=JhNZ13SzAog)

中文简报

Claritev转向智能体健康工作流，瞄准3.5倍AI回报只有当AI从边缘实验转向获得全额资助且按产品化管理的架构时，机构才能在14个月内实现每投入1美元平均获得3.50美元的回报。 Tamara Sipes阐述了人工智能的商业价值，重点指出了其预计可达数万亿美元的宏观经济影响以及平均3.5倍的投资回报率。她梳理了该学科从1950年代的符号AI到深度学习、生成式Transformer以及自主多智能体系统的演变脉络，对相关技术术语进行了解构。为获得持续价值，Sipes提出了三大不可或缺的运营支柱：构建基础就绪能力、确保企业赋能，以及将AI作为工程化产品进行管理。她以Claritev的Ben Insights平台作为案例阐释了这一实施路径，并在最后借用航海比喻，鼓励各机构大胆实验并接受快速试错。宏观预测显示到2030年AI将为全球经济贡献15万亿至16.5万亿美元；尽管92%的企业AI落地耗时长达一年、平均需14个月兑现业务价值，但企业平均每投入1美元可实现3.50美元的回报。 (1:08) 人工智能是始于1950年代的宏大超集，不仅涵盖机器学习、深度学习与生成式AI，还包括如A*寻路等非学习型确定性算法；2017年Transformer架构的突破推动了GenAI的爆发，并演进至当前的智能体系统。 (2:38) 基础AI智能体（AI Agent）代表具备自主性并执行离散任务的实体，而Agentic AI则具备感知环境、规划行动与调用工具的高阶编排能力，多智能体系统更支持专职智能体之间的对等协作与双向评估。 (6:25) 企业成功落地AI依赖三大支柱：AI就绪度（人才、数据与基础设施）、AI赋能（专项资金支持与治理卓越中心），以及将AI视为产品进行全生命周期管理（包含数据漂移监控与模块化组件复用）。 (10:14)

Reading summary

Tamara Sipes, Vice President of AI at Claritev, argues that capturing artificial intelligence's projected economic value requires treating AI as an engineered product supported by foundational technical readiness and institutional enablement rather than an ad-hoc technical experiment. Economic Urgency and Measured Return on Investment (1:08) Citing macroeconomic projections from Forbes and other forecasters, Sipes notes that artificial intelligence is projected to contribute between $15 trillion and $16.5 trillion to the global economy by 2030. Enterprise adoption reflects this outlook: 76% of companies report using AI, with 69% of those users integrating generative AI into workflows, and 92% of executives planning sustained or increased investment over the coming year. However, Sipes stresses that enterprise AI deployments cannot be compressed into conventional two-week sprints. Roughly 92% of enterprise AI implementations take up to a full year, with organizations realizing measurable business value within an average of 14 months. Despite this ramp-up timeline, enterprise benchmarks show an average return of $3.50 for every $1.00 invested, which Sipes cites as the principal justification for executive-backed funding. The AI Hierarchy and Emerging Agentic Workflows (2:38) Sipes clarifies that artificial intelligence is not a three-year-old technology born with ChatGPT, but an overarching discipline founded in the 1950s that encompasses machine learning, deep learning, and deterministic logic. Non-statistical methods, such as the A* pathfinding algorithm used in mapping applications, illustrate AI capabilities outside machine learning. Following the rise of deep learning around 2010 and the transformer architecture in 2017, the technological frontier has progressed to agentic AI. Sipes defines agents along an autonomy spectrum ranging from level-one task processors to theoretical level-five fully autonomous systems, noting that level-five systems do not yet exist in enterprise production. While basic agents handle discrete tasks such as translation, agentic AI perceives environments, formulates plans, and executes tool calls. In contrast to centrally orchestrated workflows, multi-agent systems deploy peer-to-peer collaboration, where specialized agents iteratively evaluate and refine each other's outputs. Structural Pillars: Readiness, Enablement, and Product Discipline (10:14) To bridge the gap between pilot programs and tangible ROI, Sipes outlines three mandatory organizational pillars. AI Readiness demands high-quality data infrastructure, specialized internal talent, and model evaluation environments. AI Enablement provides institutional backing, pairing dedicated capital and executive sponsorship with an AI Center of Excellence to enforce ethical governance across a RACI framework that includes legal and security teams early. The third pillar mandates managing AI initiatives as products rather than isolated experiments. This discipline entails defining target user personas, monitoring data drift to trigger proactive retraining cycles, and architecting reusable modular components across disparate business functions. Sipes advises enterprises to systematically evaluate whether to build, buy, or partner at each layer of the technical architecture. Practical Application in Claritev's Ben Insights (13:37) Claritev's health-tech platform, Ben Insights, illustrates this structured progression. Serving over 700 payer and employer organizations and ingesting more than 200 healthcare data sources, the platform launched in 2015 using purely deterministic reporting without machine learning. Five years ago, Claritev began modernizing the system across all three pillars by building unified data pipelines, predictive utilization models, and enterprise governance. Ben Insights has now advanced to testing generative and agentic proof-of-concept applications. These include natural language querying across healthcare datasets and automated "smart cards." The dynamic smart cards identify clinical utilization anomalies—such as surging demand for GLP-1 weight-loss medications—significantly faster than legacy static reporting mechanisms. Sipes emphasizes that enterprise AI implementation requires an organizational appetite for calculated failure during the proof-of-concept stage. Because production-grade deployment relies on disciplined data monitoring, component reuse, and institutional backing, organizations only need a small, battle-tested percentage of experimental projects to reach production to realize compounding capital returns.

精炼阅读摘要

Claritev 人工智能副总裁 Tamara Sipes 指出，数字化转型已使企业无法回避人工智能，但要实现平均 3.5 倍的投资回报，组织必须摆脱零散摸索，通过就绪能力、组织赋能与产品化管理三大支柱建立系统工程。经济紧迫性与投资回报周期 (1:08) Sipes 援引《福布斯》等宏观经济预测指出，到 2030 年人工智能预计将为全球经济贡献 15 万亿至 16.5 万亿美元。企业管理层正积极响应这一趋势，调查显示 92% 的受访高管计划在未来一年维持或增加对人工智能项目的投入。从普及度来看，目前已有 76% 的企业在业务中以某种形式应用 AI，其中 69% 的采用者已将生成式 AI 整合至日常工作流程中，AI 已从边缘技术实验升级为核心战略重心。然而，技术落地往往无法套用常规的两周敏捷冲刺周期。Sipes 强调，约 92% 的企业级 AI 部署需要长达一整年才能完成实施，组织平均需耗时 14 个月方能实现切实的业务价值。尽管落地周期较长，其资本效益依然显著：行业基准数据显示，企业在人工智能领域的平均投资回报率为每支出 1.00 美元获得 3.50 美元收益。这一高达 3.5 倍的投资回报率，正是各行业高管团队加速将 AI 能力推向生产环境的关键动力。技术演进谱系与智能体协同 (2:38) Sipes 澄清了业界常见的认知误区，强调人工智能并非始于三年前的 ChatGPT，而是一个建立于 20 世纪 50 年代、涵盖所有分支技术的宏大超集。机器学习属于数据驱动经验学习的子集，涵盖决策树、回归算法与神经网络等；而传统的确定性图搜索技术（例如智能手机地图导航所依赖的 A* 路径规划算法）同样属于人工智能范畴。深度学习在 2010 年前后崛起，通过多层神经网络实现无结构化数据的自动特征提取，推动了人脸识别（如 Face ID）和自然语言理解的应用。2017 年 Transformer 架构的发明催生了生成式预训练模型，从 GPT-1 快速迭代至支撑 ChatGPT 的 GPT-3.5 及后续大语言模型，并最终演进至当前的智能体（Agentic AI）阶段。在智能体领域，Sipes 将基础 AI 智能体定义为“具备自主性的 AI”，主要负责转录、翻译等特定任务执行，不需要复杂推理。相比之下，高阶智能体（Agentic AI）拥有更高的自主与决策能力，能够感知运行环境、自主制定规划、调用外部工具或 API，并从执行结果中学习反馈。她进一步区分了基于预设流程的智能体工作流与多智能体系统（Multi-Agent System）：前者类似于主教练指挥调度各类专业角色，后者则强调智能体之间的点对点协同交互。例如，一个智能体负责从 PDF 文档中抽取关键信息，另一个评估智能体则对提取质量进行审查与打回，二者通过动态对话与循环校对共同达成复杂目标。企业成功落地的三大支柱 (10:14) 针对部分组织部署 AI 取得显著成效而其他企业停滞不前的原因，Sipes 归纳了三大不可或缺的运营支柱：就绪度（AI Readiness）、赋能体系（AI Enablement）与产品化思维（Treating AI as a Product）。就绪度作为基础层，要求企业具备专业的人才储备、干净统一的数据基础设施，以及支持模型训练、评估和上线的生产环境。赋能体系则提供体制与资本保障，Sipes 直言要获取 3.5 倍的回报就必须先投入真金白银，仅靠员工在办公桌边业余摸索绝不可能成功。赋能要求明确的高管支持、清晰的战略路线图，以及负责推行负责任、公平与伦理治理的 AI 卓越中心（CoE）。最后一项支柱要求企业将人工智能视为工程化产品，而非孤立的科研实验。产品化规范要求深度理解目标用户角色（外部客户、内部开发者或客服人员），严密追踪生产环境中的各项技术指标，并主动监测数据漂移（Data Drift）以及时触发早期重训机制，避免依赖固定的定期维护周期。此外，企业必须推进模块化组件的复用设计（例如跨工作流复用 PDF 解析智能体），在各个技术层级科学评估自建、外购或合作（Build, Buy, or Partner），并在早期通过 RACI 权责框架将法务、信息安全与工程团队紧密结合。 Claritev 医疗数据产品的演进实践 (13:37) Sipes 以 Claritev 旗下医疗数据平台 Ben Insights 的演化历程阐释了上述落地框架。Claritev 专注于医疗可负担性、透明度与临床质量，连接了 700 多家支付方机构、雇主和经纪人，医疗网络涵盖 140 万名临床医生，覆盖超 6000 万名参保成员。该平台的研发部门接入超过 200 个医疗数据源，用于辅助自保雇主和经纪人优化福利方案设计。在 2015 年上线之初，Ben Insights 仅提供基于规则的自动化确定性报表，完全不涉及机器学习技术。五年前，Claritev 依据三大支柱开启了人工智能现代化转型：首先在就绪阶段重金投入专属机器学习团队、统一数据管道与可扩展建模基础设施；随后上线传统机器学习算法，实现财务风险、医疗利用率及方案推荐的预测与规范性建模；进入赋能阶段后，设立 AI 治理卓越中心并明确企业路线图；在产品化层面建立了紧密的跨职能协同、自动化漂移监测与模块化组件库。目前，该平台已进入生成式与智能体阶段，正试点交付两项生产级概念验证（POC）：一项是针对复杂医疗数据集的自然语言会话交互查询，另一项是动态“智能卡”，能够以远超传统报表的速度监测识别新型临床与市场异动（例如 GLP-1 减重药物使用量的爆发式增长）。 Sipes 借用航海比喻为企业总结了航向法则：选好船只（夯实数据与人才就绪度）、扬帆启航（配齐资金、高管支持与治理赋能）、抵达彼岸（坚持产品化生命周期与抗漂移监测）。在这一航程中，概念验证项目本质是在未知水域试错，企业应建立“快速且向前翻船”（Capsize Fast and Forward）的容错机制，将失败视为经验积累并持续迭代，最终只需将少数经过充分验证的 POC 推向生产环境，便足以创造跨越式的商业价值。

Detailed notes — zh

## Background 本演讲由 Ai4 于 2025 年 8 月 24 日发布在 YouTube 平台上，由 Claritev 人工智能副总裁 Tamara Sipes 主讲。根据简介，演讲旨在拨开机器学习、深度学习及生成式 AI 等行业流行词的迷雾，帮助企业与从业者建立扎实的 AI 核心概念认知，并掌握开启 AI 旅程的实用落地步骤与探索信心。 ## TL;DR 演讲首先通过详实的数据（如企业平均 14 个月见效、投入 1 美元带来 3.5 美元回报）论证了企业拥抱 AI 的必然趋势，并系统梳理了从经典 AI、机器学习、深度学习到生成式 AI 与智能体 AI（Agentic AI）的技术全景。随后，主讲人提出了企业成功落地 AI 的三大支柱：AI 准备度（人才与基础设施）、AI 赋能（战略投入与卓越中心治理）以及 AI 即产品（全生命周期闭环与监控复用）。结合 Claritev 医疗数据平台 BenInsights 的实战演进案例，演讲阐述了从传统报表到智能卡片（Smart Cards）与自然语言交互的落地路径，并以航海隐喻鼓励企业建立“快速试错、向前迭代（Capsize fast and fail forward）”的 AI 优先文化。 ## Atmosphere & style 整场演讲语调务实、亲切且充满启发性，既有严谨的商业架构拆解，又兼具深入浅出的科普温度。主讲人巧妙融入圣迭戈的航海隐喻，将复杂的智能体演进与企业治理具象化为选船、扬帆与掌舵的过程，有效缓解了听众面对技术术语爆炸时的焦虑感。演讲节奏明快流畅，在理性分析技术价值与鼓励大胆试错之间达成了绝佳平衡，营造出一种让人跃跃欲试、信心充沛的探索氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=JhNZ13SzAog&t=4s) AI 已经成为不可逆转的数字化浪潮，企业平均在 14 个月内见效，每投入 1 美元可带来约 3.5 美元的投资回报率（ROI）。 - [02:38](https://www.youtube.com/watch?v=JhNZ13SzAog&t=158s) AI 是涵盖机器学习与深度学习的“全集（Superset）”，最早诞生于 1950 年代，并非近几年才出现的新概念。 - [05:07](https://www.youtube.com/watch?v=JhNZ13SzAog&t=307s) 技术正从基于 Transformer 的生成式 AI 迅速演进为具备感知、规划、工具调用与自主适应能力的“智能体 AI（Agentic AI）”。 - [06:30](https://www.youtube.com/watch?v=JhNZ13SzAog&t=390s) 企业需明确区分 AI 智能体（特定任务）、智能体 AI（总控调度）、智能体工作流（预定义自动化）与多智能体系统（协同交互校验）。 - [10:10](https://www.youtube.com/watch?v=JhNZ13SzAog&t=610s) AI 落地第一支柱（AI 准备度）：企业必须在 AI 专业人才、数据与基础设施底座，以及模型开发部署平台上坚决投入。 - [11:00](https://www.youtube.com/watch?v=JhNZ13SzAog&t=660s) AI 落地第二支柱（AI 赋能）：需要高管层的战略支持与资金投入，并建立卓越中心（CoE）推行负责任且合规的 AI 治理。 - [12:00](https://www.youtube.com/watch?v=JhNZ13SzAog&t=720s) AI 落地第三支柱（AI 即产品）：将 AI 视作真实产品进行端到端闭环管理，持续监控数据漂移（Data Drift）并沉淀可复用的 Agent 组件。 - [12:45](https://www.youtube.com/watch?v=JhNZ13SzAog&t=765s) 在推进 AI 项目时，应及早通过 RACI 矩阵拉通信息安全（InfoSec）与法务合规团队，并审慎权衡自研（Build）、采购（Buy）或合作（Partner）。 - [13:37](https://www.youtube.com/watch?v=JhNZ13SzAog&t=817s) Claritev 的 BenInsights 实战展示了从基础规则报表，演进至传统机器学习风险模型，再到如今 GenAI 自然语言问答与智能卡片的完整路径。 - [17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) 概念验证（POC）应秉持“快速倾覆、向前迭代（Capsize fast and fail forward）”的航海哲学，以体系化的 AI 罗盘驱动组织迈向 AI 优先。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=JhNZ13SzAog&t=4s) 开启 AI 之旅：不可逆转的数字化浪潮与商业回报 AI 已经不再是一个可选项，数字化转型的浪潮正在重塑商业世界并创造巨大的实际价值。许多企业在面对机器学习、深度学习、生成式 AI 以及最新涌现的智能体 AI（Agentic AI）等各种专业词汇时常常感到无所适从。企业应当将 AI 视作工具箱中不可或缺的利器，认清其带来的颠覆性机遇——正如演讲所指出的，当前根本"无法置身于 AI 之外（There is no opting out of AI）"。从宏观经济与企业投入来看，AI 的商业价值正在以极快的速度释放。企业引入 AI 并非短期两周冲刺（Sprint）就能完成的任务，它需要更系统性的建设周期，但整体投资回报率非常亮眼： - **宏观经济规模**：据《福布斯》预测，到 2030 年，AI 将为全球经济贡献 15 万亿美元（其他机构估算在 16 万亿至 16.5 万亿美元之间）。 - **高管关注度**：92% 的企业高管预计今年将继续维持或进一步加大对 AI 的战略投入。 - **企业应用比例**：全球已有 76% 的企业在使用 AI，其中 69% 正在应用生成式 AI（Generative AI）。 - **价值转化周期与回报**：约 92% 的 AI 项目在一年内完成落地，企业平均在 14 个月内即可看到明确价值；更关键的是，企业在 AI 上每投入 1 美元，就能带来 3.5 美元的投资回报率（ROI）。 --- ### [02:38](https://www.youtube.com/watch?v=JhNZ13SzAog&t=158s) 厘清 AI 术语全景：从传统机器学习到深度学习在充斥着各种时髦术语的当下，必须明确人工智能（AI）是所有这些概念的"全集（Superset）"与总括。许多人误以为 AI 是三年前随着 ChatGPT 才诞生的新技术，但实际上 AI 作为一门制造智能机器的科学与工程学科，早在 1950 年代就已经创立。机器学习（Machine Learning）则是 AI 的一个重要子集，其核心在于"从数据经验中进行学习"。通过决策树（Decision Trees）、回归树（Regression Trees）、神经网络（Neural Networks）等高效算法，自动从海量数据中提取规律与洞察。值得注意的是，AI 并不等同于机器学习——比如日常手机地图应用中规划路线所使用的 A* 寻路算法，属于经典的 AI 算法而非机器学习。机器学习如今广泛应用于信用卡反欺诈、虚假与滥用识别（Fraud, Waste, Abuse）、在线精准营销及邮件过滤等场景。在 2000 年代末至 2010 年代初，行业正式步入深度学习（Deep Learning）时代。研究人员发现拥有大量隐藏层的深层神经网络能够极为高效地自动提取训练数据中的特征，从而促成了计算机视觉、自然语言处理（NLP）和自然语言理解（NLU）等领域的突破性进展。当用户使用 iPhone 通过面容 ID（Face ID）刷脸解锁手机时，背后运行的正是深度学习算法。 --- ### [05:07](https://www.youtube.com/watch?v=JhNZ13SzAog&t=307s) 从生成式 AI 到智能体 AI（Agentic AI）的技术跃迁生成式 AI（GenAI）的浪潮发端于 2017 年前后 Transformer 架构的诞生。作为一种特殊的深度学习模型，Transformer 能够根据输入的文本序列精准预测后续内容，这也构成了现代 GPT 系列模型（Generative Pre-trained Transformer）中的核心字母 "T"。从 2018 年的 GPT-1、2019 年的 GPT-2、2020 年的 GPT-3，再到 2022 年底基于 GPT-3.5 打造的 ChatGPT，生成式 AI 彻底引爆了公众与企业市场。而从去年底到今年初，技术演进迅速迈入了智能体 AI（Agentic AI）的新纪元。 ```mermaid flowchart TD accTitle: 智能体与自主系统的演进架构 subgraph Core["核心概念划分"] A["AI 智能体 (AI Agent)<br/>执行特定任务的自主系统（如文本摘要、翻译）"] B["智能体 AI (Agentic AI)<br/>具备感知、规划、工具调用、学习与自适应能力"] end subgraph Workflows["工作流与协作"] C["智能体工作流 (Agentic Workflow)<br/>替代人工的预定义 AI 自动化流程"] D["多智能体系统 (Multi-Agent Systems)<br/>多 Agent 协同、对话与互相评估校验"] end A --> B B --> C B --> D ``` 在最高维度上，智能体 AI 可以简单定义为"拥有自主性的 AI（AI with autonomy）"。为了准确区分当前易混淆的概念，可以从以下四个维度进行界定： - **AI 智能体（AI Agent）**：能够自主执行特定任务的系统。任务既可以很简单（如总结客服通话记录或翻译文本），也可以更复杂，其自主权涵盖从简单路由器/工具调用到完全自主的 5 个级别。 - **智能体 AI（Agentic AI）**：拥有更高自主性与智能水平的系统，能够感知外部环境、主动规划、调用工具、学习自适应，并协调调用一个或多个 AI 智能体。它像是一个足球队主教练，清晰了解每个球员的专长并调度全场。 - **智能体工作流（Agentic Workflow）**：由 AI 驱动、遵循预定义路径的业务流程，旨在自动化过去需要人工完成的操作，过程中可嵌入多个 Agent。 - **多智能体系统（Multi-Agent Systems）**：侧重于多个 Agent 之间的对话与协作。例如一个 Agent 负责从 PDF 中提取洞察，另一个 Agent 充当评估员检验提取结果并反馈修正，双方来回交互以共同达成目标。 --- ### [10:10](https://www.youtube.com/watch?v=JhNZ13SzAog&t=610s) 企业落地 AI 的三大支柱：准备度、赋能与产品化思维企业要成功将 AI 转化为商业价值，并非单靠个人在办公桌边零星摸索就能实现，而是需要系统化构建成功的三大核心支柱：AI 准备度、AI 赋能以及将 AI 视作产品。 ```mermaid flowchart LR accTitle: 企业 AI 成功的落地三支柱 subgraph P1["第一支柱：AI 准备度 (AI Readiness)"] T["AI 顶尖人才"] D["数据与基础设施"] PF["模型开发与部署平台"] end subgraph P2["第二支柱：AI 赋能 (AI Enablement)"] I["战略资金投入"] L["领导层支持与愿景"] G["卓越中心 (CoE) 与负责任治理"] end subgraph P3["第三支柱：AI 即产品 (AI as Product)"] U["明确用户与端到端闭环"] M["性能监控与数据漂移维护"] R["组件与 Agent 模块复用"] end P1 --> P2 --> P3 ``` - **第一支柱：AI 准备度（AI Readiness）**：企业必须储备扎实的 AI 专业人才，搭建健全的数据基础设施以保证数据可用性，并建立可进行模型开发、评估与生产化部署的统一平台。 - **第二支柱：AI 赋能（AI Enablement）**：必须有实质性的战略资金支持（"投入 1 美元才能收获 3.5 美元"）与高管层背书；同时建立清晰的战略愿景，组建卓越中心（Center of Excellence, CoE），制定负责任、公平且符合伦理道德的 AI 治理准则。 - **第三支柱：AI 即产品（AI as Product）**：明确 AI 的使用对象与应用场景（无论是外部客户、内部开发人员还是呼叫中心员工）；实现从创意到投产的端到端闭环；部署严密的性能监控机制以应对数据漂移（Data Drift）并触发维护；积极推行组件复用（如将已建成的 PDF 提取 Agent 复用到不同业务场景中）。 - **跨部门协同机制**：在推进过程中，必须及早将信息安全团队（InfoSec）、法务合规部门、数据团队等关键利益相关方纳入沟通机制（可借助 RACI 矩阵），确保全流程合规与可信。针对每一个环节，企业均可权衡"自研（Build）、采购（Buy）或合作（Partner）"的最佳路径。 --- ### [13:37](https://www.youtube.com/watch?v=JhNZ13SzAog&t=817s) Claritev 实战案例：BenInsights 产品的 AI 演进之路作为一家致力于提升医疗健康可负担性、透明度与服务质量的医疗科技企业，Claritev 拥有覆盖 140 万家医疗服务提供者的网络，服务超过 6000 万名会员，并与 700 多家支付方机构（包含大量雇主、第三方管理人 TPA、经纪人及医疗机构）开展合作。其旗下的 BenInsights 解决方案自 2015 年上线投产，汇聚了超过 200 个数据源，协助管理者为特定人群设计最佳医疗方案并挖掘可执行的业务洞察。 BenInsights 在过去近十年中经历了一场系统性的 AI 演进之旅： 1. **初始阶段（2015 年）**：系统尚未引入 AI，主要依赖自动化报表与规则处理。 2. **第一阶段：夯实准备度与基础风险模型（五年前）**：公司投资组建 AI 团队并建设数据与模型底座，成功上线了首批基于传统机器学习的预测模型、规范性（Prescriptive）与描述性（Descriptive）模型，覆盖医疗方案的财务与临床双重视角。 3. **第二阶段：深化 AI 赋能与战略对齐**：全面追加战略投资，设立 AI 卓越中心，确立公司级 AI 愿景。 4. **第三阶段：产研紧密结合与持续运营**：AI 与产品团队深度绑定，建立严格的性能监控指标，持续推出重大特性并沉淀可复用组件。 5. **第四与第五阶段：迈入智能体 AI 准备度与 GenAI 落地**：针对生成式 AI 搭建了安全合规的大语言模型（LLM）底层平台。目前 BenInsights 正在推进两项概念验证（POC），支持用户通过自然语言直接向数据提问，并利用智能卡片（Smart Cards）实时捕获市场动态（例如针对 GLP-1 类热点药物的趋势分析，比传统报表更迅速、敏捷且体验友好）。 --- ### [17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) 航海视角的 AI 总结：快速失败、向前迭代与 AI 优先结合长期生活在圣迭戈（San Diego）对帆船与航海的热爱，演讲用一套生动的航海隐喻对企业落地 AI 的路径进行了总结提炼： - **明智选船，做好准备（Choose your vessel wisely）**：一分钱一分货，企业必须在人才、数据质量与模型工作空间上坚决投入。 - **扬帆启航，战略赋能（Set sail）**：借助领导层支持、卓越中心与清晰战略，推动 AI 从实验走向企业级赋能。 - **紧盯航向，产品化落地（Reach your destination）**：把 AI 视为真实产品进行全生命周期管理与例行维护，保障长期可靠与信任。 - **快速倾覆，向前失败（Capsize fast and fail forward）**：概念验证（POC）本质上就包含试错空间，企业需要勇于探索前所未有的创新；一旦遇到失败要做到快速失败、向前迭代，将吸取的经验转化为最终成功投产的卓越产品。 - **AI 罗盘指引航程（AI Compass）**：企业面对的不是静止的陆地，而是一片不断变化的"AI 与生成式 AI 海域（Seascape）"，只有借助科学的 AI 罗盘体系，才能真正将组织建设为"AI 优先（AI-first）"的现代化企业。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=JhNZ13SzAog&t=4s) 当前根本无法置身于 AI 之外（There is no opting out of AI）。 — Tamara Sipes > [00:04](https://www.youtube.com/watch?v=JhNZ13SzAog&t=4s) 企业在 AI 上每投入 1 美元，就能带来 3.5 美元的投资回报率（ROI）。 — Tamara Sipes > [02:38](https://www.youtube.com/watch?v=JhNZ13SzAog&t=158s) 人工智能（AI）是所有这些概念的“全集（Superset）”与总括。 — Tamara Sipes > [05:07](https://www.youtube.com/watch?v=JhNZ13SzAog&t=307s) 在最高维度上，智能体 AI 可以简单定义为“拥有自主性的 AI（AI with autonomy）”。 — Tamara Sipes > [10:10](https://www.youtube.com/watch?v=JhNZ13SzAog&t=610s) 企业要成功将 AI 转化为商业价值，并非单靠个人在办公桌边零星摸索就能实现，而是需要系统化构建成功的三大核心支柱：AI 准备度、AI 赋能以及将 AI 视作产品。 — Tamara Sipes > [10:10](https://www.youtube.com/watch?v=JhNZ13SzAog&t=610s) 企业可以权衡自研（Build）、采购（Buy）或合作（Partner）的最佳路径。 — Tamara Sipes > [17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) 明智选船，做好准备（Choose your vessel wisely）：一分钱一分货，企业必须在人才、数据质量与模型工作空间上坚决投入。 — Tamara Sipes > [17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) 快速倾覆，向前失败（Capsize fast and fail forward）：概念验证（POC）本质上就包含试错空间，企业需要勇于探索前所未有的创新；一旦遇到失败要做到快速失败、向前迭代。 — Tamara Sipes > [17:08](https://www.youtube.com/watch?v=JhNZ13SzAog&t=1028s) 企业面对的不是静止的陆地，而是一片不断变化的“AI 与生成式 AI 海域（Seascape）”，只有借助科学的 AI 罗盘体系，才能真正将组织建设为“AI 优先（AI-first）”的现代化企业。 — Tamara Sipes ## Mentioned Tamara Sipes, Claritev, Ai4, Forbes (福布斯), BenInsights, ChatGPT, Transformer, GPT-1, GPT-2, GPT-3, GPT-3.5, Apple Face ID (面容 ID), A* 寻路算法, GLP-1 ## Video description Tamara Sipes, Vice President of Artificial Intelligence, Claritev Abstracts Have you ever felt lost in a sea of AI terms like machine learning, deep learning, or Gen AI? Planning to bring AI as a tool in your toolkit? This talk is for you! I will cut through the jargon and explain what these buzzwords actually mean. More importantly, you'll walk away with a solid understanding of fundamental AI concepts and practical, actionable steps you can take to begin your own journey into the exciting world of artificial intelligence. It's time to stop feeling overwhelmed and start exploring AI with confidence! [Source](https://www.youtube.com/watch?v=JhNZ13SzAog)

Source quotes

[0:04] Tamara Sipes: 当前根本无法置身于 AI 之外（There is no opting out of AI）。 [0:04] Tamara Sipes: 企业在 AI 上每投入 1 美元，就能带来 3.5 美元的投资回报率（ROI）。 [2:38] Tamara Sipes: 人工智能（AI）是所有这些概念的“全集（Superset）”与总括。 [5:07] Tamara Sipes: 在最高维度上，智能体 AI 可以简单定义为“拥有自主性的 AI（AI with autonomy）”。 [10:10] Tamara Sipes: 企业要成功将 AI 转化为商业价值，并非单靠个人在办公桌边零星摸索就能实现，而是需要系统化构建成功的三大核心支柱：AI 准备度、AI 赋能以及将 AI 视作产品。 [10:10] Tamara Sipes: 企业可以权衡自研（Build）、采购（Buy）或合作（Partner）的最佳路径。 [17:08] Tamara Sipes: 明智选船，做好准备（Choose your vessel wisely）：一分钱一分货，企业必须在人才、数据质量与模型工作空间上坚决投入。 [17:08] Tamara Sipes: 快速倾覆，向前失败（Capsize fast and fail forward）：概念验证（POC）本质上就包含试错空间，企业需要勇于探索前所未有的创新；一旦遇到失败要做到快速失败、向前迭代。 [17:08] Tamara Sipes: 企业面对的不是静止的陆地，而是一片不断变化的“AI 与生成式 AI 海域（Seascape）”，只有借助科学的 AI 罗盘体系，才能真正将组织建设为“AI 优先（AI-first）”的现代化企业。

Source quotes

Tamara Sipes: Have you ever felt lost in a sea of AI terms like machine learning, deep learning, or Gen AI? Planning to bring AI as a tool in your toolkit? This talk is for you! Tamara Sipes: It's time to stop feeling overwhelmed and start exploring AI with confidence! [1:08] Tamara Sipes: Really, there is no opting out of AI. [2:38] Tamara Sipes: If you're doing machine learning you're doing AI. If you're doing generative AI, you're doing AI. [6:25] Tamara Sipes: AI with autonomy. [6:25] Tamara Sipes: A multi-agent system on the other hand can be something like: 'Okay, let's extract insight from PDFs,' and that's one agent, and then another agent evaluates the first agent and they go back and forth. 'Hey, I extracted this from this PDF.' And the evaluator says, 'Ah, I don't see it. Go back, do it again.' And so they can converse and work together collaboratively to achieve a goal. [10:14] Tamara Sipes: Remember, you have to invest $1 to get to the 3.5. You cannot do it with zero dollars. Somebody just tinkering on the side of their desk is not going to achieve the same results. [17:08] Tamara Sipes: Choose Your Vessel Wisely [17:08] Tamara Sipes: Capsize Fast and Forward

Original transcript

[0:04] Pleasure to be here. Well, welcome. [0:06] Today we're going to cover some of the [0:08] basics of AI and see how we can get how [0:11] you guys can think about getting started [0:13] using it as a tool in your toolkit. [0:16] So, we're going to start by kind of [0:18] introducing AI and why it's so important [0:22] and what all the buzz is all about. Um, [0:25] speaking of buzz, then we're going to [0:26] unpack some of these buzzwords. AI, [0:28] machine learning, deep learning, and on [0:30] and on and on, including the most recent [0:33] buzzwords such as AI agents and agentic [0:36] AI. [0:38] Then I'm going to start talking about [0:40] some of the important pillars of success [0:43] when you're starting to think about [0:44] bringing AI to your organization. [0:48] Um, I work for for Clarative, a health [0:51] tech company. I'm going to share how we [0:53] took our one of our products, one of our [0:56] solutions, Ben's insights on an AI [0:59] journey and what we did with that. And [1:01] then we're going to conclude with [1:02] takeaways and Q&A. [1:08] Really, there is no opting out of AI. [1:11] The transformation, this digital [1:13] transformation that AI is enabling, it's [1:15] happening and it's creating significant [1:18] business value. According to Forbes, [1:22] AI is going to bring 15 trillion to the [1:26] global economy by 2030. According to [1:29] others, 16 or 16.5. [1:33] Um 92% of executives [1:37] are expecting continues continued or [1:40] increased focus on AI [1:44] this year. [1:47] So, how is AI currently being used [1:49] around the world? Well, 76% of companies [1:54] are using AI. Out of that, 69% are using [1:59] generative AI. [2:02] 92% of AI is basically done within a [2:06] year. So, it's not your regular twoe [2:08] sprint. you know, it takes a little [2:10] longer, but um within 14 months on [2:14] average, companies are realizing the [2:16] value of AI. And most importantly, for [2:19] every invested dollar, AI is bringing [2:22] 3.5 ROI, return on investment. So for [2:26] every dollar, you get $3.5 back. So [2:29] that's pretty cool. And that's why, you [2:30] know, everybody is starting to really [2:32] think about how to bring AI into their [2:36] organizations. [2:38] So let's speak about these buzzwords. [2:42] Artificial intelligence is the superset [2:44] of all of these jargon words you've [2:47] you've heard. [2:50] Artificial intelligence is is the [2:51] science and engineering of making [2:53] intelligent machines. It was it was [2:57] invented [2:59] about three years ago, right? Which had [3:00] GPT? No, it was invented in 1950s. As a [3:04] matter of fact, I've worked in AI for 20 [3:06] plus years, and that's after being [3:08] classically trained in AI. So AI has [3:11] been around for a long time. A subset of [3:14] AI is machine learning. [3:17] Machine learning is really learning from [3:19] the experience of your data. you know, [3:22] we use these um very um efficient and [3:25] effective methods such as decision [3:27] trees, regression trees, neural networks [3:29] and so on to automatically sift through [3:31] the data and come up, you know, figure [3:34] out what the insight is. So you might [3:37] ask yourself, [3:39] what's in machine learning that's not [3:42] what's an AI that's not in machine [3:43] learning, [3:45] right? Well, if you bring out your [3:48] smartphone [3:50] and choose your favorite maps app, [3:53] you're using AI, an AI algorithm that's [3:56] not machine learning. It's actually an [3:58] AI store algorithm for path finding. So, [4:03] at the bottom you see the the timeline [4:05] of when these methods were invented. Um, [4:08] and then you can see some examples that [4:10] I put in there. So machine learning most [4:12] commonly being used by your credit card [4:15] company when you get that phone call uh [4:17] for fraud, waste, abuse and error for um [4:20] online marketing, email filtering and so [4:23] on. [4:25] In the late 20s and early 2010s, the era [4:29] of deep learning began when we when [4:32] [clears throat] researchers discovered [4:33] that these deep um neural networks, [4:36] meaning neural networks with a lot of [4:38] hidden layers, can be really good at [4:40] extracting features in in the training [4:42] data. And since then, we were able to do [4:47] even more advanced things such as facial [4:50] recognition, [4:51] um, natural language processing, natural [4:54] language understanding, and things like [4:56] that. Back to the iPhone prop, you can [4:59] open it with your face. When you do [5:01] that, you're using deep learning. Um, [5:03] [snorts] and then most recently, [5:07] um, [clears throat] the era of [5:09] generative AI began. It really began in [5:12] a little before 2017 when a transformer [5:16] was invented which is this really [5:18] special deep learning method. It's able [5:20] to generate text after you put in an [5:22] input sequence like predict what's [5:24] coming after that input sequence of text [5:27] and that transformer is the T in GPTs [5:31] today right u generative pre-trained [5:35] transformers in 2018 GPT1 was invented [5:40] 2019 GPT2 [5:43] 2020 GPT3 and then of course at the end [5:47] of 2022 we had the Chad GPT the chat [5:51] experience that was based on GPT3.5 [5:54] and then most recently as of late last [5:56] year and early this year we've entered [5:59] the era of a gentic AI. So I'm going to [6:02] talk a little bit more about that. Oh by [6:05] the way what's really important to [6:07] remember from this slide is that AI [6:11] is the superset. It's as I say all [6:14] inclusive. So if you're doing machine [6:16] learning you're doing AI. If you if [6:17] you're doing generative AI, you're doing [6:18] AI. So AI covers everything. But you [6:22] know, a lot of people don't use it [6:23] properly. All right. So what about AI [6:25] agent and agentic AI in the highest [6:28] level? At the highest level, it's just [6:31] AI with autonomy. [6:35] There are a lot of different definitions [6:37] these days, a ton of them. It's an [6:40] emerging field. A lot of people are [6:42] talking about, you know, systems that [6:44] autonomously learn that um, you know, [6:47] work without human um, assistance that [6:51] sometimes people or companies define it [6:54] as using LLMs or conversational bots or [6:58] machine learning. [7:00] Really, [7:02] an AI agent is an autonomous system that [7:05] that can execute a task. That task can [7:08] be fairly simple. It can be something [7:10] like summarize this text or it can be [7:14] translate this text. So you can have an [7:17] AI agent that summarizes say a call [7:21] center conversation for you and that's a [7:24] fairly simple AI agent. We we'll see um [7:26] that there are five different levels of [7:28] autonomy of AI agents. [7:31] Um agentic AI has more autonomy and more [7:35] intelligence. So agentic AI is are these [7:39] s systems that can perceive their [7:42] environment. They can plan, they can [7:45] execute, they can learn, they can adapt, [7:48] they can call one or multiple AI agents [7:51] or they can tool call some tools to [7:54] execute certain actions. [7:57] An agentic workflow. So there are other [7:59] related terms and a lot of people use [8:00] them interchangeably but agentic [8:03] workflow is really an AIdriven process [8:06] that follows a pre predetermined path [8:09] that was usually done manually and it [8:12] can use one or multiple AI agents. [8:15] And then the final term that you guys [8:17] need to be um kind of aware of is that [8:21] [clears throat] there are multi- aent [8:22] systems and those are really multiple [8:25] agents conversing and collaborating [8:29] together to complete a task. So there's [8:32] a little bit of a gray area and we're [8:34] not going to go into details between [8:35] Aentic AI and multi- aent systems if [8:38] your Agentic AI is using multiple [8:39] agents. But basically, Agentic AI is [8:42] focused on automation and intelligence [8:46] and multi- aent systems are focused on [8:49] collaboration and conversation among [8:51] these agents. Here's an example. Maybe [8:54] you can have an agentic AI that extracts [8:59] insights from multimodal data for you [9:01] and it has multiple agents that it [9:04] employs to do that. So it's almost like [9:06] a coach that has all these soccer [9:09] players on the field for example. It [9:11] knows who's good for what and and it's [9:13] kind of orchestrating these these [9:14] agents, right? Um a multi- aent system [9:18] on the other hand can be something like [9:21] okay um let's extract insight from PDFs [9:26] and that's one agent and then another [9:29] agent evaluates the first agent and they [9:32] go back and forth. Hey, I extracted this [9:34] from this from the uh this PDF. And the [9:37] evaluator says, ah, I don't see it. Go [9:39] back, do it again. And so they can kind [9:40] of converse and work together [9:42] collaborative collaboratively to achieve [9:44] a goal. All right. [9:48] So just very briefly, different at a [9:50] autonomy levels of AI agents. In the uh [9:55] rightmost hand column, you can see how [9:58] much who's in control. So, we go from a [10:01] simple processor, router, tool calling [10:03] agent to fully autonomous agent, which I [10:06] haven't seen any of those yet, but maybe [10:07] they'll coming soon. All right. So, [10:10] where do we start? How do we do AI, [10:11] right? [10:14] Well, what's that recipe for success? [10:16] Some you guys heard Dan the um keynote [10:20] introduction. [10:22] Some companies are really good at [10:24] utilizing AI and creating business [10:26] value. How do they do it? Well, there [10:28] are three pillars of success. AI [10:31] readiness, you have to be ready. You [10:32] have to have good strong AI talent. You [10:36] have to have good data and data [10:38] infrastructure so you can get to it. And [10:40] you have to have a good platform for [10:43] where you going to develop evaluate your [10:46] models and from where are you going to [10:47] productionize your models. [10:50] The next kind of level of success is [10:52] what I call AI enablement and that's [10:55] really having the strategic investment [10:58] in AI. Remember, you have to invested $1 [11:01] to get to the 3.5. You cannot do it with [11:04] zero dollars. So, you have to have that. [11:06] Somebody just tinkering on the side of [11:07] their desk is not going to achieve the [11:10] same results as somebody that's that's [11:11] fully funded and and ready to go. Um, of [11:15] course, leadership support. And then [11:18] just like with anything, you have to [11:19] have your vision and strategy and start [11:20] executing on it. You have to kind of be [11:23] organized. Um and then the one other [11:26] thing to mention is that it's always [11:28] good to have a set of best practices and [11:30] guidelines. Your responsible um fair and [11:33] biased ethical AI governance. Often [11:36] people talk about responsible use of AI. [11:38] So that center of excellence that's [11:40] going to enable really good quality AI [11:42] to be produced. The final pillar of [11:45] success is treating AI as as product. [11:49] Knowing who, what, where, when, how will [11:52] your AI, whether they're external [11:55] customer or internal developer or [11:57] internal call center person is really [12:00] important. having that end to end from [12:03] you know end to end AI capability [12:06] uh from product product just idea to the [12:10] full final product that's launched and [12:13] in in production um knowing what your [12:16] product how your product's doing having [12:18] performance monitoring really important [12:20] because you're going to have to have [12:22] your scheduled maintenance if you're [12:24] doing AI right but if your performance [12:26] monitoring tells you that something's [12:28] off the data has drifted and so on. [12:31] You're going to have to trigger that [12:32] maintenance earlier. And then of course, [12:35] always think about component reuse. And [12:37] that's really easy to do these days with [12:39] AI agents. If you have created an AI [12:42] agent that extracts insight from PDFs, [12:44] you can reuse it in multiple use cases [12:47] in different types of PDFs. [12:50] So this is how you can create AI that's [12:53] properly built, operationalized, [12:56] consumed, maintained, and trusted. [12:59] Now, this seems like a lot, but at each [13:02] of these steps, you can consider [13:04] building, buying, or partnering. Those [13:06] are all good options. [13:09] All right. And then [13:11] it's always really important to bring [13:13] your stakeholders to your pod, COVID [13:17] analogy, to your table and, you know, [13:20] make sure that all of these folks are [13:22] involved. You can use the racy matrix, [13:25] but you need to talk to your enterprise [13:27] information security early on and your [13:29] legal and of course your your data [13:31] people and and on and on and on. So just [13:34] really important to have everybody [13:35] involved. [13:37] Okay, so let's talk about Ben insights [13:39] AI journey. So very quickly, Claritv is [13:42] a health tech company that is focused on [13:45] improving affordability, transparency [13:47] and quality of healthcare. works with um [13:50] 700 plus payer type of organizations, a [13:54] lot of employers, thirdparty [13:55] administrators, brokers, uh providers. [13:59] We have a network of 1.4 million [14:02] providers and we serve over 60 million [14:05] members. [14:07] um been around for a long time and Ben [14:10] insights is a part of the data and [14:12] decision science services business unit [14:15] and it's really um a product that's been [14:17] in production since 2015. It enables [14:21] employers, thirdparty administrators [14:25] and brokers to design um an optimal plan [14:29] for the population at hand and then to [14:32] utilize all kinds of data 200 plus [14:35] sources to um derive actionable [14:38] insights. [14:39] So when we started in 2015, we didn't [14:42] have any AI. Uh we had um a lot of um [14:46] automated reports and things like that. [14:48] But then five years ago, we actually got [14:51] on the AI journey. All right. So this is [14:54] what we've done. [14:56] We invested in an AI team, a good data [15:00] infrastructure and good model [15:01] infrastructure. And so that first AI [15:04] readiness pillar enabled us to get to [15:07] the very first risk models. And so we [15:10] added these predictive modeling models [15:12] in in step two. Um and they were [15:16] traditional machine learning models and [15:18] they're not only predictive we're able [15:19] to do prescriptive modeling also [15:21] descriptive modeling both on financial [15:24] and clinical aspects of the plan data [15:29] then we invested in AI enable enablement [15:32] pillar as I spoke earlier um created a [15:36] strategic investment in AI company's [15:38] really excited about AI um we have our [15:41] vision and strategy built our center [15:44] excellence. [15:46] We proceeded with a tighter coupling [15:48] with product really knowing what AI is [15:52] doing working with the product team [15:53] really closely having performance [15:55] monitoring um adding major features here [15:58] and there um reusing some of these [16:00] components [16:02] and then we are at the at the fifth [16:05] phase right now we're entering the [16:07] agentic AI readiness. So those three [16:09] pillars I talked about, think about [16:12] doing it for machine learning and then [16:14] going back and doing it for Gen AI. Some [16:16] of the stuff is the same, but some of it [16:18] is slightly different. So, we just [16:20] finished enabling our Gen AI platform [16:24] where we can safely and securely use [16:27] large language models and we're [16:29] currently um going through we're we're [16:32] delivering two PC's for Ben insights [16:35] product line and they're they're really [16:39] focused on um creating a delightful user [16:43] experience for our customers where they [16:45] can you know ask questions of their data [16:48] or get these smart cards that are maybe [16:50] a little bit more re revealing and [16:51] really focused on what's happening in [16:53] the market. Uh back to when GLPs popped [16:56] up, you know, there was a new report, [16:58] but now we can actually see the trends [17:00] utilizing AI much faster, quicker, in a [17:03] more friendly manner. [17:06] All right, so that's where we are. So [17:08] let's talk about takeaways. [17:11] So going on with my nautical theme, I [17:13] live in San Diego, love boating and [17:16] nautical everything. So, choose your [17:18] vessel wisely. Be ready. You get where [17:22] you pay for. You have to have strong [17:24] talent. Data is super important and you [17:27] have to have a good space to work on [17:28] your models. Um, then set sail, you [17:32] know, work on your AI enablement. This [17:34] is the first phase. When you get to the [17:36] second phase and you really have that [17:39] strategic investment, um, leadership [17:42] support, center of excellence, vision [17:44] and strategy, you're going to be able to [17:45] do much more. [17:47] then you know you are always going to [17:50] reach your destination if you think [17:52] about AI as product um and it's really [17:56] important to know how your product is [17:57] doing when it's in production what's [17:59] going on um you know to maintain it [18:02] really um with a good cadence and so on [18:06] um then I added these two capsites fast [18:11] and forward so PC's are meant to fail a [18:15] lot of them it's Okay, you want to you [18:18] want to innovate, you want to try things [18:20] that have never been done before. It's [18:22] important, but then every time if you [18:24] fail fast, fail forward. So, learn from [18:27] that experience, then it's okay. It, you [18:30] know, some of the small percentage of [18:31] your PC's will go into production and be [18:33] amazing new products for for our [18:35] company. And then keeping up with the [18:39] theme, um, AI compass, we covered that [18:42] today. We're really navigating the [18:44] seascape, not the landscape of AI and [18:47] generative AI. And I hope you learned [18:49] some things and are starting to think [18:51] about how you can make um your [18:54] enterprise an AI first enterprise. [18:59] >> [applause]