State, Federal Officials Deploy AI to Cut Government Backlogs

州与联邦官员部署AI化解政府积压工作 公共部门领导者正转向专用多智能体AI与自动化工作流，以应对激增的FOIA及许可审批工作负担，并指出如今阻碍政府效率的主要壁垒在于领导层的采纳意愿，而非技术本身。 公共部门IT与隐私主管分享了部署AI以实现政府遗留系统现代化、减轻后台工作负担并加快公共服务交付的具体策略。印第安纳州重点介绍了通过代码逆向工程和自动化处理将日常任务减少60%至70%的成效；犹他州则详细阐述了将合同审查时间缩短75%的多智能体AI架构。DHS强调利用AI在严格的法定期限内处理每年超过100万件FOIA申请，且未推高人员成本。各机构与会专家一致强调，在以90天为周期的快速技术迭代中，必须确保人类对最终决策承担责任、选用适合具体目标的模型，并克服领导层的自满心态。 印第安纳州利用 AI 对缺乏文档的历史遗留代码进行逆向工程，并部署虚拟许可智能体，将后台日常工作量减少了 60% 至 70%，把企业级系统的翻新周期压缩至 3 到 18 个月。 (3:45) 美国国土安全部（DHS）每年面临超过 100 万件 FOIA 申请的法定处理压力，通过引入 AI 数字化摄取流程，彻底终结了将邮件附件打印后再手动重扫录入的低效纸质流转闭环。 (6:05) 因现成通用大语言模型无法应对复杂的法律法规审查，犹他州开发了由 9 个专属 Agent 组成的多智能体架构，使政府合同合规分析的审查时间减少了 75%。 (8:44) 面对公共服务规模扩大与严格的数据合规要求，各机构在推动采购周期缩短至 3 到 12 个月的同时，构建了不在 AI 架构中留存任何公民数据或可识别日志的隐私隔离机制。 (29:54)

Public sector technology leaders are deploying targeted artificial intelligence to alleviate severe back-office backlogs and modernize legacy infrastructure, arguing that operational bottlenecks stem from bureaucratic complacency rather than technological limitations. Alleviating Caseload Pressures and Back-Office Workloads (3:45) Facing ballooning operational demands and strict statutory mandates, agencies are using AI to reduce repetitive clerical burdens. At the U.S. Department of Homeland Security (DHS), which receives over one million Freedom of Information Act (FOIA) requests annually subject to a 20-business-day response window, hiring enough personnel to clear backlogs is financially impossible. Automated intake workflows eliminated paper-bound bottlenecks—such as printing and rescanning emailed fingerprint scans for immigrant record requests—accelerating turnaround times for critical records from CBP, TSA, and ICE. DHS Chief Privacy Officer Roman Jankowski noted that automated processing prevents administrative delays that would otherwise disrupt public-facing programs or airport security lines. In Indiana, Chief Information Officer Rob Fulk reported a 60% to 70% reduction in routine back-office workloads across divisions regulating securities, elections, and roughly one million businesses. The state deployed AI to automate documentation checks, license approval recommendations, and virtual licensing assistance. Fulk emphasized that this efficiency shifts workers away from busywork toward providing direct constituent service, establishing a framework where automation handles data intake while human officials remain 'above the loop' to retain statutory authority and make final decisions. Tailored Architectures over Generic Chatbots (8:44) Agency leaders reported that general-purpose commercial chatbots frequently fail at complex government tasks, necessitating modular, domain-specific systems. Christian Napier, Director of AI for the Utah Division of Technology Services, noted that an initial attempt to use off-the-shelf Gemini models to evaluate thousands of state vendor contracts against statutory rules proved inadequate. In response, Utah implemented a custom architecture of nine specialized AI agents, each auditing contracts against distinct legal and regulatory standards. The system completed full compliance reviews in roughly 10 minutes, reducing analyst review times by 75%. Regarding infrastructure choices, the panelists cautioned against public agencies attempting to develop foundational large language models. Fulk argued that public-private partnerships should leverage existing hyperscalers—such as Google Cloud Platform, AWS Bedrock, and Microsoft Azure—allowing public sector IT teams to focus internal resources on workflow logic and data structure design. Napier corroborated this approach, observing that deploying smaller, fit-for-purpose models dramatically lowered development expenses and avoided the unnecessary cost of querying massive frontier models for simple tasks. Data Isolation and Privacy Mandates (29:54) Balancing automated data processing with strict statutory privacy laws requires rigorous system design. DHS manages global FOIA intake alongside sensitive law enforcement records, relying on dedicated legal teams to evaluate statutory limits on inter-agency data sharing. Indiana enforces logical and physical separation between public business filings and confidential corporate or regulatory records across 36 sectors. To prevent unauthorized surveillance and preserve public trust, Utah introduced a unified conversational AI architecture. The framework does not permanently store constituent data within the AI environment; it extracts data temporarily from primary agency systems to fulfill a transaction and returns it immediately. Furthermore, operational logs are stripped of identifiers so administrators cannot cross-reference or triangulate individual constituent records across separate state services. Procurement Speed and Emerging Security Threats (38:56) While technology adoption accelerates, panel leaders warned that legacy procurement practices and security challenges pose serious operational hazards. Fulk abandoned open-ended multi-year procurement cycles, instead enforcing strict 3-, 6-, 9-, and 12-month delivery benchmarks for vendors, and warned that agency leadership complacency is a greater constraint than technological capacity when software release cycles run in 90-day sprints. Panelists also warned vendors against vague sales pitches that fail to target defined regulatory problems. Emerging threats also challenge long-term adoption. Napier highlighted the continuous cybersecurity effort required to harden thousands of legacy government applications against AI-assisted bad actors. At the federal level, Jankowski raised concerns about 'mosaic theory,' wherein malicious actors file coordinated, disparate FOIA requests to reverse-engineer government AI architectures and uncover hidden entry points for network intrusions. Public agencies are transitioning from experimental chatbot deployments to integrated agentic workflows and strict data governance architectures. However, resolving the tension between rapid 90-day commercial development sprints and slow-moving government procurement cycles remains an active administrative challenge.

美国联邦与州政府数字化转型官员指出，公共部门引入人工智能并非单纯追求技术本身，而是为了在预算与编制受限的重压下化解日益严重的业务积压并提升行政效能，同时通过严格的架构隔离与人类最终裁量机制确保公众信任与数据隐私。 破解业务积压与现代化的落地成效 (3:45) 在联邦层面，美国国土安全部（DHS）每年面临超过100万件《信息自由法》（FOIA）信息公开申请，法定回复时限仅有20个工作日（可申请延长10日），依靠增聘人力无法承受天价成本。DHS通过AI优化高风险移民记录与指纹卡调取流程，取消了过去将电子邮件附件打印后再重新扫描录入的繁琐环节，实现了即时数字化摄取。在州政府层面，印第安纳州州务卿办公室针对近100万家监管企业与遗留系统，利用AI反向工程解析缺乏文档记录的旧代码与业务规则，在3至18个月内完成了遗留系统重构，使后台例行行政工作量减少了60%至70%，让一线人员得以转向更具服务性的工作。 犹他州则专注于专项垂直流程的自动化。该州针对大量需要核对法定法规与采购政策的政府采购合同，最初尝试通用聊天机器人但因法律规则复杂而告失败，随后构建了包含9个专项AI智能体的架构，针对合同不同条款展开合规比对。该流程在约10分钟内即可完成单份合同的筛查，为合同审核分析师节省了约75%的时间，其经验已开始向全州其他部门推广。 技术选型策略与人类监督界限 (11:18) 面对技术栈与采购决策，受访官员一致认为公共部门应聚焦于业务逻辑而非重复开发底层模型。印第安纳州首席信息官Rob Fulk指出，公共机构不应耗费数十亿美元去研发基础大语言模型，基础模型开发应由超大规模云厂商承担，政府团队的核心专长在于业务工作流与定制应用开发；同时他主张打破多年期的传统冗长采购流程，对供应商推行严格的3、6、9、12个月交付周期，并告诫供应商不要向政府领导人索求需求清单，而应直接提供具备差异化竞争力的具体解决方案。犹他州则综合运用Google Cloud Platform、AWS Bedrock与Azure等平台，指出定制化开发成本已大幅降低，小规模垂直模型在很多任务上比昂贵的前沿模型更具成本效益。 在系统操作与治理机制上，各方强调AI的职能在于生成决策建议，而不是替代行政责任。DHS隐私与信息公开官Roman Jankowski强调，出于国家安全、行动安全以及个人可识别信息（PII）的保护要求，自动化流程必须保留“人在环路中”（Human in the loop）。Rob Fulk进一步提出“人在环路上”（Humans above the loop）的监管框架，即由AI处理后台繁琐任务并提供审批建议，但全部法定责任与最终审批裁定必须由具备公权力的人员签署承担。 数据隐私边界与对话架构设计 (29:54) 各州与联邦部门在推进数字化改革时，正努力平衡政务服务质量与严苛的法定隐私限制。Rob Fulk强调将IT部门打造成“产品组织”，用私营部门的客户体验标准重塑政务办事流程；但在数据层面，跨部门数据共享受到严格的法律防火墙限制。以印第安纳州为例，其系统必须对36个监管领域、近100万家企业的数据进行物理与逻辑隔离，严格区分公开信息、机密财务记录与执法敏感数据；DHS则由于其FOIA申请人面向全球公众，必须由4至5名专业隐私律师组成的法律团队针对复杂数据披露进行专项合规评估，严防非法泄露。 为解决公民与政府互动中的信任隐患，犹他州设计了标准化的政务对话AI架构。该架构确立了三条硬性技术边界：首先，AI架构内部不永久存储任何公民个人数据；其次，数据仅在处理事务时从主管业务系统的数据库中临时调取，交易完成后立即返还且不留存副本；最后，系统运行日志会被彻底剥离身份标识信息，防止即使拥有后台访问权限的州政府雇员通过多维日志数据交叉比对来重新识别出具体公民身份。 网络安全、拼接理论与领导力滞后挑战 (38:56) 公共部门在拥抱AI的过程中面临严峻的安全与组织阻碍。犹他州AI负责人Christian Napier指出，州政府核心AI工程团队仅有2人，需要支持与培训分散在各机构的500名IT工程师，技术承载力处于紧平衡状态；而在安全防护上，各部门运行着数千个亟待加固的遗留应用程序，恶意行为者正利用AI工具加速挖掘这些系统的底层漏洞。 DHS的Roman Jankowski特别指出了一种源自国家安全层面的新型情报威胁——“拼接理论”（Mosaic theory）。恶意实体可能并不直接发动网络攻击，而是通过递交多份看似无关的独立FOIA申请，系统性索取分散的技术与运营记录，再拼凑还原出政府AI基础设施的完整内部架构，从而寻找突破口。对此，Rob Fulk总结认为，当前AI技术已进入90天的超高速迭代周期，技术本身已不再是转型的根本瓶颈，真正的最大威胁在于政府领导层面的因循守旧与治理框架、采购机制的迟缓滞后。 本场讨论展现了公共部门在AI实践中从概念验证走向深层业务重塑的过程。然而，遗留系统的脆弱安全性、针对FOIA透明度机制的新型对抗性探测，以及以90天为迭代周期的技术现实与政府传统规章体系之间的节奏错位，仍是各级公共管理者尚未完全解决的根本矛盾。

## Background 该内容由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上，视频标题为《AI for the Public Good: AI Use in the Public Sector》。根据页面简介，本次研讨会汇聚了来自美国联邦及州政府的公共部门领袖，探讨如何利用 AI 应对“用更少资源做更多事”的压力，并深入解析 AI 在欺诈检测、资源规划及公共服务交付中的真实部署案例，以及如何在创新与公平、隐私和问责制度之间取得平衡。 ## TL;DR 本次研讨会围绕公共部门部署与落地 AI 的实际经验展开，来自犹他州、印第安纳州和美国国土安全部（DHS）的专家分享了具体应用场景。讨论涵盖了利用 AI 对旧系统进行逆向工程、部署多智能体（Multi-Agent）系统加速政府合同合规审查，以及自动化处理海量 FOIA 请求与移民档案数字化。同时，嘉宾们深入探讨了“自建与外购”的权衡策略、应对供应商推销的经验、保障公众信任与“零留存”隐私架构的设计，并指出了网络安全与政府领导力跟不上技术快速迭代（约 90 天一变）的潜在风险。 ## Atmosphere & style 整体讨论气氛务实而热烈，既有专业严谨的技术与政策剖析，又不乏幽默轻松的互动。主持人与嘉宾之间默契十足，开场打趣 FOIA 催进度以及讲台遮挡视线等细节让现场氛围十分融洽。嘉宾们结合自身几十年的行业经验，直面政府内部老旧系统和漫长采购流程的痛点，展示出从传统官僚思维向“客户服务”与“产品组织”转变的积极心态，给观众带来一种兼具前瞻视野与落地可行性的深刻启发。 ## Key takeaways - [01:02](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=62s) 犹他州推行“以人为本 AI 倡议”（Pro-Human AI Initiative），将 AI 架构划分为监管与IT落地两大部分，旨在增强而非取代公职人员能力。 - [01:45](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=105s) 美国国土安全部（DHS）每年面临超 100 万份 FOIA 调阅请求，部署 AI 的核心动力是满足法定响应时限并缩短处理周期。 - [04:18](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=258s) 印第安纳州利用 AI 对缺乏文档的旧系统进行逆向工程，在 18 个月内完成云端现代化改造，并将后台工作量减少 60% 至 70%。 - [06:08](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=368s) 国土安全部通过 AI 解析并比对指纹与身份数据，消除了纸质指纹卡反复扫描的冗余流程，大幅加快移民身份核实进度。 - [08:44](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=524s) 犹他州构建了由 9 个独立 AI Agent 组成的协同审查系统，将政府合同合规分析时间缩短至 10 分钟，并为分析师节省 75% 的工作时间。 - [11:18](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=678s) 公共部门倾向于采用多云及混合模型架构，利用现有大模型解决具体业务痛点，而非自行投入巨资训练底层基础大模型。 - [13:37](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=817s) 在政企合作与软件采购上，官员偏好能够在 3 至 12 个月内快速迭代交付且具有明确合规落地路径的厂商，排斥无具体方案的泛泛推销。 - [15:39](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=939s) 基层员工对 AI 的心态已从担忧“抢饭碗”转变为渴望借助自动化工具摆脱老旧系统与繁重行政负担，从而提高服务质量。 - [19:25](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=1165s) 政府推进 AI 引入“人在环路之上”（Human above the Loop）理念，由 AI 负责自动化与建议，人类掌控最终决定权、法律责任与安全防线。 - [27:13](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=1633s) 面对服务需求增长与编制受限的矛盾，政府需要秉持财政责任，优先采用轻量化、小型化模型来降低计算算力与 Token 消费成本。 - [29:54](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=1794s) 印第安纳州将 IT 部门重构为“产品组织”，树立“客户服务”理念，致力向民众提供媲美私营部门的高品质数字政务服务。 - [33:11](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=1991s) 犹他州设计了零数据留存的州级 AI 对话架构，在保证数据仅留存在源系统的同时，防止从日志中反向追踪个人身份隐私。 - [38:56](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=2336s) 网络安全以及利用 FOIA 推演内部系统架构的“拼图理论”（Mosaic Theory）是联邦与州政府在应用 AI 时关注的重大安全隐患。 - [39:42](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=2382s) 随着技术以 90 天为周期飞速迭代，公共部门面临的最大瓶颈已从技术本身转向治理结构、采购模式与领导力的敏捷适应能力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=4s) 论坛开场与与会嘉宾介绍 本次研讨论坛由《路透社》驻白宫记者雅各布·博盖格（Jacob Bogage）主持。博盖格简要介绍了自己的报道背景以及本次讨论的主题——“公共利益与公共部门中的 AI 应用”。他幽默地称赞同台的三位专家比自己聪明得多，并依次介绍了来自州政府与联邦政府的重量级嘉宾： * **克里斯蒂安·纳皮尔（Christian Napier）：** 犹他州技术服务局（Division of Technology Services）AI 负责人。 * **罗曼·扬科夫斯基（Roman Jankowski）：** 美国国土安全部（DHS）首席隐私官兼首席自由信息法案（FOIA）官。博盖格现场打趣道，自己在 DHS 还有几个尚未处理完的 FOIA 调阅申请，如果等会儿有时间，打算当面向罗曼“催进度”。 * **罗布·福尔克（Robert Fulk）：** 印第安纳州州务卿办公室首席信息官（CIO）。 --- ### [01:02](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=62s) 各机构 AI 部署的核心目标与愿景 主持人邀请三位嘉宾各用一两句话阐述各自机构部署 AI 的核心目标与进展心态。 克里斯蒂安率先发声，他调侃现场的讲台遮挡了他将近 40% 的视线。他解释道，犹他州的 AI 架构分为两块：一端设在商务部及 AI 政策办公室，侧重于监管以及平衡安全与创新；另一端则是他所在的集中化 IT 部门，负责具体的落地实施。犹他州 AI 战略的灵魂源于州长在去年 12 月发布的“以人为本 AI 倡议”（Pro-Human AI Initiative），确保 AI 始终由人类引导、用于增强人类能力，并同时造福公职人员与普通民众。 罗曼接着介绍了联邦政府层面的挑战：他负责的 FOIA 业务是联邦政府透明化的核心，每年会收到超过 **100 万份** 调阅请求。法律严格规定官方必须在 20 个工作日内作出回应（至多额外延长 10 天）。面对庞大的业务量，DHS 部署 AI 的核心目标就是为了严格达标这一时间节点，或尽可能地缩短响应周期。 罗布则斩钉截铁地表示：“我们的使命绝不是为了用 AI 而用 AI，而是为了推动创新、降低成本并更好地服务民众。”在印第安纳州，AI 是企业现代化战略的核心工具。州务卿办公室下辖四大部门（选举、证券、汽车经销商与制造业监管、企业服务门户），服务全州近 100 万家企业。目前他们正同时对其中三个部门的四大企业级系统进行现代化改造，并将 AI 贯穿于从顶层设计到底层落地的每一个环节。 --- ### [04:18](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=258s) 印第安纳州：借助 AI 逆向工程与现代化旧系统 博盖格提到了此前与罗布长达 40 分钟的预沟通，对印第安纳州在极短时间内完成旧系统重构并嵌入 AI 表示惊叹，请罗布分享具体经验。 罗布透露，团队接手后首先将传统的“单体式”旧系统迁移至云端，并在 **18 个月内** 完成了整个系统的现代改造。由于年代久远，旧系统的文档早已丢失，相关人员离职，原开发公司历经多次转手，业务规则与逻辑全成了“黑盒”。印第安纳州利用 AI 对旧系统的数据、逻辑和业务规则进行了**逆向工程**，同时依靠 AI 辅助代码编写与自动化测试。 在面向公众的前台，他们构建了虚拟生成式 AI 许可代理，不仅能回答咨询，还能指导申请人完成许可流程并审核提交的文件；在后台，AI 则负责自动化流程并提出审核建议。这一系列举措带来了极高的降本增效成果： * **后台工作量直接减少 60% 至 70%**。 * 原本需要数年周期的系统开发，现在缩短至 **3、6、9、12 个月** 即可完成交付。 * 团队借此摆脱了繁重的表单核对与数据录入，能将更多精力投入到“白手套级”的优质政务服务中，**绝非为了裁减员工**。 --- ### [06:08](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=368s) 美国国土安全部：纸质档案数字化与移民流程提速 主持人博盖格分享了他前往 DHS 总部造访罗曼时的见闻。他回忆道，罗曼当时指着办公室的一角说“这里以前堆满了纸质文件”，即便现在仍有不便，但过去的情形更令人震撼。罗曼随后详细阐述了这些“高风险、高利害关系”的请求背景。 这些 FOIA 请求主要来自希望变更移民身份的移民人群，他们需要急切核实自己在与美国海关及边境保护局（CBP）或冰雪局（ICE）执法人员互动时登记的移民编号（A-File Number）及姓名信息。以往，申请人必须将纸质指纹卡邮寄给 DHS，工作人员扫描后再手动录入系统；后来改为允许申请人邮件发送指纹卡扫描件，结果工作人员需要先打印出来，再重新扫描进系统，反而创造了更多的“纸质垃圾”。 通过引入 AI 处理流程，DHS 实现了申请邮件接收后的即时自动解析与比对，不仅极大缩短了办理周期，也彻底改善了公众的服务体验。 --- ### [08:44](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=524s) 犹他州：智能体（Agentic）智能合规审查系统 谈及犹他州的具体应用落地，克里斯蒂安分享了一个利用多智能体解决复杂业务的典型案例——**政府合同合规审查**。 州政府每年需要与大量供应商签订合同，分析团队必须逐字核对条文，确保其符合州法律、行政法规及政策要求，过程极其耗时。起初，合同团队尝试直接使用常规的 Gemini 聊天机器人（犹他州全面接入了 Google Workspace 的 AI 体系），但由于合同条款与法规政策极其繁复，通用 Cherbot 频繁出错，导致分析师一度对 AI 产生怀疑，认为这项技术“被夸大了”。 为了打破困局，技术团队打造了一套**由 9 个独立 AI Agent 构成的协同系统**，每个 Agent 专门针对合同的某一特定领域与对应法规进行对比分析。 ``` [合同文件输入] │ ▼ ┌─────────────────────────────────────────┐ │ 多智能体协同分析系统 (9 Agents) │ ├─────────────┬─────────────┬─────────────┤ │ Agent 1 │ Agent 2 │ Agent 3...9 │ │ 法律条款审查 │ 行政法规核对 │ 内部政策校验 │ └─────────────┴─────────────┴─────────────┤ │ ▼ [自动生成高质量分析报告 (耗时约10分钟)] ``` 现在，整个分析过程仅需约 **10 分钟**，且分析结果质量极高，直接为合同分析师**节省了 75% 的工作时间**。目前，犹他州其他政府机构在看到效果后，也纷纷要求引入该系统。克里斯蒂安总结道：先让员工自由尝试发现工具局限，再针对性开发“契合特定业务（Fit-for-purpose）”的定制化方案，是推进 AI 落地的高效路径。 --- ### [11:18](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=678s) 技术选型与技术栈生态 当被问及底层供应商与技术栈的选择时，三位嘉宾展现出了极具包容性的多云与混合部署策略： * **犹他州（克里斯蒂安）：** 涵盖了 GCP、AWS 和 Azure 三大超大规模云厂商（Hyperscalers）。他们通过 AWS Bedrock 以及 GCP Vertex AI 深度调用了 Anthropic 的模型。克里斯蒂安特别提到，虽然 Token 的单价成本在上升，但自研与定制解决方案的门槛和整体成本已大幅下降。 * **美国国土安全部（罗曼）：** 幽默地透露他们“就在三天前才刚刚正式接入 Gemini”，目前正在探索中。 * **印第安纳州（罗布）：** 采用了“混合作战”策略。后台的代码生成与逆向工程大量借力 Anthropic 模型，面向公众的交互则采用 Google Vertex / Gemini 系列，同时结合 AWS 的开发工具。 罗布强调，作为政府部门，完全没有必要像私营巨头那样投入数十亿美元去训练大语言模型，应该让科技巨头去攻克底层技术，而政府机构的优势在于**深刻理解业务流程与群众需求**，利用现有大模型开发出能解决实际问题的创新方案。 --- ### [13:37](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=817s)“自建”与“外购”（Build vs. Buy）的权衡 面对政务 AI 领域经典的“自建还是外购”争议，三位嘉宾根据各自的业务特性给出了不同见解： 罗布极力主张**政企合作（Public-Private Partnerships）**。拥有 28 年私营部门从业经验的他，将创业者与创新者思维带入政府。他摒弃了传统政府 3 到 5 年的漫长采购交付周期，严格要求供应商在 **3、6、9、12 个月** 的迭代窗口内快速交付高性价比的方案，对表现不佳的供应商坚决予以清退。 罗曼指出，DHS 在 FOIA 管理软件上已外购使用了 5 年，效果符合预期。但在隐私合规管理领域，市场上现有的商业软件大多针对私营企业，无法满足联邦政府特殊的隐私保护法规要求。因此 DHS 选择**自主研发**，但同时也对市场上的优质供应商保持开放态度，毕竟完全自建的研发成本并不低。 --- ### [15:39](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=939s) 机构内部产能限制与员工心态转变 克里斯蒂安回应了在“自建与改造”过程中的团队产能挑战。他指出，技术团队的直接研发产能非常有限——他手下只有 **2 名核心工程师**，必须依靠并辅导各州级机构共计约 **500 名 IT 工程师** 才能推进落实。 同时，他也分享了州政府基层员工对 AI 心态的深刻变化： 在 2024 年试点之初，员工普遍心存恐惧与焦虑，担心“AI 会抢走我的饭碗”；而如今，这种情绪逐渐转变为“我的工作量太大了，快用 AI 帮我把繁琐的重复性劳动自动化，好让我专注于更有价值的工作”。随着工具的普及，许多非技术业务人员也开始利用 AI 进行初步的 UI/UX 构想，传统上由单一 IT 部门主导的开发工作正在向各个业务部门分散。 在讨论的最后，罗曼再次重申，部署 AI 绝不是为了裁员，而是为了确保在有限的人力资源下能够更好地应对爆炸式增长的业务需求。 <Elicitation label="查看完整研讨会第 2 部分的对谈记录" query="请继续提供该视频研讨会 Part 2 的详细逐字记录与中文分段总结。"/> ### [019:25] 人工智能在应对海量 FOIA 请求与隐私合规中的应用 面对日益增加的《自由信息法案》（FOIA）申请和隐私合规文档需求，政府部门的行政压力达到了前所未有的高度。随着各个业务部门不断推出全新的流程与系统，隐私合规的工作量呈爆炸式增长，仅靠不断招募新员工根本无法跟上这股浪潮，其人力成本也将是天方夜谭。正因如此，引入 AI 技术来减轻 FOIA 请求和隐私合规审核的堆积，从而加快行政审批进度，显得尤为迫切。如果不能及时完成审查，不仅会导致项目合同无法顺利签订，甚至还会影响到面向公众的服务与系统的上线运行，甚至带来像机场安检（TSA）长队那样的公共服务滞后。 在利用 AI 提质增效的同时，安全与合规的底线依然必须牢牢守住。无论是出于国家安全考虑，防止敏感信息泄露，还是为了保障操作安全（OPSEC），亦或是保护员工和与国土安全部（DHS）互动的公众个人身份信息（PII），都决定了必须在流程中引入“人机协同”（Human in the Loop）。 更准确地说，现在的理念正在转向“人在环路之上”（Human above the Loop）。这意味着 AI 可以用来自动化后台琐碎事务、提供智能化建议并消除冗余流程，但最终的决定权、责任和法律后果依然必须由人类承担。通过将重复性劳动交给 AI，人类官员能够被解放出来，集中精力做出更具权威性和针对性的最终决策。 - **成本瓶颈**：仅依靠增加编制无法解决指数级增长的 FOIA 请求与隐私审查，人力成本将不可承受。 - **业务延误风险**：合规审查一旦滞后，将直接阻碍政府采购合同推进及公共服务系统的按时发布。 - **人机协同演进**：提出“人在环路之上”（Human above the Loop）概念，强调 AI 负责自动化与建议，人类掌控最终决定权与法律责任。 - **安全保障范围**：涵盖国家安全（National Security）、操作安全（Operational Security）以及员工与公众的个人身份信息（PII）保护。 --- ### [021:27] 面对供应商推销：政府官员的偏好与“反向避坑”指南 论坛主持人 Jacob Bogage（来自路透社）向三位嘉宾抛出了一问题：当离开讲台后，面对络绎不绝的 AI 厂商推销，什么样的 Selling Pitch（商业推销）是真正受欢迎的，而什么样的推销又是最令人反感的？ 印第安纳州的 Rob Faulk 提出了最接地气的“反向避坑”建议：千万不要反问政府官员“你的痛点和目标是什么？我们什么都能做”。作为供应商，应该带着成熟的解决方案主动前来，明确告知产品是什么、优势何在、有何差异化特征，以及如何落地解决州政府的具体业务问题。政府官员没有时间和精力帮厂商寻找软件的适用场景。 犹他州的 Christian Napier 则强调了“人际信任”在政企合作中的核心地位。推销的形式固然重要，但能否建立良好的沟通与信任关系才是决定性的。与供应商合作若过程极其艰难，整个项目都会陷入泥潭；反之，若能找到理解政府诉求、具备高效协同能力的合作伙伴，就能实现双赢，甚至将该解决方案推广应用至更多具有相似痛点的公共部门。 国土安全部的 Roman Jankowski 表达了他对“氛围编码”（Vibe Coding）的深层担忧。很多厂商声称“只要你需要，我们什么都能做”，但这往往意味着供应商可能只是临时拼凑一套逻辑，看似比现有系统略好，但要真正接入并合规部署到政府复杂的 IT 环境中，可能需要耗费数月甚至一年的时间。相反，如果厂商能够展现出清晰的产品功能定位，并拥有与联邦政府合作的成功先例或明确的合规落地路径，便能瞬间脱颖而出。 ``` [ 供应商推销模式对比 ] 错误推销方式 (避坑) ┌─────────────────────────────────────────┐ │ "你的痛点是什么？我们可以为你做任何定制"│ └────────────────────┬────────────────────┘ │ 带来额外沟通成本与落地隐患 ▼ ┌─────────────────────────────────────────┐ │ 缺乏清晰产品定位，依靠临时拼凑或Vibe Coding│ └─────────────────────────────────────────┘ 正确推销方式 (受欢迎) ┌─────────────────────────────────────────┐ │ "这是我们的产品，它能解决州/联邦的X问题" │ └────────────────────┬────────────────────┘ │ 建立基于信任与高效协同的伙伴关系 ▼ ┌─────────────────────────────────────────┐ │ 具备成熟合规路径、差异化优势与政府合作先例 │ └─────────────────────────────────────────┘ ``` --- ### [024:38] “白手套服务”：技术赋能公职人员与提升公共利益 在公共部门，引入 AI 的核心宗旨在于服务“公共利益”（Public Good）。犹他州和印第安纳州的官员均提到了“白手套服务”（White Glove Service）这一概念——即政府应当为广大民众提供极致、高端且贴心的公共服务体验。改善政府服务不仅关乎行政效率，更蕴含着深刻的民主价值：如果政府能向公众证明自己具备高效优质的服务能力，就能极大增强社会对公共机制的信任。 如今，许多基层公职人员依然深陷上世纪 80 或 90 年代的老旧主机（Mainframe）系统中。这些公职人员本身非常优秀且充满热情，他们在日常生活中享受着银行 App、Uber 和 DoorDash 带来的便捷数字化体验，但在工作中却被迫面对繁琐低效的旧工具。他们渴望帮助每一位有需要的民众，却被沉重的行政负担压得喘不过气。因此，技术创新的核心目标是帮他们“减负”，通过降本增效将他们从无休止的机械操作中解放出来，进而提供更高质量的公共服务。 对公共透明度的追求同样体现在 FOIA 的处理上。面对外界关于“AI 自动化审查 FOIA 可能会导致信息过度隐瞒或引发灾难”的质疑，官员澄清，利用 AI 不是为了阻碍透明度，恰恰是为了加快处理速度。把原本耗时且枯燥的标准化表格交由 AI 快速处理，能够让工作人员将更多精力投入到复杂、深度的信息公开请求中，从而更快地向社会公众传递透明度。 - **基础设施现状**：部分政府部门仍在使用 20 世纪 80 至 90 年代的旧型主机系统，亟需现代化升级。 - **公职人员赋能**：通过引入现代化 AI 工具，缓解基层人员的业务过载，提升整体行政效率。 - **透明度加速**：利用 AI 快速清理单调重复的 FOIA 审查堆积，提升对公众回应的整体时效。 --- ### [027:13] 财政责任与高性价比架构设计 犹他州的数据展示了公共部门面临的严峻挑战：60年前，犹他州行政分支雇佣了约 22,000 名公职人员；而时至今日，在服务人口翻倍、服务范围大幅扩张的情况下，该州的公职人员数量依然维持在 2,0000 人左右。如果不依靠现代化技术手段，要在不增加庞大编制的前提下满足现代社会的服务需求是根本不可能的。 与此同时，政府对纳税人赋予的资金使用负有不可推卸的财政责任。犹他州拥有约 500 名软件工程师，如果这些工程师在日常工作中不去利用 AI 工具提升效率，在财政层面甚至可以说是一种“不负责任”的行为。转向使用 AI 技术、甚至采用 Vibe Coding 等现代化开发模式，是提升税收资金使用效率的必然选择。 然而，负责任地使用技术并不意味着盲目追求最昂贵的技术堆栈。在架构设计上，政府技术部门应当极力追求计算资源与 Token 消耗的高效性，遵循“打铁还需自身硬”的精准选型原则（Fit-for-purpose solutions）。绝大多数日常且具体的行政任务，并不需要动用极其昂贵的前沿大模型（Frontier Models），通过小型化、轻量化的模型同样可以高质量完成，从而为纳税人节省大量资金。 - **人员与服务对比**：犹他州在保持 2.2 万名公职人员编制不变的前提下，服务了 60 年前两倍的人口及更广泛的公共业务。 - **技术人员规模**：犹他州政府拥有约 500 名软件工程师，正全面推动 AI 辅助开发以提升产出效率。 - **算力与成本控制**：主张避免在所有场景盲目追求顶级前沿大模型，优先采用小模型处理特定任务，极大地节省计算资源与 Token 消费。 --- ### [029:54] 身份重塑与“客户”导向的现代化转变 关于如何称呼与政府打交道的对象，存在“纳税人”（Taxpayers）、“公民”（Citizens）还是“客户”（Customers）的讨论。印第安纳州积极倡导在政府内部树立“客户服务”（Customer Service）理念，并将 IT 及技术部门重新定义为“产品组织”（Product Organization）。 这一称谓的转变彻底重塑了公职人员的思维方式——无论是对内的其他业务部门，还是对外的普通民众与企业，政府团队的使命就是以更低的成本、更高效的方式，提供媲美私营部门的高质量工具与产品。传统观念认为“某个厂商从未与政府合作过就不能采用”，但现代政府应当摒弃陈规，勇敢接受新事物与新模式。 犹他州同样赞同以“客户”为中心的服务理念，但在推进数字化转型的过程中，必须牢牢守住“隐私”这一信任基石。客户在与政府互动时提交了大量极度敏感的数据，保护个人隐私是双方建立信任的基础。即使政府的出发点是出于善意（例如根据客户申请的某项服务，主动为其推荐其他可能受益的公共福利），但法律法规出于隐私保护目的，严格禁止部门间随意共享数据，政府就必须严格恪守这一边界。数据只能用于授权的特定用途，绝不能跨越红线。 - **组织定位转型**：印第安纳州将政府 IT 部门重构为“产品组织”，全面引入私营部门的标准与“客户服务”思维。 - **跨部门数据隔离**：强调即便是出于改善服务的善意目的，也必须严格遵循法律对数据共享的限制，不得擅自跨部门调阅数据。 --- ### [033:11] 隐私第一架构与多业务部门的差异化合规 为了在利用 AI 的同时绝对保障客户隐私，犹他州设计了一套统一的州级对话机器人（Conversational Agent）架构。该架构采取了极端的隐私保护策略：架构本身不存储任何关于客户的隐私数据，所有数据依然安全保留在各业务部门的源系统中。数据仅在交互瞬间流经 AI 架构进行处理，处理完毕后立即返回源系统，日志中绝不留下任何可以用来反向拼凑或推断个人身份的信息（即便是有权限的州政府员工也无法通过日志追踪到具体个人）。 ``` [ 犹他州 AI 对话机器人隐私保护架构 ] ┌──────────────────────────────────────────────┐ │ 部门源数据系统 (Source System of Agency) │ │ - 存储所有 sensitive customer data │ │ - 数据绝对保留在此，不外溢 │ └──────────────────────┬───────────────────────┘ │ 仅在交互瞬间调取数据 ▼ ┌──────────────────────────────────────────────┐ │ 统一州级 AI 对话机器人架构 (Unified Agent) │ │ - 零数据持久化 (No customer data stored) │ │ - 日志无身份关联信息 (No identity tracing) │ └──────────────────────┬───────────────────────┘ │ 处理完毕后立即归还/清除 ▼ ┌──────────────────────────────────────────────┐ │ 终端客户 │ └──────────────────────────────────────────────┘ ``` 与此相对应，国土安全部（DHS）的隐私官团队面临着更加复杂的执法与法规环境。DHS 拥有庞大的执法属性，必须在法律严格授权的范围内使用和共享数据。例如，FOIA 的申请者可能来自全球任何地方，而处理移民记录或海关信息的对象与传统美国纳税人的权益范围又有所不同。即便如此，DHS 依然在法律最大许可范围内全力保护所有相关人员的隐私。在涉及复杂敏感数据能否跨部门共享时，往往需要 4 至 5 名专业律师对具体案情进行法律梳理并提供意见，最终由首席隐私官评估法律风险与个人风险后做出决定。 印第安纳州的跨部门合规同样展现出极高的复杂性。其单个部门就涵盖了证券、汽车制造、选举、商业注册等 36 个不同的监管领域，管理着近百万家企业。不同领域的数据（从企业财务到执法卷宗）在物理层面和系统层面都有着极为严格的公私界限划分，必须针对不同的业务制定差异化的隐私保护与透明度标准。 - **零留存 AI 架构**：犹他州构建的州级 AI 平台不在中间层存储任何客户数据，日志绝不保留任何可被追踪到个人身份的标志。 - **法务风险审查**：国土安全部在决定是否共享敏感数据时，由 4-5 名专业律师组成团队提供法律评估，严控法律与个人风险。 - **多领域监管覆盖**：印第安纳州单部门监管达 36 个不同领域、近 100 万家企业，严格实行物理与系统层面的公私数据隔离。 <ElicitationsGroup message="如果您想进一步了解公共部门 AI 落地与合规的具体细节："> <Elicitation label="深入了解犹他州零留存 AI 架构的实现方式" query="请详细解释犹他州在公共部门 AI 应用中采用的零留存隐私架构与数据隔离机制。"/> <Elicitation label="对比不同政府部门在 FOIA 处理中的 AI 应用" query="对比联邦部门与州政府在处理 FOIA 请求和数据透明度时使用 AI 的策略异同。"/> </ElicitationsGroup> ### [038:40] 公众信任与系统架构设计 在讨论的尾声，嘉宾们强调了在构建政府 AI 系统时必须全盘考量各种潜在风险，例如针对《信息自由法》（FOIA）的调阅请求等问题。为了维护公众对政府机构的信任，必须从系统设计的源头做起，精心规划数据存储位置、数据库部署以及服务器的网络拓扑与地理分布，确保数据安全与隐私得到切实保障。 ### [038:56] 快速问答：AI 带来的隐忧与挑战 主持人 Jacob Bogage 提醒大家只剩下 25 秒时间，并以“快问快答”的方式抛出了最后一个引人深思的问题：“关于 AI，最让你夜不能寐的是什么？”他幽默地补充道：“我们可不需要你最深层、最黑暗的秘密”，引得全场一片笑声。 Utah 州技术服务局的 Christian Napier 率先回答，他表示目前最让他担忧的是**网络安全**。在恶意分子同样掌握并使用 AI 工具的背景下，州政府有数以千计的应用程序需要进行安全加固，以抵御日益复杂的潜在攻击。 美国国土安全部（DHS）的 Roman Jankowski 接过话头，指出**国家安全与运维安全**是他最关注的焦点。他特别提到了一种“拼图理论”（Mosaic theory）：恶意行为者可能会提出多个看似独立的 FOIA 请求，并将获取的信息拼凑在一起，从而推演解构出系统内部的工作机制，进而利用这些情报发起入侵。 ### [039:42] 领导力挑战与结语 尽管现场的红灯已经开始闪烁提示时间切断，主持人依然坚持将话筒交给了印第安纳州的 Rob Fulk。Rob 提供了不同的思考视角，他认为最让人担心的反而是**政府领导层的自满情绪**。 Rob 指出，如今技术的迭代速度远超以往，无论是私营部门还是公共领域，技术几乎每 90 天就会发生一次剧烈变革。现在“技术本身已不再是瓶颈，领导力才是”。政府部门能否建立起足够敏捷的治理结构、风险控制、隐私保护、安全防御以及采购模式，以跟上这种飞速发展的步伐并顺利部署工具，才是真正的核心痛点。 随着研讨会落幕，主持人 Jacob Bogage 感谢了三位嘉宾 Christian Napier、Roman Jankowski 和 Rob Fulk 的精彩分享，并祝愿现场观众在 AI4 大会上收获满满、尽情享受。现场在热烈的掌声中结束。 - **安全加固需求**：政府内部有数以千计的应用程序亟待加固，以防范恶意分子利用 AI 发起攻击。 - **拼图理论（Mosaic Theory）**：攻击者通过收集并组合多个 FOIA 信息请求，推演分析出底层系统架构并寻找入侵漏洞。 - **技术与治理的错位**：技术更新周期已缩短至 90 天左右，政府的采购、监管、隐私与安全治理机制必须快速跟进。 ## Notable quotes > [01:02](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=62s) 我们的使命绝不是为了用 AI 而用 AI，而是为了推动创新、降低成本并更好地服务民众。 — Robert Fulk > [08:44](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=524s) 先让员工自由尝试发现工具局限，再针对性开发“契合特定业务（Fit-for-purpose）”的定制化方案，是推进 AI 落地的高效路径。 — Christian Napier > [11:18](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=678s) 我们在三天前才刚刚正式接入 Gemini。 — Roman Jankowski > [19:25](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=1165s) 人在环路之上（Human above the Loop）。这意味着 AI 可以用来自动化后台琐碎事务、提供智能化建议并消除冗余流程，但最终的决定权、责任和法律后果依然必须由人类承担。 — Roman Jankowski > [21:27](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=1287s) 你的痛点和目标是什么？我们什么都能做。 — 未具名供应商推销代表 > [21:27](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=1287s) 只要你需要，我们什么都能做。 — 未具名供应商推销代表 > [21:27](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=1287s) 这是我们的产品，它能解决州/联邦的X问题。 — 未具名供应商推销代表 > [39:42](https://www.youtube.com/watch?v=i5DD4HsRwIw&t=2382s) 技术本身已不再是瓶颈，领导力才是。 — Robert Fulk ## Mentioned Jacob Bogage, Christian Napier, Roman Jankowski, Robert Fulk, 路透社 (Reuters), 犹他州技术服务局 (Utah Division of Technology Services), 犹他州商务部 (Utah Department of Commerce), 犹他州AI政策办公室 (Utah Office of AI Policy), 美国国土安全部 (US Department of Homeland Security / DHS), 印第安纳州州务卿办公室 (Indiana Secretary of State Office), 美国海关及边境保护局 (CBP), 美国移民及海关执法局 (ICE), Gemini, Google Workspace, Google Vertex AI, Google Cloud Platform (GCP), Amazon Web Services (AWS), AWS Bedrock, Microsoft Azure, Anthropic, 美国运输安全管理局 (TSA), Uber, DoorDash, 自由信息法案 (FOIA) ## Video description Roman Jankowski - Chief Privacy Officer / Chief FOIA Officer, Department of Homeland Security Robert Fulk - Chief Innovation and Information Officer, State of Indiana Jacob Bogage - White House Correspondent, Reuters Christian Napier - Director of AI, Utah Division of Technology Services Roman Jankowski - Chief Privacy Officer and Chief Freedom of Information Act (FOIA) Officer, US Department of Homeland Security Public sector leaders are under pressure to do more with less—and AI is quickly becoming part of the toolkit. This panel features leaders who are piloting and scaling AI to serve communities more effectively. We'll cover real deployments in fraud detection, resource planning, and public service delivery—and examine how agencies are balancing innovation with fairness, privacy, and accountability mandates. [Source](https://www.youtube.com/watch?v=i5DD4HsRwIw)

[1:02] Robert Fulk: 我们的使命绝不是为了用 AI 而用 AI，而是为了推动创新、降低成本并更好地服务民众。 [8:44] Christian Napier: 先让员工自由尝试发现工具局限，再针对性开发“契合特定业务（Fit-for-purpose）”的定制化方案，是推进 AI 落地的高效路径。 [11:18] Roman Jankowski: 我们在三天前才刚刚正式接入 Gemini。 [19:25] Roman Jankowski: 人在环路之上（Human above the Loop）。这意味着 AI 可以用来自动化后台琐碎事务、提供智能化建议并消除冗余流程，但最终的决定权、责任和法律后果依然必须由人类承担。 [21:27] 未具名供应商推销代表: 你的痛点和目标是什么？我们什么都能做。 [21:27] 未具名供应商推销代表: 只要你需要，我们什么都能做。 [21:27] 未具名供应商推销代表: 这是我们的产品，它能解决州/联邦的X问题。 [39:42] Robert Fulk: 技术本身已不再是瓶颈，领导力才是。

[0:04] Rob Fulk: not using AI for AI's sake [0:04] Christian Napier: pro-human AI initiative [3:45] Rob Fulk: white-glove service [8:44] Christian Napier: overrated [19:25] Roman Jankowski: human in the loop [19:25] Rob Fulk: humans above the loop [21:27] Roman Jankowski: vibe code [21:27] Roman Jankowski: we can do anything you want [29:54] Rob Fulk: product organization [38:56] Roman Jankowski: mosaic theory [38:56] Rob Fulk: complacency of leadership in government [38:56] Rob Fulk: technology is no longer the bottleneck; leadership is

[0:04] I'm Jacob Boggage. I'm a White House correspondent  at Reuters. Uh I cover the economic portfolio out [0:09] of the Trump administration. Uh you are at AI for  the public good, AI use in the public sector. Uh [0:15] I have three friends with me who are a lot smarter  than I am. Uh and so uh we're going to spend about [0:21] 40 minutes with maybe some time for questions  borrowing their expertise. All the way to my left, [0:27] Christian Napier, uh, who's the director of AI  at the Utah Division of Technology Services. [0:33] In the mid in the middle, uh, Roman Jenowski, the  chief privacy officer at the Federal Department of [0:39] Homeland Security and the chief foyer officer.  I do have outstanding Freedom of Information [0:44] requests uh, with DHS and if we have extra time,  I will pepper Roman about where they are. Uh and [0:50] then just to my immediate left, Rob Faulk, uh  who is the chief information officer for the [0:56] Secretary of State of the great state of Indiana.  Uh so let's start here by way of introduction, [1:02] gentlemen. And Christian, we'll start all the way  on the left and work down. Uh let's let's try to [1:06] be really brief to open up in one sentence or two.  What are your goals for AI deployment? Uh and and [1:14] how are you feeling about them? All right. Uh,  sorry I can't see like 40% of you because of this [1:21] podium here. So I'm Christian Apier, your director  of AI for the division of technology services. [1:27] We're we're kind of split into two pieces. Uh, the  role of AI and state government. So we have one [1:33] piece that sits in the department of commerce  and the office of AI policy that's focused on [1:37] regulation and balancing safety and innovation.  Uh then in the division of technology services, [1:44] we're the centralized IT shop for the state and  our team is responsible for the implementation of [1:50] AI in state government. In terms of what we want  to accomplish, a lot, but it's best articulated [1:57] by our governor's uh announcement in December of  last year of our prohuman AI initiative. And so uh [2:05] we're working to ensure that uh AI is humanguided  and is human enhancing uh and is deployed for the [2:16] benefit of both our state employees and our and  our citizens in the state. Great Roman. So, my [2:25] name is Roman Chicowski again, uh, chief privacy  officer and chief Foya officer and Foya is, [2:30] if you guys don't know, that's the transparency  side of the federal government. We get around [2:35] uh, over a million requests per year. And  unfortunately, we have to respond to them [2:40] within 20 business days. Maybe we'll get maybe 10  days extra if we can, but the law states that we [2:47] have to respond within 20 business days, maybe  additional 10. So with AI, we want to basically [2:52] be able to meet that timeline um or at least  meet as well as we can. Rob, take us home. Hey, [3:01] uh our mission is not using AI for AI's sake.  Our mission is to drive innovation, reduce costs, [3:08] and serve our citizens better. Um and AI is the  latest tool. We're able to do that and do it much [3:14] faster and much cheaper and more effectively. Um  it's embedded into our enterprise uh modernization [3:20] strategy and we've got four divisions of the  state and secretary of state. I have elections, [3:24] I have securities, um we regulate auto dealers  and manufacturers and then we are the business [3:29] portal for nearly a million businesses um across  the state of Indiana. Um and we're modernizing [3:36] three of those divisions, four different  enterprise systems um all at the same time [3:40] that AI has now come to light. So we're embedding  AI in everything we're doing from top to bottom. [3:45] Rob, let's stay with you for a minute. In our prep  calls, I really wanted to have some fun here and [3:52] talk about where you all are seeing success.  Um, and when you and I spoke individually, [3:59] um, we ended up talking for about 40 minutes on  all of these new business systems um, and portals [4:06] that you all were rolling out with AI at the core.  Can you tell us about some of those uh and and the [4:13] timeline in which they were deployed because it  was relatively rapid? Yeah, we when I came in [4:18] first thing we did was secure um some of these  legacy monolithic um systems, moved them to the [4:24] cloud and then in 18 months we modernized entire  systems. We used AI to reverse engineer the old [4:30] systems because all the tribal knowledge was gone,  the documentation was gone, the people are gone, [4:35] the company's been bought and sold three or  four times. So we had to reverse engineer the [4:39] business rules, the logic and the data of these  old systems. We used AI for that. We used AI to [4:44] help build some of the code. We're using AI to  test the code. We're building a genic AI into the [4:49] systems to automate the customer interactions. So  they actually have virtual licensing agents they [4:54] can talk to, not just a chatbot, but actually  will help them through the licensing process, [4:58] review their documentation materials. And when  we're using AI in the back office to automate [5:03] and make recommendations um of what should be done  and should these people be approved or not in our [5:08] licensing processes as well as others. So we're  we're using front to back top to bottom and it has [5:14] reduced cost driven innovation. We're seeing 67 60  to 70% reduction of workload in the back office. [5:21] Um and literally we're able to get stuff done in  3 6 9 12 months. Um entire enterprise systems we [5:26] built in 18 months. What is that freeing people  up to do that they were spending more time doing [5:32] before? It's improving the customer service.  That's one thing I want to I want to talk about. [5:37] We're not using AI to replace people. We're using  AI to get rid of the busy work that we can let [5:42] computers do. All of us, the things that they're  doing, reviewing forms, inputting data, reviewing, [5:47] let's let the computer do all that what it's  designed to do and let the human beings interact [5:50] with people. So, we're not reducing our workforce.  What we're doing is freeing up them so they can [5:55] provide that white ghost service to our citizens  and our customers. So, it's improving our human [5:59] interaction. It's improving our customer service  and improving the speed and and responsiveness we [6:05] have internally and externally with our customers.  Roman, when I visited you at DHS headquarters, [6:11] we were walking around the office a little bit and  you pointed to a corner and said, "This used to [6:15] have a lot more paper in it." Uh there was already  a lot of paper there, but I I kind of shiver to [6:22] think at what it [clears throat] did look like  once upon a time. Tell me what was I seeing? What [6:28] are you putting in practice? Uh and and again,  help us understand the benefits that you're seeing [6:36] and what you might still want to keep working on.  So what you're seeing are uh two systems colliding [6:43] at each other and not be able to talk and  communicate. Um so what he we're mentioning is in [6:49] my office uh we process uh a privacy act request  or foyer request uh in regards to immigration [6:55] records specifically for interactions with uh  CBP or ICE agents and for people to understand [7:02] what names they might have provided those agents  to see what their immigration number is. So Roman, [7:08] who are these people that are that are putting  in these requests? These are immigrants who [7:11] had encounters with immigration officials or  enforcement or who are these people cuz these are [7:17] I I don't want to underell this. These are high  stakes requests. Yeah. So these are immigrants [7:22] who are who came to this country who are probably  trying to change their immigration status. So we [7:27] they need to figure out what their immigration  number is as quickly as possible. They need to [7:32] figure out what their interactions were, what name  they provided. Sometimes you know people's names [7:36] change or they provided one name which maybe is  not the same as they provided another name when [7:40] they met somebody. So maybe they met met in  the border, maybe they talked to a TSA agent. [7:44] All these interactions all happen and they all get  documented. So they ask for an information and the [7:49] only way they get information is that they provide  us uh fingerprints. And so back in the days this [7:55] how we used to do it where somebody would mail  in their fingerprint card. We would take it and [8:00] then scan that card and put it into a different  system. And that system would uh have an output of [8:06] uh their interactions with uh the department and  uh what their a file number was. And then we said, [8:12] hey, let's modernize the system and let's send  it through email. You can send your your card [8:15] through the email. So they scan the card before  they even provide it to us. And then what we had [8:21] to do, we had to print it out and then rescan it  again. So it kind of caused even more paperwork. [8:26] Um but then uh through using AI um we're able to  solve the process pro process reduce the amount [8:32] of time it takes to uh do this. So once they send  us the email we'll be able to process right away. [8:36] So it was definitely a fantastic way to use it and  improve the customer service. And then Christian [8:44] you and I were just talking uh about a number  of implementations uh in in Utah. One of them [8:51] we talked about was contract analysis. Can you  share a little bit more with us on that? Yeah, [8:56] so uh we have a lot of contracts in the state with  a lot of vendors and whenever we get contracts in, [9:04] uh we've got an an analyst team that goes through  those contracts to make sure that the terms and [9:09] conditions of those contracts, the clauses that  are in there are aligned with uh state statute, [9:17] aligned with administrative rules and policies.  And that process is a time-consuming process. [9:23] And initially our contract analysis team  tried to use uh the regular old Gemini [9:32] uh chatbot to see if it could use Gemini because  we're a we're a Google shop and so we all the [9:38] employees have access to AI using Gemini for  Google Workspace in a ring fenced environment [9:43] and everything but it wasn't working very well.  So their attitude kind of soured like ah this AI [9:48] stuff is overrated. Uh and it wasn't a problem of  the AI necessarily. It was just hey we're asking [9:54] the AI to do something that's really complicated  because there are a lot of statutes and there are [10:00] a lot of rules and there are a lot of policies  and to get an a the kind of analysis that that an [10:07] analyst can trust is hard using some off-the-shelf  tools. So, we ended up building an agentic uh [10:14] solution that had nine different agents that were  each targeted at analyzing different areas of the [10:19] contract and comparing to specific rules. It takes  this agent, depending on the size of the contract, [10:25] maybe 10 minutes to run the agent, right? But  the analysis that comes out of it is really good. [10:33] and our contract analysts are now saving about  75% of their time as opposed to the way they [10:40] did it before. We've only implemented that in  our division, but all the other agencies that [10:46] have contracts are like, "Hey, can we do what  you're doing over there because that is cool, [10:51] right?" So that's just a very small uh you know  a very very small example of of which I think is [11:01] roll out AI to folks let them experiment  discover the limitations figure out how we [11:09] can overcome those uh those those uh limitations  by uh deploying more fitforpurpose solutions. [11:18] You mentioned you're a Gemini shop. To the  extent that that you all are able to share, [11:23] what what vendors do we like? Are we are we on  anthropic? Are we open AI? Are we experimenting [11:30] with AWS? Like what do we what do we like? So uh  we have agencies that are on all three of those [11:38] hyperscaler platforms. Uh GCP, AWS, and Azure.  So we rely heavily on our partners uh for those [11:46] platforms. We also use uh Anthropic, but we use it  through uh AWS Bedrock or through um GCP Vert.Ex [11:56] AI, which now has a really long name that I can't  remember. Uh our partners have been fantastic to [12:02] work with and there's dozens of other vendors who  who have more specific solutions that uh uh that [12:09] we've worked with. But there's an argument to  be made for seriously considering now in-house [12:16] developed solutions. Uh as has been alluded  to here, the cost to develop those solutions [12:23] has come down quite a bit. Uh even though the  cost of tokens has gone up. Uh but uh uh yeah, [12:30] for us we we rely on uh a large ecosystem of of  partners. Yeah, I I want to get to the in-house [12:36] solutions question in a minute, but real quick,  what what do we like what do we like to use? Uh [12:41] we just got Gemini just three days ago. So uh I'll  let you know. [laughter] Ours ours it's a mixture [12:47] you a lot of our back office uh code generation  reverse engineering with some of the anthropic [12:51] models. Um we're using a lot of Google uh facing  you know whatever they're called this week you [12:57] know the vertex Gemini whatever family enterprise  something or another. Um and we're also using some [13:02] AWS tools. So a lot of the big hyperscalers. Um  you know I think well the one big thing you know [13:07] I'll argue with uh you know at this point is um I  we use uh our internal resources or our vendors to [13:15] develop custom solutions. That's where the real  innovation occurs. We don't need as a government [13:19] far as I'm concerned to invest billions of dollars  that has already been done in private sector into [13:23] the hyperscalers and large language models. Let  them be the experts at that. Let us be the experts [13:28] on the business process, what our customers  need, what the state and government needs, [13:32] and we'll use those tools to develop the real  innovation solutions for our citizens. Well, [13:37] let's keep talking about that because a lot of the  conversations I've had on the sidelines this week [13:42] um and that I had with with you all in prep as  well was buy it or build it. Um and even even [13:49] private sector um what can we build on our own and  what do we need to buy? So, what are you buying? [13:55] what are you building? Um, and and as you  mentioned, Rob, that that kind of building custom [14:04] solutions, what's your process to accomplish  that in the most efficient way possible, finding [14:11] good partners that can implement um I'm a strong  believer in public private partnerships. I spent [14:16] 28 years in the private sector. This is my first  time in government, three and a half years in. Um, [14:21] you know what I've done and I've demanded. I've  removed some vendors that I came in and I wasn't [14:26] happy with their performance. I brought new  vendors in. Um, I operate in a 3 6 9 12-month [14:31] time frame, not 3 to 5 years. So, I expect high  delivery, great partners delivering things fast, [14:36] quick, cheap with AI and the tools out there.  You can do it and I've done it. I've been an [14:40] entrepreneur. I've been an innovator and we can  do the same thing in government. We can serve [14:44] our citizens better and reduce costs and provide  great services. Roman, what are you buying? What [14:49] are you building? So, we bought uh Foye management  software and we're we we've had it for around uh [14:55] 5 years now and it's uh meeting our needs. Uh but  we're always open to seeing what other vendors are [15:01] having. When it comes to privacy, unfortunately,  there's a lot of privacy management software out [15:06] there. Um but it's not specifically made for  the federal government. Um and unfortunately, [15:10] we do privacy in a different way. And so, we're  open to talking to new vendors, but at this point, [15:17] we're building our own. And worst case  scenario, if it's not working the way we want, [15:20] we'll still approach more vendors. Everything  changes. We're always open to to working with [15:25] somebody else. There's a cost associated with  building it yourself. It's not cheap. So, [15:29] we'd rather work with somebody else who uh  will just take on that cost and it'll be [15:34] easier for us to implement. Christian, you you  kind of answered this question a minute ago, [15:40] so slightly different twist on it. when you are  building or when you are taking an off-the-shelf [15:48] product and fitting it for proper use uh  for you and your team, what's the capacity [15:56] of your department or of the agencies and and  teams that you work with to really mold these [16:04] solutions to exactly what you need. The capacity  is limited. I remember when AI came onto the scene [16:16] and in 2024 when we started piloting Gemini in  various agencies, there was a lot of fear and [16:25] trepidation among the state employees that this  technology is going to come and take my job. [16:31] Some of that is still there to be honest, but  I've seen a shift and and it's more anecdotal, [16:38] but I've seen a shift in our attitude from  this technology is scary and is going to [16:45] come take my job to I'm a state employee  and I'm overwhelmed. I have so much to do. [16:54] Can this technology help me automate some of these  processes? Take some of the load off so that I can [17:01] focus on the higher value transaction. In terms  of the capacity for this to happen in the state, [17:10] we have limited capacity. I have a team of two  engineers. That's all I have. So, we have to work [17:17] with the agency IT teams. They have collectively  about 500 engineers to help them to get up to [17:25] speed on the technology and mentor them which  is challenging because it's changing so quickly. [17:33] The business users now have access to tools that  allow them to to do some of the initial thinking [17:41] about what would the user experience look like.  So there's kind of a dispersion of work that was [17:50] traditionally held uh within a single IT function  now being spread across multiple functions in this [17:56] build versus buy scenario. And then another level  of it is if we don't have expertise or capacity in [18:03] house, as uh Rob and Roman have mentioned, there's  a there's an ecosystem of partners out there that [18:12] can deliver as well. And so it's kind of a a  mixture of the of all those things. I hope that [18:20] kind of answers the question. Yeah, it absolutely  does. Roman, I I want to ask you Oh, Rob, if you [18:24] had something you want to do you want to No, from  a federal standpoint, and I'm not going to ask you [18:30] to speak on behalf of the Trump administration,  but we've seen the Trump administration pour [18:37] a lot of resources into um increasing capacity,  especially around AI. And when when you and I met, [18:48] you were telling me about when you have problems  or questions or need solutions, there are people [18:55] you can go to to say, can you build this for me?  Can you tell tell us about that? Yeah. So, um, [19:03] first, uh, one of the things that I I want to also  say that we're not firing people because of AI. It [19:11] unfortunately when it Why are you firing them?  I'm kidding. We're [laughter] That's so we're [19:16] just checking to make sure everyone's awake.  So in like I said in our office it it it's it [19:20] was so important for us to keep the people that  we have to ensure that we had employees there [19:25] who could basically process these foyer requests,  look at our privacy uh uh regulations and do our [19:31] privacy compliance documentation and uh we're even  hiring at this point. Um but unfortunately due to [19:38] the volume of a foyer request and the volume  of privacy compliance documentations because [19:43] everybody's coming up with a new process system  uh we can't keep up you know I mean the cost would [19:47] just be astronomical just to keep hiring more and  more people and so that's the need to use AI and [19:53] so we're we're doing this to decrease our volume  of foy request and privac privacy compliance [20:00] documentation and be able to provide a quicker  product because if we don't if we can't review [20:04] it in in time it the contract might not be able  to go out. Uh we have issues with hey we won't [20:11] be able to uh if there is some type of program  or system that we need to have out there for [20:18] maybe the public it's for public interaction it  it just won't be out there. Um and we'll just be [20:23] slower maybe longer lines of TSA. Um so uh we're  we're trying to do our best and like I said we're [20:29] hiring. We need to have the human in the loop. Um  and we we need to have that for multiple reasons. [20:34] once uh for and to ensure that you know for  national security reasons we don't want to [20:39] make sure information goes out that'll hurt our  national security uh other stuff operational [20:44] security and also uh PII of individuals um our  employees and uh people who interact with DHS [20:51] Rob anything you want to add here no I I I think  I just want to echo what you said you know it's [20:57] keeping you know and I I I got a new phrase where  humans above the loop in charge of it you know you [21:02] know it's actually moving to you The humans are  responsible for still for the decisions and things [21:08] being made. We can use AI to make recommendations  to automate all this work to remove all this back [21:13] office, but ultimately the humans are responsible  for the AI. They're responsible for the decisions [21:18] being done and they make the final determinations,  but we're freeing up their time and giving them [21:22] the ability to do so. You know, utilizing these  tools. So, folks, you just walked in. You're in [21:27] AI for the public good. I'm Jacob Bogage uh from  Reuters. To my left, Rob Faulk from the state of [21:33] Indiana. In the middle, Roman Jenowski from the  Department of Homeland Security. And all the way [21:37] on the left, Christian Napier, Utah, Division  of Technology Services. Uh, this is a pulse [21:42] check question for our audience. So, uh, oh, I saw  some heads pop up, so that's good. Uh, this was, [21:48] of all of the things that we have talked about,  the funnest thing for me because it gets the most [21:54] reaction out of you guys. So, let's have some fun.  Um, I guarantee you when you step off this stage, [22:00] people will come up to you and and try to talk  to you about your work. Uh, and there's a lot of [22:04] people here who have really great solutions. Don't  come talk to me. I have a flight to catch. Um, [22:12] what is a good sales pitch to you and what is a  bad sales pitch to you? What should people not [22:16] be telling you when you step off this stage? Rob,  yours is the most entertaining, so please lead us [22:23] off. Um, don't ask me what my problems and goals  are. Um, and you're here to solve them and we can [22:30] do anything for you. Repeat that. Um, do not ask  me what my goals and what my problems are. You [22:35] come to me, give me a pitch. What solution do you  have? Why is it better? How is it differentiate [22:40] itself? And how can it solve things at the state?  You come to me with a solution. Don't ask me to [22:45] to solve the problem or or to find a solution  for you where to put your your project or your [22:50] software in the state. We're going out of order.  Christian, I want to work with people that I can [22:57] trust can deliver. And so it's not necessarily  about the structure of the pitch. It's for me it [23:10] all comes down to people working with people.  So when I have an interaction with with you, [23:16] do I feel like we have a connection and you can  help me and in turn we can be helpful for you [23:24] uh because maybe we have a problem uh that you can  solve but maybe it can be solved for other people [23:30] as well the way that you solve it. So for me it's  really about establishing personal connection and [23:35] building a relationship of trust uh because that  makes all the difference in working with the [23:39] vendors. uh if if it's hard to work with them,  it's hard. It's such a struggle. So, you know, [23:49] we want to find vendors that can that can really  partner with us and work with us in a in a smooth [23:57] and an efficient way. Roman, yeah. So, uh we've  had AI partners approach us and say, "Hey, we [24:04] could provide you this software. We could provide  you this and we could do anything you want." [24:07] And my biggest fear is like, oh, there's going  to, you know, I met your whole team. Are you just [24:10] going to vibe code this and basically provide us a  product that maybe is better than what we're deal [24:16] dealing with right now, but it's going to take  us months to maybe a year to implement properly. [24:21] And you know, that's but if you provide us, you  say, hey, we have a product and this is what it [24:25] does. And you know, hey, we've worked with the  federal government previously or we we we know [24:30] we could also work with a different company to  work with the federal government. you show us a [24:34] solution that you can work with us and there's a  path forward and that we like your product. Yeah, [24:38] that's that's definitely a plus for us. One of the  reasons I love this panel is because we get to use [24:44] the words public good and it is really striking  to me as we kind of talk about these solutions. [24:52] I think both Christian and Rob when in our prep  sessions you mentioned you use the phrase white [24:57] glove service. this idea of um a a almost deluxe  delivery from the government that serves you. And [25:08] I wonder if you could talk to us a little bit  more about that aspect of service and [snorts] [25:17] the the there is a small d democracy element  to this that if we can prove we can serve you [25:24] better that is a good thing for our government  and our society. And I wonder if you could all [25:29] all you tell us about your experience with that.  I I'd echo what some of you said. You know, look, [25:35] our government workers, they're there. They're  great people. They're overburdening. They're [25:40] stuck with legacy old systems. We have some using  mainframe systems from the 80s and 90s. Look, [25:44] these are good people and they all know what  they're using. You know, they go to their bank, [25:48] they go use their Uber, they use their Door Dash.  Why can't they have those same type of tools and [25:52] stuff in their hands? As well as a citizen  interacting. um they want to do great work. [25:58] They want to talk to every customer that needs  help, but they're overburdened and overwhelmed, [26:02] most of these government workers. So, my job, as  I see it, is to drive innovation and efficiency, [26:07] reducing cost to free them up so they can  serve those citizens better um and do their [26:12] job more effectively because they are they're  overwhelmed and they're great people. Roman, [26:19] do you want to jump in anything there? Yeah, I  was just thinking about this uh a couple days [26:23] ago. I uh some publication came out and said,  "Oh, you guys are using AI for FOYA. This is, [26:29] you know, it could be a catastrophe. Are is there  going to be a human in the loop? Are you going to [26:33] be releasing withholding too much information  or, you know, and that's not going to be the [26:38] case. We're we're here to provide uh transparency  when it comes to privacy to the American public [26:43] and protect everybody's privacy. But in other at  least for our transparency side comes a foy and [26:49] privacy. We just want to process faster guys. Uh  when we see uh forms that are um very monotonous [26:57] that we could process in a very quick amount  of time. We'd rather be doing that. Um and it [27:02] would be able to provide it to the American  public faster and maybe we could basically [27:06] focus on Jacob's foyer requested after that. So  Christian, anything you want to add there? Yeah, [27:13] just kind of continuing on this theme. 60 years  ago, the state of Utah employed about 22,000 [27:22] workers uh in the state executive branch. Today,  the state of Utah employs 22,000 workers to serve [27:31] twice the population and deliver a much wider  range of services than existed 60 years ago. [27:42] You cannot do that with these old technologies.  So we have to modernize so that our citizens can [27:49] get the experience, our constituents can get the  experience that they deserve to have. One other [27:56] thing that I want to mention about this is uh we  have a responsibility to uh spend efficiently to [28:09] honor the taxpayers's commitment. Everybody's  paying money to the state government uh in a [28:16] variety of ways and and we have a responsibility  to make sure that that money is spent responsibly. [28:23] As I mentioned, we have 500 software engineers  in the state. It's irresponsible from a financial [28:32] perspective for these people to not be using  AI because they can do their jobs faster and [28:39] better. Uh, you know, vibe coding the thing or  whatever. But, uh, you know, we have to be we have [28:48] to pivot to using these kinds of technologies. Uh  so you pro part of providing good services, white [28:56] glove services to the citizen is to make sure that  they're getting good value for their money, right? [29:02] And we want to be responsible in that way, not  wasteful. Uh we want to design architectures that [29:09] are very efficient in terms of compute and token  usage. Uh and we want to uh use fitforpurpose uh [29:19] solutions where possible. We don't need to  spend huge amounts of money using the latest [29:24] and greatest frontier model for everything. Uh  we can use smaller models to accomplish very [29:30] discreet tasks and that results in significant  cost savings. So, it's kind of a wandering answer [29:36] to your question, but I think it's important to  to point out that it's not just about the enduser [29:41] experience, but everything that happens behind it  uh to ensure that that experience is delivered not [29:49] only effectively, but costefficiently as well and  responsibly. This might be a question that only I [29:54] care about as someone who covers government  operations, but I'm going to ask it anyway, [29:59] and sorry if you don't care. Um, I hear you  [clears throat] all, but also other policy [30:07] makers, federal, state, and local, kind of  interchangeably use taxpayers or customers. [30:16] Rob, tell me about because you and I talked about  this a little bit. What do you like and why does [30:22] it matter? I've been using the term customer  service and getting all my employees to do that [30:27] as well. Um and even inside my organization  I call it a product organization because that [30:31] changes their focus and mindset that we are there  to serve who are using our systems whether it's [30:36] an internal customers and the different divisions  or our external customers whether are the citizens [30:42] consistent constituents sorry and taxpayers but  it changes it that we are there to serve we are [30:49] there to provide them good service good tools  at a reduced cost in an efficient way and the [30:54] way they expect you know that and and that we are  competitive with other systems out there. Again, [31:00] there's no reason the winning government  cannot provide the same level of service, [31:04] tools, software, and technologies  that the private sector has. You know, [31:08] I got asked this I'll say that by a previous state  IT director, you know, we're separately elected, [31:13] so we run things a little different that he  said, "Well, Rob, this vendor has never worked [31:16] for the state before." And I said, "Is is that a  problem?" You know, it's a you know, it's we need [31:22] to look at things differently um and we need to  approach things differently. Christian taxpayers, [31:27] citizens, or do we not care? It's okay to say you  don't care. Yeah, I don't know. Terminology is an [31:35] interesting thing, right? Uh because different  words can have different meanings for different [31:40] people. So maybe I'll echo the the customer  focused aspect of it and use the term customer. [31:47] Uh I want to come back to something that uh Roman  was talking about though and that's privacy. [31:58] Our customers when they interact with government  entrust us with some of their most sensitive [32:09] information. And one of the underlying principles  that really anchors us is that the privacy of the [32:18] individual is paramount. We have to honor that  because that is the foundation of trust. Even [32:26] if we're trying to be benevolent by saying, "Oh,  well, uh, you know, we've had a lot of discussions [32:33] in government. Well, you have you have requested  this particular service. You could potentially [32:39] benefit from other services. But statute doesn't  allow us to share data with agencies in those ways [32:46] and there's good reason for that in a lot of there  is good reason for that right so even when you are [32:51] thinking I'm trying to do a good thing and provide  a high level of service to customers we have to [32:57] remember that first and foremost we have to honor  their privacy. We can only use the data that they [33:05] provided us for the reason they provided it and  nothing more. And so we have you know created [33:12] a a unified architecture for state uh you know  conversational agents for any agency can plug in [33:19] and use that thing. We don't store anything about  a customer in that architecture. All of that data [33:29] is stored in the source system of the agency that  provides it. It comes in here, we use it, and then [33:35] we put it back. And we store nothing in logs  that would allow anyone, even a state employee [33:43] who had access to triangulate the identity of  an individual by looking at the logs. So for me, [33:50] this is a really, really important thing. The  foundation of trust is honoring your agreements [33:57] with your customers, citizens, constituents,  whatever. Uh, and one of those agreements is we're [34:04] only going to use the information you provide for  the reason you provided it and that's it. Roman, [34:11] talk to me a little about privacy. I I  before we do that, I want to mention, [34:15] you know, there's been a lot of conversation  about privacy and and data around uh DHS and [34:21] immigration enforcement. And we should just point  out [snorts] Roman's chief privacy office does not [34:27] handle policy or even implementation. you handle  parameters um and and access and oversight. So [34:34] talk to me a little bit about um a lot of the  data that comes through your team is extremely [34:43] personal um and could have immediate real world  consequences in the wrong hands. So talk to us [34:50] about how you're you're using AI or AI is a factor  in your privacy conversations. Um that it's really [34:58] good what you're saying and it's important to  understand uh when we use our data we have to do [35:03] it for a lawful purpose. It has to be uh what is  what does regulation allow us to do and so that's [35:10] where we start off from. Well what what do our uh  delegations policies basically state and those are [35:16] the rules that we have to uh abide by or you know  sometimes there's laws that say hey we can't even [35:21] share this information with another agency even  if we really wanted to. So it's it's tough when [35:28] we're having this cuz you know you're talking  about the conversation you know are who who am [35:32] I dealing with a foyer requester? What is a foyer  requester? That could be anybody. And you're like [35:36] oh anybody in America? No it could be anybody in  the world is allowed to send us a foyer request. [35:41] So who are constituents? US you know are we trying  to save money for US tax payers? Yeah. Are are are [35:49] those the same as the foyer requesters? No. are  the people who send us uh uh who request their [35:54] immigration records uh paying into the system?  Sometimes yes, sometimes no. So, you know, to be [35:59] able to refer to them. But the one thing I want to  say, no matter what, we try to protect the privacy [36:04] of all those types of individuals um to the  greatest extent of the law and uh but you know, [36:11] we are an agency. We are a law enforcement agency.  We do have to share this information and are we [36:16] truly allowed to share it? And you know, we look  at the law, we we we check it out and we see, hey, [36:20] can we share it and can we share or not? A lot of  these decisions, they come up to me, they're like, [36:24] "Hey, can we are we allowed to share this?" We've  got three or four issues right now with sharing. [36:30] So, you were telling me that you have some uh  literally get phone calls going this individual [36:37] case, this this single discrete piece of  information, yes or no. Yeah. And so you we when [36:44] I get uh a question about that uh and I'm like  okay I I look at the fact pattern and it's just [36:51] not me who's deciding this. We have attorneys.  We have around uh four to five attorneys who [36:56] review every case when we get something like  that and they provide me a legal readout and [37:01] from there I make the decision if we be able to  uh share it, how to share it, what are the risks [37:05] of sharing it or I would tell the program offices  requesting this information, what risk there are [37:10] associated with type of sharing um legal risks,  risk for the individual. Um you know you know if [37:17] you somebody does somebody even have a warrant  after their arrest you know should we even put [37:21] post something like that you know these are some  of the questions I get Rob you have four divisions [37:26] you do manufacturing automotive elections business  registration shoot security there we securities [37:36] um tell me about the privacy requirements across  those divisions how they vary they vary all over [37:42] the place there's certain data even within  them some data is public, some's private, [37:46] and we actually demarcate it even in the systems  themselves. What is even accessible physically [37:52] in those systems and where that data is located um  that's publicly available and not and very secure. [37:57] So, we follow a lot of the same things you were  talking about. Um and it it it varies. I mean, the [38:02] securities we got everything from, you know, the  money, the investment, loan brokers, reg you know, [38:07] um investment advisors, it's 36 different things  we regulate, investigate, and enforce. We actually [38:12] have law, you know, enforcement attorneys and  and as well as um officers that go out um our [38:18] auto dealers, manufacturers. We got one, you  know, nearly a million businesses we regulate. [38:22] Um and we know everything about those businesses,  their financials, you know, where they are, where [38:26] they're located, their officers. Some of it's  public, some isn't. Um you know, we do trademarks, [38:31] UCCC's, there's a lot of things, but it, as you  talked about, they all have different regulations. [38:36] They all have different privacy requirements.  They all have different transparency requirements, [38:40] different foyer requests. So all that has to be  considered in these systems and it's even into [38:45] the systems designs where things are located,  where the databases are, where the servers are [38:50] um to protect uh that trust that the public  puts in us. We have 25 seconds. Uh let's go [38:56] really rapid fire. Uh what keeps you up at night,  Christian, about AI? We don't want your deepest, [39:03] darkest secrets. [laughter] Uh right now  it's security. Uh we've got thousands of [39:10] applications that need to be hardened in the  event that uh uh the bad actors are using [39:18] uh these AI tools and and so security is the  thing that's really keeping us up at night. Roman, [39:23] what keeps you up at night with AI? uh national  security issues, operational security. When it [39:28] comes to uh individuals taking a foyer request,  basically uh using a mosaic theory of putting [39:33] multiple foy quests together to figure out how a  system works, using that intelligence to figure [39:38] out how to break into the system. The red light's  blinking, but we're still going to go to Rob. Rob, [39:42] what keeps you up at night? Um I'd flip it. Now,  what I would say is probably the the complacency [39:48] um of leadership in government. um our ability  technology is moving faster than we've ever seen [39:53] in government as well as private sector. We're  talking cycles of 90 90 days things are changing. [39:58] Um and what our ability to have the governance,  the control, the privacy, the security, the [40:05] procurement models to adapt to that speed which  we need to do to deploy these tools is what keeps [40:11] me up at night. You know technology is no longer  the bottleneck. Leadership is um friends Christian [40:17] Napier Roman Jenowsky Rob Faulk. I'm Jacob Ogage.  Have a great AI4. Have some fun. [applause]