Data-Driven Marketing with AI: Extracting Value from Customer Data

AI驱动的数据营销：挖掘客户数据价值 来自 AEG Vision、Northwell Health 和 Uniphore 的营销与分析负责人探讨了利用 AI 从结构化与非结构化客户数据中挖掘价值的实操现状。与会嘉宾分享了实际应用案例，包括利用机器学习通话跟踪避免对现有客户浪费广告支出、分析网站数字化趋势作为临床早期预警信号，以及提供微妙且动态排序的内容推荐。贯穿各个主题的讨论均强调，企业应放弃追求完美无瑕、完全统一的数据架构，转而开展直接与“北极星”业务成果挂钩的小型、范围明确的试点项目。 呼叫中心记录、SOP与复盘报告等非结构化企业数据，与结构化关系型数据相结合，能为微调营销AI模型提供至关重要的真实业务上下文支撑。 (2:24) 利用AI分类模型甄别咨询来电是新客户还是既有老客户，能有效防止广告投放网络根据虚假转化信号进行错误竞价，避免付费获客预算浪费。 (11:03) 营销团队无需等待完美统一的客户底层数据架构即可展开行动，基于现有数据子集启动紧密契合业务目标的小型试点是建立AI飞轮的关键。 (17:25) 线上数字消费行为能够作为领先的运营指标，例如医院网站上对急诊与等待时间的搜索量变化，提前两周预示了线下就诊高峰的到来。 (23:25)

Enterprise marketing leaders emphasize that generating tangible business value from artificial intelligence requires moving past disconnected experiments to ground predictive models in contextual customer data and defined organizational objectives. Drawing from deployments at Northwell Health, AEG Vision, and Uniphore, the panel argues that companies achieve faster returns by launching focused, incremental pilots that optimize journeys and eliminate spend waste rather than waiting for pristine data architectures. Merging Multi-Channel Engagement and Unstructured Context (2:24) Paul Lambson of Northwell Health explains that clinical records alone present an incomplete portrait of customer needs; synthesizing clinical markers with multi-channel engagement data from websites, CRM systems, and call centers allows analytics teams to determine the precise cadence, channel, and tone needed for effective outreach. Jeremy Lyon of Uniphore observes that while enterprises have digitized structured records, significant operational intelligence remains trapped in unstructured assets such as call center transcripts, standard operating procedures, presentation decks, and campaign retrospectives. Both Lyon and Lambson advise against delaying deployments in pursuit of flawless enterprise warehouses or universal customer identifiers. Teams generate earlier returns by deploying targeted pilots on clean subsets of available data that directly support strategic outcomes, treating AI models as continuous operational assets that require sustained coaching rather than one-time technical launches. Eliminating Marketing Waste Through Operational Attribution (11:03) AI-driven attribution can protect customer acquisition budgets and uncover unexpected conversion pathways. Rachelle Kuebler-Weber describes how AEG Vision, which oversees 509 independent optical practices across 30 states, trained an AI classification model over three months on call-tracking data. The system distinguished existing patients calling for basic office details from prospective customers, passing conversion signals back to advertising networks only for net-new inquiries and preventing paid search algorithms from optimizing around redundant spend. Operational efficiencies can also create secondary acquisition channels. Lyon highlights an insurance initiative where an AI model trained on claims adjudication analyzed unstructured records in real time to guide claims examiners on missing information. While the primary goal was shortening payout cycles, beneficiaries who received swift, transparent service during stressful claims frequently chose to purchase their own personal policies from the carrier, converting a routine back-office cost center into an organic acquisition engine. Predictive Targeting and Behavioral Early Warnings (7:43) Replacing generic outreach with predictive personalization allows organizations to match communications to individual user context. Lambson details how Northwell shifted from broad email campaigns promoting high-acuity procedures like liver transplants toward dynamic Salesforce templates powered by Python scripts that recommend statistically probable care lines, such as orthopedics or cardiology. Over a six-month measurement period, recipients demonstrated measurable increases in multi-service participation. Beyond marketing conversion, external digital signals can inform core operations. During the spring 2020 COVID-19 surge in New York, Northwell's analytics team applied machine learning to Google Analytics traffic, discovering that search behavior shifted approximately two weeks prior to regional hospital volume spikes: traffic for elective treatments dropped off while visits to emergency wait times and parking pages surged. The resulting early warning model was shared across the health system and open-sourced to GitHub, permanently integrating marketing analytics into hospital operational governance. Aligning Martech Execution with Business Objectives (35:50) Panelists caution that AI use cases must function strictly as tactical execution vehicles linked to core key performance indicators and high-level strategy, rather than standalone experiments. Lambson outlines Northwell's operational progression from custom standalone scripts to an integrated native martech stack that provides automated send-time optimization and channel selection, supported by a low-cost waterfall model that escalates from email and SMS to direct mail and call center outreach only as needed. Sustaining organizational buy-in requires rigorous transparency and quantifiable metrics. Kuebler-Weber notes that admitting when bounded experiments fail builds credibility with operators, citing analytics showing clinical practices with a 4.8 Google rating were twice as likely to grow revenue and four times as likely to expand EBITDA compared to those at a 4.6 baseline. Similarly, Alicia Shapiro reports that restructuring AiNews editorial content for LLM ingestion—using clear heading hierarchies, key takeaway bullets, and factual Q&A blocks—doubled article audience reach from 10–20% to 30–50% across AI search engines. While AI enables automated content production and granular segmentation, enterprise implementation faces organizational friction, including swings between data storage expenses and latency, as well as tensions between strict governance and time-to-market. With Northwell assessing its current AI maturity at approximately three out of five, the panel concludes that sustainable advancement relies on transparent experimentation, native martech consolidation, and grounding every initiative in verifiable attribution.

企业营销在引入人工智能时正从被动的静态报表转向深度的业务流运作。来自Northwell Health、AEG Vision与Uniphore的行业专家指出，释放客户数据价值的关键在于打通临床、多渠道交互及非结构化企业语料，破除对“完美数据”的等待迷思，以紧密锚定核心业务KPI的小步快跑试点实现精细化触达、广告浪费过滤及跨部门的运营反哺。 多维数据整合与非结构化语料激活 (2:24) Northwell Health的Paul Lambson指出，医疗机构与支付方传统上过度依赖临床病历来判断患者的就医风险与复诊需求，但该单一视角无法全面反映个体行为。Northwell通过将核心医疗记录与网站浏览行为、呼叫中心记录及CRM触点等多渠道交互数据相融合，不仅能明确患者在医学上何时需要干预，还能判断出触达该患者的最佳沟通频率、语气和渠道，以满足消费者对对标其他商业领域的个性化体验需求。 Uniphore的Jeremy Lyon进一步区分了关系型数据库与非结构化数据的潜能。企业目前虽已擅长整合结构化记录和应用端事件流，但大量关键商业智能仍封存在非结构化资产中，包括客服通话录音与文本、标准作业程序（SOP）、汇报演示文稿以及历史营销复盘记录。Lyon认为，整合这些材料能够构建出反映企业真实经营的语料库，为微调AI模型与优选触达策略提供不可或缺的业务上下文支撑。 细分意图建模与付费广告无效转化过滤 (11:03) 在精细化获客与转化环节，泛化的信息推送与失真的广告归因会造成严重的预算浪费。AEG Vision的Rachelle Kuebler-Weber分享了针对旗下全美509个独立眼科品牌的管理经验：团队发现大量老患者习惯点击付费搜索广告来查找既有医生的电话，导致广告竞价平台误将老客户来电识别为高价值转化并加倍竞价。为此，AEG部署了结合AI分类模型的通话追踪系统，历经约3个月的行为信号训练后在重点州推广，仅当来电确认为净新增患者时才向广告系统回传转化信号，彻底切断了老客带来的虚假正向反馈并遏制了广告浪费。 在患者入组营销中，Paul Lambson提到Northwell曾默认向所有新用户群发复杂的肝移植手术等重疾服务介绍。随后团队运用Python脚本与预测模型分析患者首次就医画像，动态推荐最可能需要的次要服务线（例如40岁男性在统计上更需要初级保健、骨科或心脏科，而非器官移植），在不改变基础触点的情况下动态编排Salesforce模版内容。在为期6个月的监测期内，接收动态推荐的患者在不同科室间的交叉就医比例显著提高。 破除完美数据迷思与管理组织落地摩擦 (17:25) 针对企业在AI部署中的执行阻力，Jeremy Lyon与Paul Lambson警示不要陷入“等待完美数据”或“一劳永逸”的技术陷阱。AI模型上线后如同新员工，需要持续的培训和资源投入；企业无需等待全渠道统一客户档案或无瑕的集中式数据仓库，而应基于可用的局部干净数据集启动微型生产试点。例如Northwell早期采用由低至高成本的阶梯瀑布流策略（邮件、短信、直邮、人工外呼逐级推进），单季度带动数万次预约；随后耗时一年半将散落的独立脚本升级整合为原生MarTech平台，借助内置算法实现发送时机与渠道组合的自动化配置。 在组织治理层面，拥有约11.5万名员工的Northwell面临着存储成本与查询延迟、严格数据治理与业务敏捷性之间的周期性摆动。Lambson主张通过实时营销通话监控看板与严谨的增量归因研究来构建信任。Kuebler-Weber则强调必须对业务线管理层保持完全透明，主动通报未达预期的测试并果断止损；通过清晰展示每一美元预算的流向、具体就诊转化与整体财务回报，换取业务方对AI通话质检等进阶技术方案的支持。 营销信号外溢与企业经营指标的量化关联 (23:25) 营销端收集的用户数字足迹能够外溢转化为关键的临床运营预警。在2020年4月至5月纽约疫情暴发初期，Northwell分析团队将Google Analytics历史数据接入机器学习分类算法，发现大众搜索行为在区域医院确诊激增前约两周发生急剧变化：整形外科或医生介绍等择期需求搜索骤降，而急诊室候诊时间与医院停车指引等突增。该模型转化为全院预警工具并在GitHub开源，证明了外部数字足迹的临床运营价值，使营销分析团队得以正式进入医院数据治理与管理委员会。 在业务增长归因方面，Jeremy Lyon指出Uniphore在保险理赔中利用垂直领域AI知识模型实时分析缺失文件并引导审核员提问，不仅缩短了赔付处理时间，更因高效关怀的解决过程让高满意度赔付人转化为新增保单客户。Kuebler-Weber在此前执掌的临床机构研究中发现，公开的Google评论评分是诊所业务扩张的最强预测指标：保持4.8星评分的诊所相比4.6星者，次年实现营收增长的概率高出2倍，实现EBITDA增长的概率高出4倍，其预测力远超复杂的内部患者指标。 生成式实践与面向LLM的内容检索重构 (33:03) 面对生成式AI，企业应优先寻找能切实消除工作摩擦的落地点。AEG Vision针对509个分支机构人工拍摄视频成本过高的问题，构建结构化内容库结合生成式工具批量生产本地化视频；Northwell Health则为约7000名医生试点AI生成职业头像，消除了医生穿越交通拥堵前往集中拍摄地的通勤痛点，迅速赢得了临床一线人员的认同。 Alicia Shapiro从信息分发角度提出了面向LLM搜索的内容架构改革。由于AI摘要功能导致传统网站搜索流量大幅下跌，AiNews对文章排版进行了工程化改造以适应算法抓取：在文首设置核心要点摘要、规范H1至H3层级标题、加粗特定多词关键短语并在文末嵌入结构化问答模块（Q&A）。这一结构调整使受众触达率从原先的10%–20%提升至30%–50%，并带来了大量由LLM直接引流的高质量新订阅用户。Jeremy Lyon最后强调，所有AI和智能体项目都必须依附于战略与KPI目标，杜绝脱离商业产出的纯技术自嗨。 本场研讨表明，释放客户数据与AI潜力的核心不在于算法复杂性，而在于能否以确定性的财务归因推动小步快跑，并在多渠道交互与非结构化语料中捕捉真实意图；但如何在受严格隐私监管的行业中平衡高度个性化与合规红线，以及如何长期维护生成式资产的品牌一致性，仍是企业在扩展AI能力时有待持续解决的课题。

## Background 本期内容由 Ai4 于 2025 年 8 月 28 日在 YouTube 平台发布，记录了一场题为“AI 驱动的数据营销：从客户数据中挖掘价值”的高管圆桌论坛。视频简介显示，本次研讨会汇集了医疗、视力保健与 AI 战略咨询领域的资深专家，重点探讨营销人员如何利用 AI 理解多维客户数据、识别潜在行为模式，并将预测洞察融入实际业务工作流以实现高效获客与留存。 ## TL;DR 论坛围绕企业如何将 AI 融入客户数据营销展开深入交流，嘉宾们分享了将临床数据与全渠道互动数据融合、借助 AI 区分新老患者电话以避免广告浪费，以及构建自适应动态推荐的实战经验。讨论强调企业不必等待数据“完美”才启动项目，而应通过微小切口快速验证成效，并依靠透明的量化归因体系打破数据孤岛并建立跨部门互信。此外，嘉宾们还探讨了面向生成式大模型的内容重构优化（GEO）、Google 评分对门诊营收的放大效应，以及将 AI 用例自上而下对齐企业商业北极星目标的落地框架。 ## Atmosphere & style 整场圆桌讨论务实而高效，嘉宾们没有停留在宏大的技术概念宣传上，而是紧扣一线业务中的真实痛点与落地细节展开交流。对话风格兼具专业深度与幽默感，无论是关于“抠门式”渠道漏斗设计的自嘲，还是免除医生通勤拍照的生动案例，都让严谨的数据分析充满了人情味。各方交流步调默契、观点层层递进，既有对数字化转型“钟摆效应”的理性审视，又传递出一种鼓励“从小切口入手、边玩边学”的积极探索氛围。 ## Key takeaways - [02:14](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=134s) 将临床诊疗数据与全渠道互动数据融合，能够让医疗机构在明确患者需求的同时，精准预测最佳触达渠道、频率与沟通时机。 - [02:14](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=134s) 非结构化数据（如通话日志、SOP 文档和营销复盘报告）构成了企业真实运营的知识语料库，在微调大模型以制定渠道战略方面极具潜力。 - [07:43](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=463s) 北岸医疗通过轻量级 Python 脚本将 AI 预测的服务条线偏好推送至 CRM，以不着痕迹的动态内容重排显著提升了患者的跨科室就诊参与度。 - [11:03](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=663s) AEG Vision 利用 AI 电话追踪区分拨打电话的是新患者还是老患者，仅将新客通话标记为转化信号回传给广告平台，避免了获客广告预算的浪费。 - [12:53](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=773s) 渐进式患者触达策略按照渠道成本由低到高（邮件 -> 短信 -> 直邮 -> 呼叫中心外呼）推进，只要患者在低成本环节完成预约即可节省昂贵的后续支出。 - [16:04](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=964s) 企业无需等待数据清洗完美无瑕才启动 AI，应先将具体用例与业务目标紧密挂钩，通过小步迭代验证成效并自然形成自增强的飞轮效应。 - [19:14](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1154s) 跨部门推行 AI 的阻力需通过清晰的归因体系和透明度来化解，营销团队开诚布公地披露测试失败反而有助于赢得高管与业务线的长期信任。 - [23:27](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1407s) 疫情初期北岸医疗通过分析医院官网实用页面流量异动提前两周预警确诊激增，使营销数据成功进入医疗管理层的决策与治理体系。 - [28:28](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1708s) 回归分析发现 Google 评分是驱动门诊诊所营收和 EBITDA 增长的核心指标，评分从 4.6 提升至 4.8 能带来成倍的盈利增长。 - [33:03](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1983s) 针对大模型搜索（GEO）优化内容结构（如顶部要点摘要、清晰层级标题与结构化 Q&A），能显著提高内容被大模型抓取引用的概率并带来订阅转化。 - [35:50](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2150s) AI 用例必须严格沿着“企业商业目标 -> 战略与 KPI -> 落地用例”的层级向上对齐，避免沦为脱离实际业务的科学实验。 - [38:49](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2329s) 嘉宾一致建议初涉 AI 营销应从小处着手（Start small），即便数据不完美也可以立即开始实践并在实际操作中摸索迭代。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=4s) 论坛开场与嘉宾介绍 论坛主持人、AiNews.com 首席营销官 Alicia Shapiro 向观众致开场白，宣布本次圆桌讨论的主题为“AI 驱动的数据营销：从客户数据中挖掘价值”。Shapiro 指出，如今的营销人员拥有前所未有的海量客户数据，但真正的挑战与机遇在于如何理解并利用这些数据。随着人工智能的介入，营销数据分析已经超越了传统的静态仪表板和报表，转向能够识别深层模式、预测用户行为并实现超个性化体验的智能工具。本次讨论的核心即是将技术潜力转化为业务实践，探讨如何借助 AI 提取有价值的洞察，将其融入工作流，从而有效提升获客与客户留存率。 随后，主持人介绍了本场圆桌的重磅嘉宾： - **Rachelle Kuebler-Weber**：AEG Vision 首席营销官（CMO），在医疗健康及视力保健领域拥有丰富的营销团队领导经验，擅长通过数据驱动策略推动企业规模化增长与价值提升。 - **Paul Lambson**：美国大型医疗系统北岸医疗（Northwell Health）客户分析助理副总裁（AVP），专注于运用高级分析与数据策略帮助医疗机构理解、互动并留存患者。 - **Jeremy Lyon**：Uniphore 战略 AI 咨询总监，与企业高管团队紧密合作制定并落地企业级 AI 战略，致力于打通数据、技术与业务目标以实现可衡量的成果。 --- ### [02:14](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=134s) 临床与全渠道互动数据的融合价值 在被问及当前工作中哪些类型的客户数据最具价值时，Paul Lambson 首先从医疗服务机构的视角分享了核心洞察。他强调，将临床数据（Clinical Data）与互动数据（Engagement Data）结合能够释放出极大的协同价值。传统医疗机构和保险公司往往高度聚焦于“流失患者”的临床风险指标，但仅凭临床指南无法解决患者沟通的断层。北岸医疗通过将网站浏览、呼叫中心记录、CRM 等全渠道互动数据与临床指引层层叠加，构建出预测模型：不仅明确患者需要接受何种诊疗，更能精准判断应当通过何种渠道、何种沟通频率以及何种方式在合适的时间节点触达患者，避免冷冰冰的说教。Lambson 补充道，现代消费者的预期已经发生了转变，他们希望医疗机构的触达策略能像非医疗行业的成熟体验一样，充分体现对其已知喜好和历史互动的理解。 ```mermaid flowchart TD accTitle: 医疗数据融合与触达策略流程图 accDescription: 展示临床数据与互动数据融合后，通过 AI 模型生成个性化触达策略的过程。 subgraph DataSources ["多维数据源融合"] A["临床诊疗数据<br/>(病历、就诊记录、脱落风险)"] B["全渠道互动数据<br/>(网站行为、呼叫中心记录、CRM)"] end subgraph AIModeling ["AI 建模与上下文分析"] C["AI 预测与推荐模型"] D["非结构化数据提炼<br/>(SOP、复盘报告、通话录音)"] end subgraph ActionableOutput ["精准触达执行"] E["个性化服务推荐<br/>(骨科 / 心脏 / 全科)"] F["动态渠道与节奏匹配<br/>(短信 / 邮件 / 最佳时间)"] end A --> C B --> C D --> C C --> E C --> F ``` 随后，Karim Touma 补充了非结构化数据在现代营销中的关键作用。他指出，尽管数据团队在关系型数据库中集中结构化数据，以及整合来自网站与移动应用的半结构化数据方面取得了长足进展，但蕴藏在非结构化数据中的价值同样不可忽视。这包括客户沟通速记、通话日志、标准作业程序（SOP）文档、演示文稿乃至历史营销复盘报告。这些海量资料构成了“企业真实运营的知识语料库（corpus of your business's reality）”，在微调大模型以精准理解营销目标、制定渠道战略方面展现出了巨大的应用潜力。 --- ### [04:51](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=291s) 多品牌运营、合规挑战与主动式患者体验 Rachelle Kuebler-Weber 结合 AEG Vision 的复杂业务场景阐述了数据整合的挑战。AEG Vision 旗下拥有遍布全美的 509 家视光诊所，且每一家诊所均采用独立品牌运营，相当于营销团队需要同时管理 509 个独立品牌。视光行业兼具医疗与高度零售属性，因此 AEG Vision 的数据策略是将患者临床数据、诊所一线运营数据、营销平台表现数据以及第三方数据紧密串联，完整监控从潜在患者的第一触点、转化就诊到诊所内消费的全流程吞吐量，从而向目标人群精准投放匹配的医疗产品与服务。 ``` AEG Vision 业务特征与数据全景： • 运营规模：覆盖全美 30 个州，拥有 509 家独立品牌视光诊所 • 数据交织：患者数据 + 诊所运营数据 + 营销漏斗数据 + 第三方人口属性数据 • 核心诉求：实现“从潜客初次触达 -> 转化为患者 -> 诊所内消费”的全流程数据闭环 ``` 针对 Paul Lambson 提出的关于在“一对一营销”层面使用第三方及理赔数据是否会遭遇组织阻力的问题，Kuebler-Weber 坦言，在 30 个州展业必须严格遵循 HIPAA 等法律法规。企业之所以能推行该战略，得益于卓越的法务团队支持以及高管层在合规与业务创新之间达成的文化共识。谈及个性化信息是否会引起患者反感时，Kuebler-Weber 引用了核心法则：“不要让我把一句话重复说两次（Don't make me tell you twice）”。她指出，当患者提供了个人信息后，就天然期望品牌能够预判其需求；在患者尚未意识到自身需求之前，于其偏好的渠道精准送达复查提醒，正是让患者感受到被关怀的“甜蜜点（sweet spot）”。 --- ### [07:43](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=463s) 北岸医疗的个性化服务推荐与意图捕捉 Paul Lambson 分享了一个生动的实际案例：过去在面向新就诊患者发送“欢迎系列邮件（Welcome Series）”时，医疗机构往往盲目推送高客单价服务，“我们过去总是在内容里硬塞进顶尖科室介绍，比如宣传‘看我们做肝移植手术多厉害’，但现实中有谁会随便在市场上物色肝移植服务呢？”这种脱节的沟通促使北岸医疗引入 AI 模型。系统通过分析患者的初次就诊记录，预测其后续最可能需要的服务条线。例如针对一名 40 岁超重男性患者，模型会推断其潜在需求集中在骨科（Ortho）、心脏科（Cardio）和全科门诊（PCP）。 在落地执行上，团队并未大动干戈重构内容库，而是采用了一种轻量且渐进的技术路径： - **极简起步**：初期仅依靠一个 Python 脚本将个性化排序的数据扩展字段（Data Extension）推送至 Salesforce CRM。 - **隐形个性化**：复用现有的邮件模板与内容库，但根据模型给出的服务条线推荐分数为每个患者动态重排内容区块，不刻意向患者强调“这是为您定制的”，保持触达的自然与低调。 - **业务验证**：运行 6 个月后，数据团队创建了名为“独立服务条线参与度（distinct number of service line participation）”的指标来衡量钱包份额（share of wallet）。结果表明，接收个性化动态推荐的患者群体，其跨科室就诊参与度显著高于接收通用肝移植等硬推内容的对照组。 在谈到如何利用 AI 获取新客户时，Jeremy Lyon 概述了获客逻辑的演进。传统获客依赖于捕捉网站与 App 上的零散意向信号，并将多次会话的多维信号归因聚合至单一标识符（Identifier）上进行预测；而如今引入 AI 后，企业能够快速消化非结构化数据与历史营销复盘资产，将过去需要耗费数周甚至数月才能完成的洞察提炼过程大幅缩短。 ### [11:03](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=663s) 利用历史数据与实时信号实现精准营销 在讨论如何利用模型提炼关键信息时，Jeremy Lyon 指出，现代 AI 模型已经具备了对大量历史数据进行推理（Inference）的能力。企业可以将过去五年内针对不同客群画像和目标受众的营销战役成效，与当前所有新进入的实时信号相结合，从而生成高度精准、高转化率且覆盖面广的触达信息。这种高效能的精准营销在两年前几乎是无法实现的，如今在医疗健康、保险以及各大行业都展现出了巨大的应用潜力。 针对这一趋势，AEG Vision 的首席营销官 Rachelle Kuebler-Weber 分享了一个极其生动的落地案例。当她刚加入 AEG Vision 时，发现公司的大部分营销预算本意是获取“新患者”，但实际转化而来的却大多是“老患者”（Established Patients）。由于官网上的诊所信息非常容易被检索到，许多老患者看到付费搜索广告后便顺手点击并给自己的医生打电话——这白白消耗了针对获客投放的广告预算。 为解决这一痛点，AEG Vision 引入了电话追踪技术和基于 AI 的识别系统，利用各种信号来区分拨打电话的是新患者还是老患者： - **模型训练与试点**：团队对 AI 模型进行了大约 3 个月的训练，并极其谨慎地率先在几个核心州进行了小范围部署。 - **转化信号回传过滤**：只有当系统确认通话来自新患者时，才会向 Google 等广告平台回传“转化成功”的信号；老患者的通话数据则被直接过滤。 - **成效**：老患者依然可以通过广告找到诊所并拨打电话，但广告算法（如 Google Ads）不会再将老患者的点击误判为“优质转化（You did good, kid）”，从而将预算精准引导至真正的新客获取上。 --- ### [12:53](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=773s) 医疗体系的渐进式触达与原生平台优化 Northwell Health 的客户分析副总裁 Paul Lambson 详细介绍了他们在疫情后通过渐进式触达策略吸引流失患者（Lapsed Patients）回归诊所的实践经验。面对“让患者重新坐回诊室（putting butts in seats）”的挑战，Paul 团队采取了一种按渠道成本从低到高的策略： ```mermaid flowchart TD accTitle: 渐进式低成本患者触达漏斗 A["流失患者名单"] --> B["1. 电子邮件 (最低成本)"] B -- "已预约" --> Z["终止触达 / 节省成本"] B -- "未预约" --> C["2. 短信推送 (SMS)"] C -- "已预约" --> Z C -- "未预约" --> D["3. 直邮信件 (Direct Mail)"] D -- "已预约" --> Z D -- "未预约" --> E["4. 呼叫中心外呼 (最高成本)"] E --> Z ``` Paul 幽默地解释道，这一策略的核心逻辑在于“我很抠门（I'm cheap）”——只要患者在收到第一封低成本邮件后完成了预约，就能为医疗系统节省后续所有昂贵渠道的支出。该项目落地后反响极其热烈，单个季度就能为 Northwell Health 带来数万次的就诊预约量。 在这个高成效项目的基础上，团队进一步思考如何利用 AI 对其进行深度优化： - **架构演进**：此前团队仅依靠简易的 Python 脚本，上传至云函数（Cloud Function）并推送到数据扩展包中，这种方式十分简陋。 - **拥抱原生能力**：团队花费了一年半的时间推动 MarTech 平台整合，并培训员工全面利用平台内建的原生 AI 模型，直接启用最佳渠道推荐（Channel Optimization）、最佳发送时间预测（Right Time Send）以及渠道组合优化（Mix Optimization）。 - **耐心沉淀**：团队在推进过程中坚决遏制了“为了求快而仓促上线”的冲动，虽然前期标准化建设较慢，但为后续的敏捷优化和规模化落地打下了坚实基础。 --- ### [15:12](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=912s) 小切口见大成效：微小 AI 功能的巨大业务价值 在谈到规模化应用 AI 时，Rachelle 坦诚地表示，目前大多数企业其实都还处于“涉水摸索阶段（dipping our toes in the water）”，AEG Vision 也不例外。然而，不必一开始就追求翻天覆地的大型变革，很多看似基础、微小的工作往往能带来巨大的业务价值。 她特别提到了现场参会的公司 IT 负责人 Satic 所做的一项改进：利用计算机视觉让患者直接拍照上传医保卡，系统自动提取识别信息并填充至后台系统。这项功能免去了患者手动输入繁琐医保信息的痛苦，不仅极大地提升了患者体验和建档效率，也降低了前台录入错误率。Rachelle 强调，AI 的落地“不一定非要是具有划时代意义的庞大集成，哪怕只是微小的杠杆，也能撬动可观的价值”。 --- ### [16:04](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=964s) 营销 AI 框架：决策、行动与飞轮效应 Uniphore 的战略 AI 咨询总监 Jeremy Lyon 分享了咨询体系中经过实战验证的 AI 营销框架。他指出，营销人员的核心在于在正确的业务背景下将 AI 用例与业务结果相匹配： ```mermaid flowchart LR accTitle: 营销 AI 决策与优化飞轮 subgraph Flywheel ["AI 营销闭环"] direction LR C["1. 语境化数据 (Context)"] --> D["2. 智能决策 (Decisions)"] D --> M["3. 渠道/Agent 动员 (Mobilize)"] M --> O["4. 效果迭代优化 (Optimize)"] O --> C end ``` 针对大型企业往往误将 AI 项目视为“一劳永逸（one and done）”工程的普遍担忧，Jeremy 强调：AI 更像是一个新入职的员工，需要企业持续对其进行培训、引导、辅导和资源投入。 同时，Jeremy 澄清了两个长期存在的认知误区： - **无需追求“完美数据”**：企业在启动 AI 项目之前，不需要等待所有数据都清洗得完美无瑕，也不需要将所有孤岛数据统统汇聚到一个唯一的客户标识（Persona/ID）下，完全可以从某一部分数据切入并逐步迭代。 - **飞轮是结果而非起点**：企业的首要目标不应该是“构建一个抽象的飞轮”，而应当是将 AI 用例与具体业务目标紧密挂钩，先将不脱离实际业务的试点项目推向生产环境并验证成效，随后再将成功的试点向不同业务职能横向复制，以此自然形成自增强的飞轮效应。 --- ### [19:14](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1154s) 突破数据孤岛：从建立归因信任开始 在讨论跨部门数据孤岛与复杂组织架构下的 AI 落地阻力时，Rachelle 指出，市面上所有的营销技术供应商都会宣称自己能解决所有业务问题，但他们往往严重低估了将“患者全链路漏斗”与“企业实际运营数据”深度融合的复杂性。 因此，建立跨部门信任的最佳路径是**从小处着手，建立清晰明确的归因体系（Attribution）**： - **可量化的商业闭环**：能够清晰地向业务利益相关者证明“我们在某处投入了 1 美元，这笔钱具体流向了哪个门诊、带来了什么类型的患者、该患者产生了多少消费，以及后续展现出了怎样的就诊行为模式”。 - **降低推行阻力**：一旦通过扎实的归因建立了业务层面的信任，后续再推行更高级的技术（如 AI 智能通话追踪、呼叫中心 AI 质检）时，来自内部的阻力就会大幅减小，各方的买账度、支持度以及对诊所业务流程调整的容忍度都会显著提高。 Paul 对此深表赞同，他补充道，将 AI 用例推向真实市场的核心价值不仅在于做一个“Hello World”式的技术通路验证，更在于通过严密的实验设计来量化业务归因，从而稳步提升企业的 AI 成熟度。 ### [21:27](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1287s) 数字化转型的钟摆效应与跨部门互信构建 大型企业的数字化转型往往伴随着庞大的资源投入和层出不穷的分歧意见。Paul Lambson 形象地将其比作一个永无休止的“钟摆”：当数据存储成本上升时，组织会全力推进行政控制以压缩开销；而一旦存储成本下降，系统的数据延迟（latency）问题又会暴露出来；随后，数据治理部门介入以降低合规风险，但这又不可避免地拉长了业务的上市周期（go-to-market）。在 Northwell Health 这样拥有 11.5 万名员工的庞大机构中，要推动各方协作并调转庞大的组织巨轮，最核心的依托在于同各级业务主管建立扎实的人际信任。 为证明业务价值，团队不能仅停留在静态图表上，而是拿出了硬核的数据指标——例如能够实时展示营销带动呼叫量转化的监控看板，以及证明营销增量投入如何直接拉动患者预约门诊的大型归因研究。Rachelle Kuebler-Weber 补充强调，建立信任的关键还在于“敢于承认失败”。如果某项测试彻底失败，营销团队必须开诚布公地向业务运营负责人说明：“我们花掉了你们的预算，但测试彻底失败了，因此我们已终止该项目。”由于营销人员通常擅长讲故事和推销，过度乐观往往会损耗信任资产；唯有全面摊开账本，坦诚展示哪些有效、哪些待定、哪些绝不再碰，才能在跨部门之间建立起长久牢固的信任关系。 ```mermaid flowchart TD accTitle: 数字化转型中的组织权衡钟摆 subgraph Storage["存储与性能权衡"] S1["存储成本上升"] --> S2["压缩成本策略"] S2 --> S3["存储成本下降"] S3 --> S4["出现延迟 (Latency) 瓶颈"] end subgraph Governance["治理与速度权衡"] G1["引入数据治理"] --> G2["降低合规与数据风险"] G2 --> G3["业务上市周期 (GTM) 放缓"] G3 --> G4["业务要求加速并重新反弹"] end Storage <--> Governance ``` - **组织规模与沟通阻力**：在拥有 115,000 名员工的医疗系统中，协调技术、合规与业务各方意见需依赖深度的人际协作与硬核量化指标。 - **透明度原则**：营销部门主动披露测试失败案例，通过小规模试错控制风险，以此维持高管与业务线对创新项目的长期信心。 --- ### [23:27](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1407s) 意想不到的洞察：医疗搜索预测疫情与保险理赔逆向获客 主持人 Alicia Shapiro 提问，生成式 AI 或机器学习曾帮助各家企业挖掘出哪些令人意想不到的客户洞察。Paul Lambson 分享了一个发生在 2020 年春季（4 至 5 月）新冠疫情初期的案例。当时纽约市受到疫情重创，一位被紧急调往儿科急诊前线的健康信息学主管找到他，询问能否像 Google 预测流感一样，提前预警新冠疫情的暴发。团队利用当时的 Google Analytics 网站行为数据，结合机器学习算法进行聚类与趋势挖掘，并与区域内的临床确诊病例进行交叉对比。 结果发现，在社区确诊病例激增前的约两周，公众访问医院网站的行为模式会出现剧烈反常：常规的整形外科（plastic surgery）、保险查询以及医生个人履历浏览量骤降；相反，急诊等待时间、医院停车位等实用页面访问量出现暴涨。这项被称为“黑盒”挖掘出的营销数据指标，随后被转化为医院日常运营的预警信号，甚至开源发布至 GitHub，并引来全美多套医疗系统的连线交流。这次经历不仅挽救了生命，更彻底扭转了由专业医师主导的医院管理层对营销数据的认知偏见，让营销团队正式进驻企业数据治理与决策委员会。 Jeremy Lyon 则分享了保险理赔裁定（claims adjudication）场景中的意外发现。传统的理赔流程对受益人、理赔员和审查员来说都极其漫长痛苦，收集齐核赔所需的非结构化文档往往需要耗费数周。利用理赔领域的专业知识库大模型，AI 能够快速识别缺漏材料，并为理赔员生成针对受益人的精准询问清单。这不仅大幅缩短了赔付周期，还促成了一个意想不到的“二次获客”闭环：许多原本仅作为被动受益人的客户，因在全流程中获得了极佳的理赔服务体验，在后续自身产生投保需求时，会主动选择该保险公司，将成本中心的理赔流程直接转化为了低成本获客渠道。 --- ### [28:28](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1708s) 诊所增长的核心密码与规模化 AI 内容生成 Rachelle Kuebler-Weber 分享了她在前一家公司通过多变量模型分析得出的颠覆性结论。当时团队将净推荐值（NPS）、产品组合、销售额、患者构成以及 Google 评分（Google Score）等大量数据全部输入模型进行回归分析，试图寻找驱动门诊增长的核心动力。令人震惊的是，预测收入增长以及 EBITDA（税息折旧及摊销前利润）增长的唯一决定性指标既不是 NPS 也不是客群组合，而是 Google 评分。 数据对比显示，Google 评分达到 4.8 分的门诊诊所，其年营收增长概率是 4.6 分诊所的两倍；而在 EBITDA 维度的增长倍数更是高达 4 倍。这一发现证明了微小的线上声誉差异对实体门诊盈利能力的非线性放大效应。 ```mermaid flowchart LR accTitle: Google 评分对医疗诊所业务增长的驱动关系 subgraph Inputs["输入评估因子"] I1["净推荐值 (NPS)"] I2["患者及产品组合"] I3["销售体量数据"] I4["Google 评分"] end subgraph CoreDriver["核心驱动引擎"] CD["模型筛选唯一关键变量:<br/>Google Score"] end subgraph BusinessOutcomes["业务增长回报 (4.8分 vs 4.6分)"] O1["年营收增长概率: 提升 2 倍"] O2["EBITDA 盈利增长: 提升 4 倍"] end Inputs --> CD CD --> BusinessOutcomes ``` 随后话题转向 AI 驱动的内容策略与个性化实践。Rachelle 谈到，AEG Vision 旗下拥有 509 家视光诊所，依靠传统人力根本不可能为每家诊所定制高频的本地化视频内容。目前团队正致力于搭建基础内容库，并根据患者人口统计特征对 509 家诊所进行聚类分群，计划全面引入 AI 工具生成视频。她特别提到近期行业内的一个惊人案例：一家公司利用 AI 工具链，仅凭 1 名人力在 1 小时内便制作出一条达到可直接上电视播出标准的商业广告（ready-for-TV commercial）。 Jeremy Lyon 进一步阐述了 AI 在人群洞察方面的革新。在传统营销流程中，营销人员依赖个人直觉与过滤条件来圈定重叠客群和受众细分（cohorts）；而 AI 能够“看清视线死角”，主动挖掘出人类直觉未曾设想过的 5 个全新细分人群或战略提问。这种跳出固有框架的能力，直接为营销团队解锁了转化率更高、获客成本可控的全新获客战役。 - **评分倍增效应**：Google 评分从 4.6 分提升至 4.8 分，诊所的营收增长可能性提升 2 倍，EBITDA 增长空间扩大至 4 倍。 - **内容生产极效化**：行业前沿实践展现了“1 人、1 小时制作出电视级商业广告”的生产力跨越，使多分支机构的矩阵式视频分发成为可能。 - **受众洞察升维**：AI 超越人类营销人员预设的交叉人群交集，能够自主推演出数个未被发掘的高转化潜力子人群。 ### [32:11](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1931s) 解决痛点的小切口：用 AI 证件照消除医生的通勤负担 在谈及 AI 如何充当思考伙伴与实验助手时，Paul Lambson 分享了 Northwell Health 正在推进的一个极具现实意义的试点项目。该医疗系统拥有约 7,000 名在职雇员医生以及大量志愿医生，过去医院总是要求医生们亲自驾车前往长岛（Long Island）拍摄标准职业证件照。对于工作极为繁忙的医生而言，在拥堵的高速公路上耗费数小时去拍照是一件极不受欢迎的差事。许多医生抱怨道：“我明明知道可以用 AI 生成证件照，为什么非要逼我开车？” Paul Lambson 幽默地指出，如果营销团队今年能顺利完成这个 AI 头像生成试点，在未来五年里这很可能成为营销部门最受医生赞誉的政绩——“除掉营销团队做过的所有其他事情，最让医生满意的恰恰是我们从他们的生活中拿掉了一次不必要的通勤”。这个例子生动地说明，有时 AI 最大的价值并不在于多么高深的算法重构，而在于切实解决用户日常中的具体痛点。 - **组织规模与痛点**：覆盖 7,000 多名在职医生及大批志愿医生，彻底免去了专程前往长岛拍摄职业证件照的交通成本。 - **内部影响力**：通过微小但感知极强的 AI 工具试点，极大地提升了临床医生对营销与数字化团队的好感度。 --- ### [33:03](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1983s) 面向大模型的新闻重构：AiNews 的生成式引擎优化（GEO）实践 主持人 Alicia Shapiro 分享了自己在 AiNews.com 的亲身探索。传统新闻写作讲究为人类读者营造故事弧光（narrative arc）和沉浸式阅读体验，但随着 ChatGPT、Perplexity、Gemini 以及 Google AI Overviews 等生成式引擎的普及，整个搜索生态发生了剧变。大模型并不会通读整篇娓娓道来的长文，它们只是高速扫描互联网内容，精准抓取少数核心事实、数据与定义；与此同时，随着 AI 直接给出总结答案，用户不再下滑点击网页链接，导致传统网站的访问流量普遍腰斩。 为了应对这一危机，Alicia Shapiro 直接向 ChatGPT 请教如何让新闻内容更易被大模型发现与引用。根据建议，AiNews 全面调整了文章结构：保留清晰的主副标题层级、用粗体高亮核心长尾短语（如 "data-driven marketing"）、在文首放置核心观点摘要列表（key takeaway bullets），并在文末附上结构化的 Q&A 问答板块。这套专为大模型“可发现性”设计的版式改造带来了显著回报，平台的内容触达率（reach）直接翻倍，受众覆盖范围从原先的 10%~20% 跃升至 30%~50%，并持续从 ChatGPT 和 Perplexity 等 AI 平台源源不断地获得新订阅用户。 ```mermaid flowchart TD accTitle: 面向大模型优化的内容结构 subgraph Traditional["传统新闻叙事模式"] T1["线性故事铺垫与叙事弧光"] --> T2["长篇叙述散落数据"] T2 --> T3["大模型抓取效率低 / 零点击导致流量腰斩"] end subgraph LLMOptimized["面向大模型（GEO）的重构模式"] L1["顶部核心观点摘要 (Key Takeaways)"] L2["清晰层级标题 + 粗体核心词组"] L3["底部结构化问答板块 (Q&A)"] end subgraph Outcome["成效与转化"] O1["大模型快速解析事实与引用"] --> O2["内容触达率翻倍 (最高达 50%)"] O2 --> O3["持续从 ChatGPT / Perplexity 获取新订阅"] end LLMOptimized --> Outcome ``` --- ### [35:50](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2150s) 避免沦为“科学实验”：自上而下的 AI 用例对齐框架 针对营销团队如何挑选正确 AI 用例的提问，Jeremy Lyon 强调必须将所有用例牢牢锚定在企业的“北极星目标”（North Star objectives）与商业结果上。他直言，如果用例脱离了业务目标，团队最终只会是在消耗宝贵的时间与资源做无意义的“科学实验（science projects）”。 Jeremy Lyon 提出了一个自上而下的金字塔对齐框架：企业商业目标居于顶层，由此派生出达成目标所需的战略；为了衡量战略落地效果而设定具体的 KPI；而 AI 或智能体（Agentic AI）用例，本质上只是帮助团队达成 KPI 和业务指标的“载体与工具”。他建议团队在迷茫时随时“沿着梯子往上倒推”，审视当前所做的 AI 用例是否真正服务于核心业务结果。 ```mermaid flowchart TD accTitle: AI 用例与业务战略对齐层级 subgraph Hierarchy["战略对齐阶梯 (Hierarchy Framework)"] A["顶层：企业商业目标 (Business Objectives)"] B["中层：达成战略与关键指标 (Strategy & KPIs)"] C["底层：AI / 智能体实施用例 (AI & Agentic Use Cases)"] end A --> B B --> C C -.->|"自下而上校验：是否支撑 KPI 与商业目标？"| A ``` - **防范资源浪费**：脱离商业目标的 AI 尝试极易变成自嗨型的孤立项目。 - **层级追溯法则**：无论技术形态如何演进（从传统分析到 Agentic AI），用例必须经得起向 KPI、战略与业务成果的层层验证。 --- ### [37:06](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2226s) 从单一版本到微观定向：赋能营销团队的 A/B 实验思维 谈及个性化落地，Paul Lambson 阐述了 Northwell Health 在营销技术栈（MarTech stack）上的具体举措。作为技术底座的负责人，其明确目标在于提升打开率与点击转化率（CTR）。为此，他致力于为一线协调员和总监解绑，移走必须立刻产出“唯一最终版本”的交付压力，并在排期中预留充足的时间，鼓励大家针对研发过程中产生的假设进行对照验证。 这一实验思维同样延伸到了 UX 设计团队。借助升级后的 A/B 测试工具，团队不再强行将所有前期的用户洞察压缩成单一折中方案，而是可以直接将开发过程中考量过的两到三种版本推向市场进行真实验证。Paul Lambson 指出，这要求团队完成一次深刻的思维转变：从过去受限于边界、只能产出一个版本的惯性，转变为运用聚类分析划分 7 个左右的客群，并最终迈向更精细的“微观定向（micro-targeting）”多版本并行实验。 - **减压与留白**：在生产流程中为团队预留测试时间，鼓励将开发阶段的推测与分歧直接转化为上线实验。 - **分群演进路径**：推动营销与产品人员从“单一版本交付”思维，跨越到“经典 7 客群划分”，进而深入到“微观定向实验”。 --- ### [38:49](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2329s) 快问快答与总结：从小处着手，在不完美的数据中边玩边学 在讨论尾声的闪电问答环节，主持人邀请各位嘉宾用一句话总结给初涉 AI 客户数据营销者的建议： - **Rachelle Kuebler-Weber**：坚定地建议“从小处着手（Start small）”，优先选择那些能够快速见效的容易用例，让团队和整个组织快速积累实战经验与手感（reps）。 - **Jeremy Lyon**：“数据不必等到完美才能取得进展（Your data does not need to be perfect to make progress）。”只要从小做起，即刻启动并在过程中持续学习。 - **Alicia Shapiro**：补充强调了“保持玩心（Start playing）”的重要性。直接动手尝试各种 AI 工具、向不同的大模型提问互动，往往能碰撞出未曾预料的全新创意。 最后，Alicia Shapiro 代表主办方感谢了全体嘉宾与观众的参与，鼓励大家放下顾虑、积极动手探索 AI 在客户数据中的潜能。 ## Notable quotes > [02:14](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=134s) 企业真实运营的知识语料库（corpus of your business's reality） — Karim Touma > [04:51](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=291s) 不要让我把一句话重复说两次（Don't make me tell you twice） — Rachelle Kuebler-Weber > [07:43](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=463s) 我们过去总是在内容里硬塞进顶尖科室介绍，比如宣传‘看我们做肝移植手术多厉害’，但现实中有谁会随便在市场上物色肝移植服务呢？ — Paul Lambson > [12:53](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=773s) 我很抠门（I'm cheap） — Paul Lambson > [15:12](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=912s) 涉水摸索阶段（dipping our toes in the water） — Rachelle Kuebler-Weber > [15:12](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=912s) 不一定非要是具有划时代意义的庞大集成，哪怕只是微小的杠杆，也能撬动可观的价值 — Rachelle Kuebler-Weber > [21:27](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1287s) 我们花掉了你们的预算，但测试彻底失败了，因此我们已终止该项目 — Rachelle Kuebler-Weber > [32:11](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=1931s) 除掉营销团队做过的所有其他事情，最让医生满意的恰恰是我们从他们的生活中拿掉了一次不必要的通勤 — Paul Lambson > [35:50](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2150s) 科学实验（science projects） — Jeremy Lyon > [38:49](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2329s) 从小处着手（Start small） — Rachelle Kuebler-Weber > [38:49](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2329s) 数据不必等到完美才能取得进展（Your data does not need to be perfect to make progress） — Jeremy Lyon > [38:49](https://www.youtube.com/watch?v=ZUxG3tWoe9w&t=2329s) 保持玩心（Start playing） — Alicia Shapiro ## Mentioned Alicia Shapiro, Rachelle Kuebler-Weber, Karim Touma, Paul Lambson, Jeremy Lyon, Ai4, AiNews.com, AEG Vision, DeAcero, Northwell Health, Uniphore, Salesforce CRM, Google, Google Ads, Google Analytics, Google AI Overviews, ChatGPT, Perplexity, Gemini, GitHub, Python, HIPAA ## Video description Alicia Shapiro, CMO, Head of News Reporting, AiNews.com Rachelle Kuebler-Weber, Chief Marketing Officer, AEG Vision Karim Touma, Group Chief Digital and Data Officer, DeAcero Paul Lambson, AVP of Customer Analytics, Northwell Health Jeremy Lyon, Director, Strategic AI Consulting, Uniphore Abstracts This session will focus on how marketers can use AI to make sense of customer data and turn insights into actionable strategies to more effectively acquire customers. Panelists will discuss the types of customer data that hold the most value, how to identify and access new data sources, and the role of AI in analyzing this information. Learn how AI can enhance your understanding of customer behavior, uncover hidden patterns, and generate predictive insights that drive better marketing decisions. The discussion will also cover practical approaches for integrating AI tools into existing data workflows, helping marketers extract maximum value from their data and create data-driven campaigns that attract and retain new customers. [Source](https://www.youtube.com/watch?v=ZUxG3tWoe9w)

[2:14] Karim Touma: 企业真实运营的知识语料库（corpus of your business's reality） [4:51] Rachelle Kuebler-Weber: 不要让我把一句话重复说两次（Don't make me tell you twice） [7:43] Paul Lambson: 我们过去总是在内容里硬塞进顶尖科室介绍，比如宣传‘看我们做肝移植手术多厉害’，但现实中有谁会随便在市场上物色肝移植服务呢？ [12:53] Paul Lambson: 我很抠门（I'm cheap） [15:12] Rachelle Kuebler-Weber: 涉水摸索阶段（dipping our toes in the water） [15:12] Rachelle Kuebler-Weber: 不一定非要是具有划时代意义的庞大集成，哪怕只是微小的杠杆，也能撬动可观的价值 [21:27] Rachelle Kuebler-Weber: 我们花掉了你们的预算，但测试彻底失败了，因此我们已终止该项目 [32:11] Paul Lambson: 除掉营销团队做过的所有其他事情，最让医生满意的恰恰是我们从他们的生活中拿掉了一次不必要的通勤 [35:50] Jeremy Lyon: 科学实验（science projects） [38:49] Rachelle Kuebler-Weber: 从小处着手（Start small） [38:49] Jeremy Lyon: 数据不必等到完美才能取得进展（Your data does not need to be perfect to make progress） [38:49] Alicia Shapiro: 保持玩心（Start playing）

[4:51] Rachelle Kuebler-Weber: Don't make me tell you twice. [7:43] Paul Lambson: Hey, look at us. We're really good at liver transplants, but anyone in the market for a liver? [12:53] Paul Lambson: slapping wrists when people are quick to put something to market [21:27] Rachelle Kuebler-Weber: We have something we tested; it has failed miserably. We won't be moving forward with that particular project. And being able to say to my operators, 'Hey, we spent your money. It didn't work. It was a test. It was limited. There's a reason it was limited.' I think being able to show both sides of the coin... goes a long way. [23:25] Chief Technology Officer in Health Informatics: Paul, I know that Google can spot flu. Can we spot this, please? [32:11] Paul Lambson: If we get this done this year... this will be what we're known for for like five years is marketing made a headshot AI. [35:50] Jeremy Lyon: walk that ladder back up [38:48] Rachelle Kuebler-Weber: Start small. Start with something that's an easy win that gets your team a few reps, that gets your organization a few reps. [38:48] Jeremy Lyon: Your data does not need to be perfect to make progress. Just get started. Get started small. Learn. [38:48] Alicia Shapiro & Paul Lambson: start playing

[0:04] Hello everybody. Thanks for joining us [0:07] today and welcome to today's panel of [0:09] datadriven marketing and AI. We're going [0:11] to um talk about extracting value from [0:13] customer data. My name is Alicia [0:16] Shapiro. I'm the CMO of a newsw.com. You [0:19] can subscribe to our newsletter on the [0:21] website and I'm so thrilled to be here [0:23] today to help moderating this panel. [0:25] It's very exciting. [0:27] So to get started as marketers, we have [0:31] never had access to more customer data [0:34] than we do right now. And the challenge [0:37] and perhaps an opportunity is how we [0:41] make sense of that data. Right? So with [0:43] AI, we're not just talking about [0:45] dashboards and reports anymore. We're [0:48] talking about the tools that can [0:49] identify patterns, predict behavior, and [0:52] personalize experiences in ways that [0:54] were just impossible a few years ago. [0:57] Today's conversation is going to be [0:59] about turning that potential into [1:01] practice. How to use AI to extract [1:04] meaningful insights, integrate them into [1:07] workflows, and ultimately drive customer [1:09] acquisition and retention. I'm joined [1:12] today by some ex uh excellent panelists [1:16] here whose work spans multiple [1:18] industries and approaches in AI and [1:20] marketing. First is Michelle Kubler, [1:23] chief marketing officer at AEG Vision. [1:27] Michelle has built and led [1:28] high-erforming marketing organizations [1:30] across healthcare and vision care, [1:32] consistently delivering datadriven [1:34] strategies that scale growth and [1:36] enterprise value. [1:38] Paul Lamson, AVP of customer analytics [1:41] of at Northwell Health. Paul applies [1:44] advanced analytics to help one of the [1:46] largest healthcare providers in the US [1:48] to understand, engage, and retain [1:51] patients through datainformed [1:52] strategies. And Jeremy Lion, director of [1:56] strategic AI consulting at Unifor. [1:59] Jeremy works closely with executive [2:01] teams to develop and implement [2:02] enterprise AI strategies with a focus on [2:05] aligning data, technology, and business [2:07] goals to achieve measurable outcomes. [2:11] So you guys ready to kick things off? [2:14] This is a question for all of you. Uh [2:16] what types of customer data are proving [2:18] most valuable in your work today and [2:20] why? And whoever wants to start can go. [2:24] >> Sure, I'm happy to happy to start. So uh [2:27] being a healthc care provider we find [2:29] that that clinical data and um [2:32] engagement data marrying the two of [2:34] those has proven a lot of value. So we [2:36] we as healthcare providers have often [2:38] known that laps lapsed customers laps [2:42] lapsed patients have indicators that the [2:44] insurers are aware about. There's like a [2:46] hyperfocus along um along the within the [2:50] provider community of what makes what [2:53] puts a patient at risk. um because [2:56] that's just what the providers have [2:57] done. But what we found is layering in [2:59] and marrying other forms of data, [3:02] engagement data, whether that's web, [3:04] whether that's call center data, CRM [3:07] data, bringing those together have has [3:10] allowed us to make models that take the [3:13] clinical guidance and just add one more [3:15] layer on top of that, right? It lets us [3:17] understand, okay, well, we all know that [3:19] this person is due for this type of [3:22] clinical encounter, clinical visit. What [3:25] we're trying to do on top of that, say, [3:26] well, how can we best bring them back [3:28] in? What channels do they want? What is [3:31] the right cadence? What is the what how [3:33] do we meet them where they are and not [3:36] just something something cold and [3:38] something that that that that they could [3:40] get from anyone else. We're finding that [3:42] the data that we have about them across [3:44] the ecosystem, there's an expectation [3:46] from consumers now that our strategy and [3:50] outreach should really mimic all the all [3:53] the things they've told about told us [3:56] about them that kind of matches what [3:58] other industries outside of healthcare [4:00] are doing. [4:01] >> How about you K? Yeah, I think there's [4:03] been a lot of progress in uh marketing [4:05] organizations, data organizations with [4:08] centralizing structured data in [4:10] relational databases and semi-structured [4:12] data from a website and mobile apps. But [4:14] there's a lot of value in unstructured [4:16] data. uh things like transcripts and [4:19] call logs, uh standard operating [4:20] procedure documents, even PowerPoint [4:23] decks and and sort of campaign [4:24] retrospectives that in in total become [4:28] this corpus of your business's reality. [4:32] uh and that becomes uh really essential [4:35] for uh defining and fine-tuning models [4:38] that can contextualize campaign [4:41] objectives can help you to mobilize the [4:43] right uh channel strategies and we're [4:47] seeing a lot of traction uh with [4:48] unstructured data [4:51] >> you think similar to you um the [4:55] juxtaposition of so first I have to [4:58] explain AEG vision has 509 optical [5:00] practices across the us they're all [5:03] independently branded. So we are [5:05] literally marketing 509 separate brands. [5:09] >> That's a lot. [5:10] >> Um with that the the juxosition of the [5:14] patient data because optical is highly [5:17] retail also from more of a customer [5:19] focus and then operational data in our [5:23] practices [5:24] um and marrying that with marketing [5:28] platform performance. So being able to [5:30] see the throughput from the time you get [5:32] the first touch of a potential patient [5:35] into how that person converts into a [5:39] patient, what they do in our practices [5:42] and then we marry third party data with [5:45] that. So who they are and where we can [5:49] ultimately then target the right uh [5:52] population for the offerings that our [5:54] practices have. Can [5:55] >> I ask a follow-up question to that? Um, [5:58] do you find that like marrying the third [6:00] party data and looking at things like [6:02] claims and other data sources for us [6:04] strategy has always loved that at large, [6:07] right? If they want to know where to put [6:09] a new practice or where to or where to [6:12] invest, they're really comfortable with [6:14] that at a large scale level. Have you [6:17] found um like organizationally any [6:21] resistance to using that on a more [6:23] personal, more onetoone marketing level? [6:27] Well, yes and no. So, we have a [6:29] fantastic legal team. Um, operating in [6:32] 30 states means navigating the uh [6:35] legalities of 30 states with HIPPA and [6:37] everything like that. I think um, you [6:40] know, we've got a pretty fantastic [6:41] culture. So, finding a way to make sure [6:45] that we are compliant, that we are [6:47] within the bounds of the law always and [6:49] that we are also doing everything we can [6:52] for the business. I think pretty much [6:54] our entire seauite, our senior [6:56] leadership is all aligned around that, [6:58] but it it's taken some pretty [7:00] intentional uh cultural nurturing to get [7:03] there. [7:04] >> How do the patients feel when they get [7:06] those personalized messages? Do they [7:08] feel freaked out or do they like it? [7:10] >> Well, I think similar to what Paul said, [7:14] don't make me tell you twice, right? [7:16] Don't make me tell you anything twice. [7:18] If I give you my information, you should [7:20] know me. You should know who I am. You [7:22] should know what I want. you should know [7:23] what I need. And if you can anticipate [7:26] my needs, even better. So, if you can [7:29] put that recall reminder in the [7:31] patient's hand, in the channel that they [7:34] look at at the time they need it before [7:36] they even realize they need it, that [7:38] that's where the sweet spot is where [7:41] they feel like you actually care about [7:43] them. [7:43] >> So, we we found a lot of um benefit in [7:46] things like a like a welcome series, [7:48] like a like a new user, a new patient. [7:51] We would traditionally just place in all [7:54] that content high value service lines. [7:57] Hey, look at us. We're really good at [8:00] liver transplants, but anyone in the [8:03] market for a liver? [8:05] Okay. So, what what we found is if we're [8:08] trying to expand it's it's healthare, so [8:11] it's always a little awkward to talk [8:13] about these marketing terms, but if we [8:15] want people to to get as much access to [8:18] as much care as they can for their for [8:20] their own health journey, well, we're [8:22] going to take a look at that first visit [8:24] and we're going to use some AI models [8:26] and we're going to find out what are [8:27] their next most likely service lines. [8:30] Right? So me as a a 40year-old [8:33] overweight man like there there's this [8:35] there you can look at that and say okay [8:36] I'm probably in ortho cardio and PCP [8:40] markets and we take that content that [8:43] we've always sent out right the [8:44] content's always been there it's been in [8:46] an email template and and we're just [8:48] injecting a personalized ordering of [8:50] that content say feature the service [8:52] lines that are most relevant to that [8:54] person. So it's a touch point that [8:56] already existed. It was a content [8:58] library that was already there and we [9:01] had to start with like a pi a Python [9:04] script that pushed a data extension into [9:06] Salesforce that was really rough when it [9:09] started but we could show after 6 months [9:12] that that group that received the [9:14] personalized messaging of hey here's our [9:17] three services at Northwell that maybe [9:20] we told them we didn't tell them it was [9:22] personalized to them but it was content [9:24] that they were receiving that was [9:25] person. So we found that there was a a [9:28] larger um we made up a metric distinct [9:32] number of service line participation [9:33] right so a share of wallet measure had [9:35] increased over the folks who were [9:37] getting the the ampute or the the liver [9:41] transplant messaging so that was one way [9:43] that that we were able to take that like [9:45] what we knew about them they're [9:47] expecting people are expecting [9:49] personalized content and not doing it in [9:52] a way that was overly dramatic or in [9:55] their face But like let's just subtle a [9:57] subtle touch point that was already in [9:59] place customizing that. [10:03] >> Jeremy, we're gonna start with you for [10:05] this next question. Um, how are you [10:07] using AI to analyze data and turn it [10:10] into actionable insights or customer [10:12] acquisition strategies? [10:14] >> Yeah. So, [10:16] traditionally acquisition has been and [10:18] still the case in many uh in many [10:20] respects. You're trying to capture [10:22] intense signals on a website, on a [10:24] mobile app, on a mobile experience, and [10:27] take all of those disparate signals and [10:30] multi-session signals and consolidate [10:33] them and rationalize them into a single [10:35] identifier [10:37] with which can be very um uh predictive [10:40] and very uh [10:43] effective in terms of outreach. But now [10:46] with AI, with unstructured data, with [10:49] lots of campaign retrospective uh [10:51] analysis that used to take weeks if not [10:55] um well weeks if not a couple of months [10:57] to really synthesize and and give you [10:59] some actionable insights. Now you can [11:01] take that data and you can make it [11:03] inference. You can make it uh sort of [11:05] inferencing ready if you will by a model [11:08] to distill down the key takeaways. So [11:11] what you can do now is take hey here's [11:12] what worked in the past 5 years in the [11:15] past you know series of campaigns for [11:17] these targets for these profiles and you [11:20] can take all the different signals that [11:21] are incoming and you can combine them [11:23] and then you can activate really precise [11:26] really accurate messaging high efficacy [11:28] high reach uh high conversion in a way [11:31] that wasn't possible even two years ago [11:34] I think that's a really exciting [11:35] opportunity um in insurance in [11:37] healthcare and across multiple verticals [11:39] every vertical [11:40] Michelle or Paul, do you guys have any [11:43] anything to add to that? [11:45] >> Yeah, I think a pretty tangible example [11:48] of that. So, um, when I joined AEG, the [11:52] majority of our marketing spend was [11:55] targeting new patients, but acquiring as [11:58] established patients. So, we had people, [12:01] we made it really easy to find our [12:03] information on on the web and they'd [12:06] click on our ad and call their doctor, [12:07] which is exactly what we don't want them [12:09] to do. Um, so we brought in call [12:13] tracking. We brought in an AI based [12:16] system that allowed us to identify new [12:18] versus established patients based on [12:20] different signals. um trained the model [12:22] for about three months and then deployed [12:25] it with our advertising very very [12:27] carefully only in key states um where we [12:31] were able to then let advertising know [12:34] this resulted in a conversion for a new [12:37] patient and filter out the draws of the [12:40] noise for established patients. They can [12:41] still find us. They can still click on [12:43] the ad. But, you know, if it's Google or [12:47] whoever, they're not getting the signal [12:48] that says you did good, kid. [12:52] >> That's great, Paul. [12:53] >> Cool. Uh, Paul, what are some practical [12:56] ways that you've integrated AI into your [12:58] existing marketing or data workflows and [13:00] what's worked well for you? [13:02] >> Yeah, I I I appreciate the question. So, [13:04] one of the things that that we had to [13:05] deal with was [13:08] uh putting, you know, butts and seats [13:10] post pandemic, right? had we had to get [13:12] people back into into the doctor's [13:14] offices and and again we we'll talk a [13:17] lot about the life cycle. So it's that [13:19] lapsed customer. So what we what we did [13:22] was like a pretty crude least to most [13:24] expensive channel outreach strategy, [13:26] right? We're going to we're going to [13:27] email them. We're going to send them a [13:29] text, a direct mail, then we're going to [13:30] send them to the call center based on [13:32] the fact that I'm cheap and that was if [13:35] if they if they booked after the email, [13:38] I would save the expense of the next the [13:40] next channels. So, that was a a good [13:42] platform for us to get to get started on [13:45] this kind of proactive outreach. Um, and [13:48] that's been wildly popular in our health [13:50] system. And as a platform that's driving [13:53] tens of thousands of appointments a [13:55] quarter, it's been really really great [13:57] for us to say, okay, we have this high [13:59] performing program. How can we optimize [14:01] them? And so because of the martekch [14:03] stack we're on and kind of the [14:05] consolidation that we did to get all [14:06] those channel outreaches in that native [14:09] platform, there are models that are [14:12] there ready to turn on for channel [14:14] optimization, for right time send, for [14:17] mix optimization. So that's been really [14:19] a benefit to us. We talked about we [14:22] talked about that that model earlier [14:24] about ranking the content in a welcome [14:26] series and that was a Python script that [14:29] got uploaded to a cloud function and got [14:32] pushed to a data extension that's was [14:33] really rudimentary. So knowing that that [14:36] was kind of the less than ideal state, [14:39] we took a lot of time training the team [14:41] into getting getting as much of this [14:44] native functionality out of these [14:46] platforms as we could. And now that we [14:48] we've done that, we've spent a year and [14:50] a half of, you know, slapping wrists [14:52] when people are quick to put something [14:54] to market when there's a a slower way to [14:57] get up but is faster for optimization. [14:59] That that's been a real benefit to us. [15:01] So, it's h half of it is knowing what [15:04] the the vendors are able to do and and [15:07] building in a way that allows you to [15:09] take advantage of that as quickly as you [15:10] can for go to market. [15:12] >> Awesome. Relle, you've integrated AI at [15:14] scale. What's worked well for you? [15:17] The example I gave is probably the best [15:20] one. I think, you know, to be honest, we [15:22] don't have it figured out yet. I mean, [15:23] we're we're we are just dipping our toes [15:26] in the water. I think um the head of IT [15:30] for my company is here, Satic, and he's [15:32] done [15:34] things that seem really basic and small, [15:37] but at the same time add they're not [15:40] basic and they're not small, and they [15:41] add a ton of value. um as simple as uh [15:45] building capabilities for a patient to [15:48] take a picture of their insurance card [15:49] and not have to enter all their [15:51] information in and recognize that and [15:53] fill it out in our system. So it doesn't [15:56] have to be some monumental gamechanging [16:00] integration. It can be yeah a small [16:03] needle mover. [16:04] >> Yeah, that's important to know. And [16:06] Jeremy, you have may have some existing [16:08] frameworks that have worked well. What's [16:10] worked well with you guys? Yeah. So, any [16:14] good consultant, we've got lots of [16:15] frameworks. Um, for for marketers, it's [16:19] always been important to have the right [16:20] context. Um, what's worked well in the [16:23] past and how are we going to refine that [16:25] and and take what's worked and just hone [16:28] it. What we'd like to think about is [16:31] first contextualizing the right data for [16:33] the use case to achieve that outcome, [16:36] right? aligning use cases to outcomes [16:38] but then taking all the context and [16:41] using that to make the right decisions [16:44] but the decisions have to um be [16:47] mobilized right so once you have context [16:49] how do you take that context and make [16:51] the right decisions once you have the [16:53] right decisions in in sort of [16:54] established how do you mobilize them on [16:56] which channels with which agents how are [16:58] you orchestrating across those agents [17:00] and then really importantly Paul you [17:02] said it how do you optimize and feedback [17:05] iteratively [17:06] what's worked and the learning so you [17:08] have that uh that flywheel you can check [17:11] flywheel off your bingo cards but that's [17:14] that's really important is is taking [17:15] that that sort of complete loop and [17:17] going from optimization back to context [17:19] and improving the decision-m and [17:21] improving uh the means with which you [17:23] mobilize those agents [17:25] >> Jeremy as you've as you've done this [17:28] because you've you've watched rinse [17:30] repeat a few times [17:32] how do you find organizations are [17:34] thinking about the flywheel piece of it [17:38] because one of the things that I think [17:40] I'm concerned about as a large [17:42] enterprise organization is the thought [17:45] that it's one and done [17:47] >> and not that it is another employee that [17:50] you have to train and guide and coach [17:53] and come back and and work with and it [17:56] takes resources. [17:57] >> Yeah, it takes it takes resources and it [17:59] takes a lot of iteration uh and sort of [18:02] incremental progress. Um, even before, [18:06] you know, AI was top of- mind for [18:08] everybody, there was still this [18:10] misconception that your data had to be [18:12] perfect and it had to be complete for [18:14] you to use it and harness it and get it [18:16] out there and achieve these outcomes. [18:18] But you don't have to have one, there's [18:21] a couple points. One, you don't have to [18:22] have perfect data to make progress with [18:25] AI. Um, you don't have to have all of [18:27] the data in one place. you don't have [18:29] have all the data aligned to a customer [18:32] persona or unique identifier. You can [18:35] get started incrementally with a subset [18:37] of the data. The other piece that's [18:39] really important is that as you I I [18:42] think sometimes organizations are [18:43] looking at well how do I achieve the [18:44] flywheel? That's that's my goal when [18:47] really the goal should be how do I align [18:50] AI and use cases to achieving business [18:52] objectives and how do I get pilots in [18:55] production that are not isolated that [18:58] are not sort of decoupled from the [19:00] business objectives but how do I get [19:02] those in production realize results and [19:04] then start to scale those pilots across [19:07] business functions uh or even within a [19:09] business function that's how you start [19:11] to generate uh that flywheel effect [19:13] >> I like [19:14] Uh Michelle, what we'll start with you [19:16] on this one and uh Jeremy and Paul can [19:19] chime in. Uh what challenges have you [19:21] faced in scaling AIdriven marketing [19:23] efforts, especially when working with [19:25] your siloed data or you know your [19:27] complex organization? [19:28] >> Yeah, I think uh every business is [19:32] unique. Um, and [19:36] every marketing company, every, it [19:39] doesn't matter if they're AI or not, is [19:41] going to tell you they can solve your [19:43] business problems. [19:44] >> Yeah. [19:44] >> And they underestimate the need to be [19:47] able to really integrate [19:50] um the full funnel of the patient into [19:55] uh your operational data. So, I think [19:59] one of the things that has been really [20:01] important to me is starting small and [20:05] and it's it's been a recipe for success [20:07] over and over and over again. Starting [20:09] small and giving um my business [20:12] stakeholders attribution [20:14] >> and being able to point to I spent a [20:17] dollar here. This is where it landed in [20:19] our practice. This is what they did. [20:21] This is the type of patient. This is how [20:23] much they're spending. This is the [20:25] pattern we're seeing ongoing. I think [20:27] when you're able to do that, it [20:29] establishes trust [20:31] >> and then the next time that you want to [20:34] bring something more advanced like [20:36] AIdriven call tracking or AIdriven QA [20:39] for our call center or any of those are [20:41] both initiatives that we either have [20:43] done or will do. Um the the push back is [20:46] significantly less. the buying's higher, [20:49] the support's higher, and then the the [20:51] tolerance for how you need to shift [20:55] behaviors even in our practices is [20:57] higher. [20:58] >> And I'm sure that attribution is so [21:00] helpful to your clients. [21:01] >> Yeah. [21:01] >> Yeah. They need that. [21:03] >> If I can add to that, I think there's a [21:05] a huge benefit to getting both of you [21:07] said this to getting AI use cases in [21:10] market, right? Not even just as a hello [21:12] world as a hey, we can do this. the [21:15] pipes can work, but set up an [21:17] experiment, right? Make make sure that [21:18] you can show that attribution. And then [21:21] in our kind of journey, if you look at [21:23] that as a maturity curve, we're probably [21:25] three out of five on this journey to the [21:27] point where it's kind of triggered a a [21:30] digital transformation with with the [21:32] firm. And that brings so many resources, [21:36] so many people with so many opinions. [21:39] And we're we're caught in this pendulum, [21:40] right? And the pendulum of it is data [21:43] storage costs increase time to push them [21:45] down. When data storage costs decrease, [21:48] well then latency issues pop up, right? [21:50] And then governance comes in to decrease [21:52] risk, but then you're slower to go to [21:53] market. So there's always going to be [21:55] some swing back and forth and and and to [21:58] solve that, man, I'm at a I'm at a [22:01] 115,000 [22:03] person firm and it's just personal [22:05] relationships with all these leaders who [22:07] are getting brought in and and there's [22:09] so many ships to turn around trying to [22:12] work with them to to show like hey this [22:14] is this functionality can work. We have [22:16] the bonafetus. We have the we talked [22:18] about how this is not all about [22:19] dashboards. We have the dashboards that [22:21] show real-time calls being driven. We [22:23] have the we have the the larger studies [22:26] that show incremental spend in marketing [22:29] leads to appointments and having those [22:30] harder metrics can help form that [22:32] conversation. But it is constant tending [22:35] on on on that spectrum to make sure that [22:37] we're doing right by the company. [22:39] >> I think the same thing being transparent [22:41] when it doesn't work. We have something [22:43] we tested [22:45] it has failed miserably. [22:47] uh we won't be moving forward with that [22:50] uh particular project. And being able to [22:53] say to my operators, hey, we spent your [22:55] money. It didn't work. It was a test. It [22:57] was limited. There's a reason it was [23:00] limited. I think being able to show both [23:03] sides of the coin because, you know, as [23:04] marketers, we kind of were really good [23:06] at selling things. We get really excited [23:07] about things and it can impact how much [23:12] people trust you. But when you're [23:13] willing to say, "Hey, full open books. [23:16] here's what working, what's working, [23:18] here's where we don't know yet, and [23:20] here's what's not that we won't do [23:21] again. I think it it does go a long way. [23:25] >> That's great. Uh this next question is [23:27] for all of you. Uh it's a little bit [23:29] lighter, so maybe you guys will be able [23:31] to enjoy it. Uh can you share a [23:33] surprising insight uh that AI helped [23:36] uncover about your customers? [23:39] >> Uh yeah, I can. I'm happy. So, this is a [23:41] couple years dated, but I really like [23:43] this example. So, um, this is the second [23:45] time I brought it up, but there was [23:46] there was a pandemic a few years ago. [23:48] Um, and we're in New York City, so we [23:50] got hit really, really hard. And one of [23:52] the things that that there's this this, [23:54] uh, one of the one of the CTOs, the [23:57] health informatics guys came to me and [23:59] he had gotten called back from the back [24:00] office to the front line and was working [24:02] in um, pediatric ER. This is like April, [24:07] May of 2020. Like, Paul, I know that [24:10] Google can spot flu. Can we spot this [24:15] please? Right? So what we did is we [24:17] grabbed all of our the time we were [24:19] using Google Analytics because it was [24:21] okay to do that back then and we and we [24:23] took that and we ran it through some [24:25] machine learning algorithms. We [24:27] categorized it. We said, you know, find [24:29] out some trends. We put it against test [24:32] cases in our region. And we saw that at [24:37] about two weeks before a spike in test [24:39] cases, the way that people a community [24:44] uses uh hospitals website changes, [24:48] right? And it was all black box to me. [24:49] But we looked at some examples. People [24:52] don't care about plastic surgery, right? [24:54] That that that drops. People don't care [24:56] about um about insurance. They care [24:58] about weight times. They care about [25:00] parking. They don't care about doctor's [25:02] profiles anymore. And so we were able to [25:04] show that oh man there's like real value [25:07] in marketing data to the operations of a [25:10] health system. So that became this [25:12] indicator that we started to publish in [25:14] our health system. Um we got some [25:17] attention for this model. It got went on [25:19] to GitHub. I had calls with a bunch of [25:21] health systems around the country and [25:22] then we're started looking at this of [25:23] like wait I think there's like value [25:27] there's value to marketing data inside [25:29] of inside of a healthare system. For us [25:31] that was like a really a really like I [25:34] mean life-saving but but an important [25:37] thing for for institutional memory to [25:40] have right of like hey when we are all [25:43] for us it's all trained doctors running [25:45] the health system which is wonderful but [25:47] there's a little bit of like oh no there [25:49] there there are other silos that are [25:51] outside of the health record that might [25:53] be valuable and so you can see that like [25:57] example and me and that that that that [25:59] that that [26:01] seat that um health informatics officer [26:03] and our our relationship now has kind of [26:06] helped the company move to a place where [26:09] we are brought in more on these large [26:11] data decisions and we're on governance [26:13] committees and we're on u we are data [26:16] stewards and so for us look finding that [26:20] um and the value that could add was like [26:22] a a really fun example for us of of what [26:24] could happen if we integrated more. [26:26] >> That is cool Jeremy. [26:27] >> Yeah. So one of the unique uh use cases [26:31] we've worked on with a insurance client [26:34] is on the claims adjudication process [26:36] which is uh sort of can be long and and [26:39] frustrating and um both for the uh the [26:42] claimant or the beneficiary for the [26:44] examiner for everybody involved. It it [26:47] can take weeks. It can take weeks just [26:49] to get the right data and get it all [26:51] complete for the examiner to review it [26:54] uh and for the payout to occur. [26:57] What we've what we've encountered, what [26:59] we've sort of learned in this process is [27:01] that for a beneficiary who isn't even [27:04] directly involved but is, you know, on [27:07] the on um or on the hook to to fulfill [27:10] this claim or process this claim, [27:12] getting all of the right documents and [27:15] unstructured data and information for [27:18] the examiner can be difficult. And so [27:20] with AI, you can now have a a a corpus [27:25] of knowledge, a a very domainspecific [27:27] model that is an expert in claims [27:30] adjudication and claims processing that [27:33] can tell the examiner, hey, here's all [27:35] the missing information and here are the [27:37] questions to ask the claimant or the [27:39] beneficiary so that you can give them a [27:42] good experience, but also get the [27:43] information you need to process the [27:44] claim in a timely manner. And the [27:49] outcome of that is one a faster time to [27:51] pay out. But what is really interesting [27:52] is that for these insurers what they [27:55] what they are experiencing with a better [27:58] processing with a better claims [27:59] processing experience is that the [28:02] claimant or the beneficiary really the [28:04] beneficiary is going to seek them out [28:06] the insurer for insurance. So they're if [28:09] they can give a good experience in that [28:12] process they open the door to acquire a [28:15] new customer. So it's a really um [28:18] fantastic secondary outcome or sort of [28:20] uh you know related outcome of that [28:24] claims experience that was driven or [28:26] improved through AI. [28:28] >> What is uh surprised you with AI? [28:31] >> So this is my former role. I've been at [28:33] AEG for seven months. Um in my former [28:37] role we did an evaluation where we took [28:40] uh multiple factors and this this is it [28:43] is patient but it's not patient. patient [28:45] input and business output. [28:47] >> Um, we took net promoter score, we took [28:50] Google score, we took product mix and [28:53] sales volume and a bunch of other things [28:57] that I can't really get into and threw [29:00] it into the hopper with patient mix. And [29:02] what we found was there was one single [29:06] predictor of growth [29:08] >> uh both in revenue and even more so in [29:11] Eida and that is Google score. Really? [29:15] >> Yep. So, uh, a practice that had a [29:18] Google score of, and it's funny because [29:21] we could have done this analysis just [29:23] flat if we had known, you know, that it [29:24] was just that. Um, but a practice that [29:27] had a score of 4.6 was twice as likely [29:31] to grow in revenue year-over-year than a [29:33] PRA or sorry, 4.8 than a practice at 4.6 [29:36] and 4x in EVA. [29:38] >> Wow. [29:39] >> Yeah, [29:40] >> that is cool. Hey. Um, for this next [29:44] question, we'll start with you, [29:45] Michelle. Uh, what AI powered content [29:48] strategies or personalization efforts [29:50] have you seen success with? Kind of the [29:52] same thing, but or maybe what you're [29:54] exploring now. [29:55] >> Uh, I'd have to go with exploring. [29:57] >> Okay. [29:58] >> Um, so 509 practices, uh, there is no [30:02] way we can produce that volume of [30:04] content, [30:04] >> right? So, especially because video is [30:08] what gets people's attention now and [30:10] video is very uh if you do it oneonone, [30:12] it's very resource intensive to produce [30:15] right now. What we're working through is [30:17] creating the content libraries, creating [30:19] the segmentation around our practices [30:21] and their patients and the demographics [30:23] and what we're targeting so that we can [30:26] then segment our practices into groups [30:29] and produce content, [30:31] >> okay, [30:31] >> through AI. Um, I was at a great [30:34] conference a couple weeks ago where I [30:36] saw this done and it was literally a [30:38] team was able to produce and I saw the [30:41] video last night on uh on TV. They were [30:44] able to produce a commercial, a [30:46] full-blown ready for TV commercial in [30:50] one hour with one resource. [30:52] >> Wow. [30:53] >> Yeah. [30:53] >> And so that's that's kind of where we'll [30:55] be leaning in. [30:56] >> Nice. [30:58] Are there any specific tools that you're [31:00] thinking of using or can you say? [31:02] >> Can't say right now. We're still [31:04] exploring. [31:06] >> Okay. [31:06] >> I'll be attending a few sessions here. [31:08] >> Okay. [31:09] >> How about you guys? [31:10] >> Yeah. So um what has been really uh [31:14] interesting is using AI to look around [31:18] corners and identify not blind spots but [31:21] sort of hidden opportunities that [31:23] otherwise preai a marketer may not uh [31:27] you know surface or or or illuminate. [31:30] What I mean by that is when they're [31:32] creating audiences and defining these [31:33] cohorts, they've got filters and and [31:36] criteria that they're using to ideulate, [31:38] right? Their own intuition to define [31:40] these cohorts. And they want to look at [31:42] inter in inter um intersecting segments [31:45] and overlaps and you know what what [31:48] percentage of the population is also in [31:49] this audience. [31:51] What AI can unlock for them is [31:54] five additional audiences or questions [31:56] to think about that they may not have [31:57] uncovered or thought about which can [31:59] unlock entire new campaigns or new [32:02] audiences or new um sort of activation [32:05] strategies which are you know high [32:07] converting contain costs well uh [32:11] preserve margin and so it's it can [32:13] become a a thought partner an [32:15] experimentter working alongside you. [32:19] Do you have any? [32:19] >> I I would add to both those are great [32:21] great examples. For one, we are just [32:23] like 7,000 employed physicians, a bunch [32:26] more voluntary head shot. [32:29] >> That's it. We are We are trying to get [32:31] people to come to Long Island to take a [32:33] head shot. No one wants to do it. Yeah. [32:34] >> So, they're just so used to like, I know [32:37] that I can do an AI head shot. Why are [32:39] you making me drive on the BE? And so [32:43] for for us, this is if we get this done [32:46] this year, and we're we're working on [32:47] the pilot right now, this will be what [32:49] we're known for for like five years is [32:51] marketing made a a headshot AI outside [32:53] of everything else we've done, that's [32:55] the thing that the physicians will be [32:56] most pleased by is we've removed one [32:59] commute from their lives. And [33:01] >> that's that's big. Yeah, [33:03] >> I know for us at AI News, um we just [33:07] started writing our news articles for [33:10] AI. So, typically in journalism, you [33:13] write for the humans and you tell a [33:15] story and it's got this arc and it takes [33:17] the readers on a journey. Well, uh, AI [33:22] AI doesn't do that. And especially with [33:25] the LLMs with Chad GBT and Perplexity, [33:28] Gemini, all that stuff, they um, they [33:32] scan they quickly scan the articles that [33:34] are on the internet. They just look for [33:36] a few things, few facts, few statistics, [33:39] and they're not reading the whole [33:40] article. So your storytelling doesn't [33:43] matter to AI. And so I just asked chat [33:48] GPT one day, I'm like, how can I write [33:50] my news articles so that our articles [33:53] will be more discoverable by AI? Because [33:55] you know with Google searches they only [33:57] have those AI overviews or you go to [34:00] chat GPT or Perplexity and they just [34:02] give you the answer and no one's [34:05] scrolling down to click on any website [34:07] links and so website traffic is dropped [34:09] by like half and so I just went to chat [34:11] GBT and I said how can I write my news [34:14] articles so that they're more [34:16] discoverable by the LLMs and Chad GBT [34:20] gave me a big list of things. So besides [34:24] making sure you know you have a main [34:26] header and then a couple of little [34:27] header twos and a little three and then [34:30] you're going to bold some key terms [34:33] you could kind of think of them as SEO [34:36] keywords kind of but they're a little [34:38] bit longer little phrase more phrases. [34:41] So datadriven marketing would be a [34:44] phrase that would be highlighted in the [34:46] article but there's bolded terms [34:47] throughout. There's some key takeaway [34:49] bullet points at the top and a Q&A [34:52] section at the bottom. And AI looks for [34:54] those quick facts, quick questions and [34:57] answers, quick statistics. And so once I [35:01] started writing our news articles for AI [35:03] discoverability, [35:05] we doubled our reach. We went from maybe [35:09] having 10 to 20% see our articles. Now [35:13] we're seeing 30 to 40 to 50% sometimes. [35:17] So, our reach has increased a lot. And I [35:20] know it doesn't maybe sound like a lot, [35:23] but it is for us. And to me, just [35:26] writing for those LLMs and being able to [35:29] get our content discovered more so that [35:31] they can recog re recommend our AI news [35:34] platform. That was huge for us. And [35:36] we're constantly seeing new subscribers [35:38] from chat GBT and perplexity. So, AD [35:41] just ask chat GBT what you should do. [35:46] Um, okay. That was just a side note. [35:50] >> Uh, what advice, uh, do you have, [35:52] Jeremy, for marketing teams that are [35:54] trying to pick the right use cases for [35:56] AI? [35:59] >> I'll sound like a broken record. [36:02] You should always anchor your use cases [36:05] to those northstar objectives and [36:09] outcomes. Um, if they're not aligned to [36:13] those, then you're off doing science [36:15] projects and consuming resources and [36:17] time. But in all seriousness, the use [36:19] case needs to align to the business [36:21] objective and then align to the the KPIs [36:25] and the strategies that are going to [36:26] help you achieve those business [36:27] objectives. And so the use case is [36:28] really uh the vehicle to help you [36:32] achieve the KPI or or hit the target or [36:35] the metric. Um, and so you you can think [36:38] about that really like a hierarchy [36:40] business objective and strategy and KPI [36:42] and use case and with AI and with [36:45] agentic AI that that sort of scaffolding [36:47] that framework still holds true. Um, but [36:50] if you're ever uh questioning, you know, [36:52] is this the right use case or am I [36:55] focused in the right areas, you should [36:56] just walk that ladder back up and ask [36:59] yourself, what are the KPIs and those [37:00] KPIs help me achieve a strategy and are [37:03] those strategies aligned to my business [37:04] outcomes and objectives? [37:06] >> How about you, Paul? Do you have any [37:07] personalization use cases? [37:09] >> Yeah, I think um for for me again owning [37:13] the Martekch stack is very very clear [37:15] what what I'm intending to do. I I have [37:17] to increase clickthrough rates. after [37:18] include o opens and so for me pointing [37:22] pointing my coordinators my directors [37:25] the folks who sit in these tools towards [37:27] that and enabling them with the ability [37:30] to do experiments has been something [37:32] that's that's opened it up a lot for me [37:34] is like removing the pressure to produce [37:38] and and padding their time with like hey [37:41] you made some decisions on the way to [37:43] production you want to test any of them [37:46] and we're doing the same thing with our [37:47] UX team this year because we we've made [37:49] some advancements in in some AB tools um [37:52] is is telling them take those decisions [37:54] you did along the way and and and go to [37:57] market with a val with a validation of [37:58] them and so for us that that kind of [38:01] opens up well what variables [38:04] have did I think about during [38:05] development when I was doing research [38:08] what what did what assumptions did I [38:10] make on the majority of of them are this [38:13] then backing that up to well what if you [38:16] didn't have to narrow down into one. [38:18] What What if you could actually do two [38:21] or three versions and and and [38:24] it's a it's a shift in mindset for a lot [38:27] of these folks who are just so used to [38:29] to working within bounds where they have [38:31] to make a version allowing them to think [38:34] in in seven segments which is kind of [38:37] where most folks end up in a in a c in a [38:39] in a in a cluster analysis but then [38:41] moving to microtargeting which is beyond [38:44] that. So, it's changing the way they're [38:46] thinking is a lot of what we're working [38:47] on this year. [38:48] >> Gotcha. Uh, finally, this one last [38:51] question. We're going to make it a [38:52] lightning round. In one sentence, what's [38:54] your top advice for marketers getting [38:56] started in AI with customer data? [38:58] Michelle, [39:00] >> start small. [39:02] >> Start with something that's an easy win [39:04] that gets your team a few reps, that [39:05] gets your organization a few reps. [39:07] >> I like that. That's great advice, [39:08] Jeremy. Your data does not need to be [39:11] perfect to make progress. Just get [39:14] started. Get started small. Learn. [39:18] That's the sentence that rambled. That's [39:19] the sentence. [39:21] >> Just so fun. It's so fun to have to do [39:23] this stuff. [39:24] >> It is. [39:24] >> Start playing. [39:25] >> Yeah. [39:26] >> I know. If I get just start playing [39:28] around with the tools or just start [39:30] experimenting or talking to a different [39:32] LLM, I find so many new ideas that I [39:35] hadn't even thought of. So, yeah, just [39:37] get used to it. Start small. I love it. [39:40] Well, uh, thank you everyone for joining [39:43] us and coming out today and thank you to [39:45] Michelle and Jeremy and Paul. Uh, it's [39:49] this has been really great conversation. [39:51] Whether you're just getting started, I [39:53] hope today's discussion has given you [39:55] some good ideas.