Panel 8 - 3: Governance Before Automation: A Practice-Based Framework for AI Adoption in

Ai4

AI-assisted summary of Panel 8 - 3: Governance Before Automation: A Practice-Based Framework for AI Adoption in — original source. Check the original for full context.

Source published: 2026-08-21

Detailed notes

## Background Published on YouTube by Ai4 on August 21, 2026, this presentation features Jason Makarow, Chief of Staff to the President and Chief AI Officer at Sacred Heart University. The session outlines a practice-based AI governance framework constructed in 12 weeks on existing infrastructure, demonstrating how organizations can align with the NIST AI Risk Management Framework to govern shadow AI and prevent operational liability.

## TL;DR Jason Makarow explains how Sacred Heart University developed an enterprise-wide AI governance framework for $114,000 in 12 weeks using native Microsoft 365 tools instead of costly external consultants. The initiative addresses both governance blind spots and underutilized software capabilities across an institutional footprint of nearly 400 applications. Grounded in organizational change theory and the NIST AI RMF, the model establishes an 'enablement over restriction' approach featuring a two-track liaison system, automated intake, and a strict 5-question vendor compliance checklist. Real-world scenarios demonstrate how human oversight catches silent vendor feature creep, prevents rogue script data leaks, and guards against algorithmic mission drift.

## Atmosphere & style The presentation carries a pragmatic, lucid, and reassuring tone that balances scholarly organizational theory with gritty operational reality. Makarow speaks with engaging candor and accessible authority, using vivid workplace scenarios—such as well-intentioned interns running local Python scripts and unexpected vendor toggles—to ground abstract risks in everyday organizational life. The talk feels empowering and refreshingly practical, transforming what is often framed as an expensive, overwhelming compliance nightmare into an achievable, low-cost operational discipline.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=4s) Sacred Heart University successfully designed and deployed an enterprise-wide AI governance framework in 12 weeks for $114,000 using existing Microsoft 365 infrastructure. - [01:59](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=119s) Legacy procurement reviews miss embedded AI risks, creating both a governance gap of unmonitored shadow tools and a capacity gap of paid AI features sitting idle without training. - [05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) With 88% of organizations deploying AI and 95% of pilots failing without governance, adoption must be treated as an organizational change challenge grounded in enablement over restriction. - [07:09](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=429s) Sacred Heart's modular governance model maps six practical components directly to the four NIST AI Risk Management Framework functions: Govern, Map, Measure, and Manage. - [10:33](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=633s) Shadow AI manifests through two primary vectors: silent vendor feature creep in core software updates and well-intentioned staff automating tasks with personal API keys. - [10:33](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=633s) Traditional SIEM and DLP security systems routinely fail to detect outbound API calls and data exfiltration from local employee scripting environments. - [13:42](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=822s) A two-track liaison model resolves enterprise stalemates by separating programmatic/academic autonomy from administrative compliance under a Chief AI Officer. - [15:56](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=956s) Vendor compliance requires contractual enforcement across five criteria, particularly mandatory 24- to 48-hour breach notification windows and binding no-training clauses. - [19:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1144s) Technically accurate AI models can still cause mission failure and algorithmic bias if demographic and equity considerations are not actively audited by human overseers. - [20:43](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1243s) Modern workforce preparation must shift from teaching basic prompt generation to developing employees who know how to critically govern AI and maintain human-in-the-loop verification. - [20:43](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1243s) Scalable AI governance does not require expensive data science teams or external software, relying instead on automated intake, centralized registries, and active committee champions.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=4s) Governance Before Automation: The Sacred Heart University Model

Jason Makarow, Chief of Staff to the President and Chief AI Integration Officer at Sacred Heart University, introduces the practical reality facing modern institutions: whether leadership realizes it or not, every organization is already utilizing artificial intelligence. Sacred Heart University, a Catholic institution in Connecticut with roughly 10,000 students, 1,300 employees, and an annual revenue between $500 million and $600 million, set out to establish strict "governance before automation" prior to rolling out any enterprise-scale AI tools. Rather than commissioning expensive external consultants or hiring an army of data scientists, the university designed and deployed a comprehensive governance framework in just 12 weeks on an operating budget of $114,000 by leveraging existing infrastructure.

Makarow highlights that in an enterprise with half a billion dollars in revenue, a $114,000 investment represents less than two-tenths of one percent of operating revenue. The initiative began as an internal research project commissioned by the university president to prevent ungoverned deployments and manage organizational risk. While Makarow’s accompanying academic paper provides formal documentation and scholarly rigor, his presentation focuses on the practical application, offering a blueprint that leaders across higher education, healthcare, financial services, and corporate sectors can implement immediately.

- **Institution Profile:** Sacred Heart University (Fairfield, CT), ~10,000 students, 1,300 employees, $500M–$600M annual revenue. - **Project Scope:** Comprehensive AI governance framework constructed in 12 weeks. - **Budget:** $114,000 total expenditure utilizing existing Microsoft 365 infrastructure (<0.02% of university revenue). - **Core Directive:** Mandate robust governance prior to deploying enterprise AI systems.

---

[01:59](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=119s) The Hidden AI Footprint: Uncovering the Governance and Capacity Gaps

To understand their baseline risk, the administration asked the Vice President of IT a fundamental baseline question: how much was the university spending on software? The answer was eye-opening—$22 million distributed across 396 distinct applications. While every single platform had passed through standard procurement checks for cost, functional utility, and technical compatibility, legacy procurement processes had never been designed to interrogate generative and autonomous technologies. Nobody was asking the critical modern questions: Which applications are executing external API calls? Where is the data payload routed? Is proprietary data hitting an unmonitored third-party endpoint? Does the vendor's terms of service grant them rights to ingest institutional data to generate model derivations? And when marketing materials claim a tool is "AI-powered," is it secretly routing prompts to OpenAI or Anthropic under the hood without contractual disclosure?

```mermaid flowchart TD accTitle: Dual Gaps in Legacy Software Procurement accDescription: Diagram showing how legacy procurement overlooks third-party API exposure and creates simultaneous governance and capacity gaps.

subgraph Audit["Software Landscape ($22M / 396 Apps)"] Legacy["Legacy Procurement Review (Cost, Functionality, Compatibility)"] Hidden["Unasked Technical Questions • External API Calls? • Payload Routing & Endpoints? • Vendor Data Rights / Training? • Undisclosed LLM Dependencies?"] end

subgraph Gaps["The Dual Gaps"] GovGap["Governance Gap 'You can't govern what you don't know about'"] CapGap["Capacity Gap 'You can't use what no one has taught you'"] end

Legacy --> Hidden Hidden --> GovGap Hidden --> CapGap ```

The university's subsequent audit revealed embedded AI components lurking across standard contract management systems, security camera networks, donor propensity modeling software, and student retention tracking platforms. The issue was not employee negligence; rather, commercial software capabilities evolved far faster than institutional procurement mechanisms could adapt. Simultaneously, the audit uncovered an equal and opposite failure: the capacity gap. Employees were actively using platforms containing paid AI features, but because nobody had received training or strategic guidance on them, those operational capabilities sat completely idle. As Makarow observes, "The governance gap and the capacity gap really became two sides of the same coin. You can't govern what you don't know about, and you can't use what no one else has yet taught you."

- **Software Audit Findings:** $22 million annual spend across 396 active software applications. - **Embedded AI Identified In:** Contract lifecycle management, campus camera systems, advancement/donor modeling platforms, and student tracking software. - **Legacy Procurement Blind Spots:** Unmonitored API calls, data payload routing, third-party model endpoints, intellectual property/derivative rights in terms of service, and undisclosed third-party foundation models (e.g., OpenAI, Anthropic). - **The Twin Gaps:** The governance gap (unmonitored risk) paired directly with the capacity gap (underutilized paid features).

---

[05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) The Change Management Imperative and Academic Foundations

Industry data underscores that ungoverned adoption is a widespread crisis. The McKinsey 2025 State of AI survey revealed that 88% of organizations now utilize AI within at least one business function. Concurrently, MIT’s Project Nanda reported that 95% of enterprise AI pilots fail when implemented without governance. Makarow stresses that overcoming these odds requires treating AI adoption not as an IT initiative, but as an organizational change challenge. The framework draws directly upon classic organizational behavior and leadership theories:

1. **Kotter’s 8-Step Change Model:** Structures the phased sequencing of implementation to ensure enterprise buy-in and sustainability. 2. **Rogers’ Diffusion of Innovations:** Informs the structure of committee members serving as active internal champions across departments. 3. **Heifetz’s Adaptive Leadership:** Separates technical architecture problems from the adaptive challenge of altering how faculty and staff evaluate daily tools. 4. **Bolman and Deal’s Four-Frame Model:** Ensures the governance structure operates simultaneously across structural, political, symbolic, and human resource dimensions. 5. **Schein’s Organizational Culture Framework:** Establishes a philosophy grounded in "enablement over restriction."

Makarow jokes about a meme circulating around Sacred Heart showing his photo with the caption "I start negotiating it," clarifying that his mandate as Chief AI Integration Officer is not to act as an institutional roadblock, but to safely enable operational transformation. As a practice-based case study rather than an experimental trial, the resulting framework provides an operating model supported by seven testable propositions for future empirical research, requiring zero new hires, zero external software platforms, and no incremental licensing fees.

- **McKinsey 2025 AI Survey:** 88% of organizations deploy AI in at least one business unit. - **MIT Project Nanda Finding:** 95% of enterprise AI pilots fail in the absence of governance. - **Operating Philosophy:** "Enablement over restriction" rather than restrictive IT gatekeeping. - **Resource Constraints:** Designed to succeed without specialized data science teams or external consulting retainers.

---

[07:09](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=429s) Architecture: Mapping the Six Components to the NIST AI RMF

To ensure rigorous compliance and broad interoperability, Sacred Heart structured its six-component governance model around the four core functions of the National Institute of Standards and Technology AI Risk Management Framework (NIST AI RMF): **Govern**, **Map**, **Measure**, and **Manage**. Every functional requirement was implemented directly inside the university's existing Microsoft 365 ecosystem, eliminating the need for bespoke enterprise software.

```mermaid flowchart TD accTitle: NIST AI RMF Alignment of the Sacred Heart Governance Framework accDescription: Mapping the four NIST AI RMF functions to the six operational components built inside Microsoft 365.

subgraph NIST["NIST AI RMF Core Functions"] direction TB G["GOVERN"] Mp["MAP"] Ms["MEASURE"] Mg["MANAGE"] end

subgraph SHU["SHU Modular Framework (Built in M365)"] direction TB C1["Committee Charter & Guiding Principles"] C2["Three-Tiered Risk Classification System"] C3["Shadow AI Inventory"] C4["Registry Data Capture & ROI Tracking"] C5["Vendor Review & Compliance Checklist"] C6["Automated Intake Workflow & Ongoing Oversight"] end

G --> C1 Mp --> C2 Mp --> C3 Ms --> C4 Mg --> C5 Mg --> C6 ```

Each component directly addresses a distinct operational governance vulnerability while fulfilling a specific cultural and behavioral role within the enterprise:

- **Govern:** Executed through a formal AI committee charter and clear institutional guiding principles that set strategic direction. - **Map:** Operationalized through a three-tiered risk classification system paired with an exhaustive shadow AI inventory to identify existing technical debt. - **Measure:** Maintained via an intake registry data capture process and dedicated ROI tracking methodology to gauge business value. - **Manage:** Enforced through standard vendor compliance reviews, automated intake workflows, and ongoing lifecycle monitoring.

---

[10:33](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=633s) Real-World Scenarios: Silent Feature Creep and the Well-Intentioned Script

To demonstrate how ungoverned "shadow AI" manifests as a present operational reality rather than a speculative future risk, Makarow outlines two scenarios common across both corporate and academic settings.

The first scenario involves silent feature creep in core enterprise platforms. An organization runs standard HR software such as Workday or Dayforce. During a routine software update, the vendor quietly introduces an automated AI module designed to screen, score, and rank job applicants. No HR manager submitted a procurement ticket or purchased new software; an HR coordinator simply noticed the toggle, assumed it was helpful, and enabled it. For months, the algorithm evaluates candidates in production while legal remains entirely unaware, no testing for disparate impact is conducted, and the institution has no idea whether applicant resumes are being ingested into the vendor's central training set.

```mermaid flowchart LR accTitle: Two Vectors of Shadow AI accDescription: Flow showing silent enterprise vendor feature creep versus internal script automation with personal API keys.

subgraph Scenario1["Scenario 1: Vendor Feature Creep"] U1["Routine HR Platform Update (Workday / Dayforce)"] --> F1["Coordinator Enables Silent AI Ranking"] F1 --> R1["Months of Production Scoring (No Legal Review / No Disparate Impact Audit)"] end

subgraph Scenario2["Scenario 2: Well-Intentioned Automation"] U2["Finance / Ops Task Assigned to Intern"] --> F2["Local Python Script + Personal OpenAI API Key"] F2 --> R2["ERP Data Exported Off-Network (Uninspected Endpoint / Training Data Ingestion)"] end ```

The second scenario stems from well-intentioned employee problem-solving. A tech-savvy recent college graduate or intern in finance or operations is asked to clean and summarize complex institutional data. Naturally leveraging skills learned in school, the employee sets up a local Python environment on their laptop, plugs in a personal API key, and writes a script that extracts bulk records directly from the enterprise ERP, passes them to a public LLM endpoint for classification, and outputs a formatted spreadsheet. The operational problem is solved efficiently, but sensitive corporate data has quietly exited the network perimeter. Standard Security Information and Event Management (SIEM) tools and Data Loss Prevention (DLP) filters rarely flag the activity without deep endpoint traffic inspection. The employee was not acting maliciously; they simply used the best tool available to complete an assigned task because leadership failed to provide a governed alternative.

- **Scenario 1 (HR Feature Creep):** Workday/Dayforce auto-update silently introduces candidate ranking; activated without legal review, bias auditing, or data privacy verification. - **Scenario 2 (Rogue Script Automation):** Tech-savvy new hires run local Python scripts using personal API keys to process ERP exports through external LLMs. - **Security Visibility Failure:** Traditional SIEM and DLP systems routinely miss outbound API calls generated from local scripting environments without granular endpoint payload inspection. - **Root Cause:** Operational staff solving business problems with unapproved tools in the absence of sanctioned, well-governed alternatives.

[13:42](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=822s) Enablement Over Restriction and the Two-Track Liaison Model

When unvetted tools or rogue scripts circulate within an organization, data frequently gets exposed without proper oversight. Rather than assigning blame to an individual or department, an effective framework catches these activities through a designated committee champion who flags the issue. The goal is "enablement over restriction"—the tool is not killed; it is governed. The organization exists not to say no, but to ensure that everyone knows what they are using and how to use it safely.

```mermaid flowchart TD accTitle: Two-Track Governance Liaison Model subgraph Stakeholders["Competing Domain Perspectives"] A["Academic / Clinical / Programmatic (e.g., Dean: Academic Priority)"] B["Compliance / Operations / Legal (e.g., Officer: Data Privacy)"] end

subgraph Mediation["Governance Mediation Layer"] C["Chief AI Officer / Integration Liaison"] end

subgraph Resolution["Governed Outcomes"] D["Valid Vendor Contracts"] E["Preserved Curriculum / Clinical Autonomy"] F["Compliance & Security Verification"] end

A --> C B --> C C --> D C --> E C --> F ```

This model directly resolves common enterprise stalemates. For example, if a dean wants to deploy an AI tutoring program for students, a compliance officer might block it because it processes sensitive student records. The dean views it as an academic decision, while the compliance officer treats it as a data decision. Both are right, and historically the victor was simply whichever party had the ear of senior leadership. A Chief AI Officer serves as the neutral liaison: domain leaders do not sign vendor contracts, and compliance officers do not dictate academic curriculum. At Sacred Heart University, this is structured as a two-track model—one track for administrative AI integration and a separate track for faculty, acknowledging faculty autonomy. This dynamic applies far beyond higher education: - **Higher Education:** Academic vs. Administrative / Staff governance - **Healthcare:** Clinical decisions vs. Administrative compliance - **Nonprofits:** Programmatic mission vs. Operational management - **Government:** Policy creation vs. Field operations

---

[15:56](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=956s) The 5-Question Vendor Compliance Checklist

Enterprise leaders frequently attend vendor demos where sleek presentation slides boast SOC 2 Type II compliance, FERPA compliance, HIPAA readiness, and promises that no customer data is used for model training. Impressed CEOs and university presidents often want to sign immediately, but organizations must systematically evaluate five foundational criteria before executing any contract:

| # | Governance Requirement | Critical Criteria & Contractual Obligations | |---|---|---| | **1** | **Data Handling Agreement** | The vendor must be contractually bound to specific regulatory roles (e.g., designated as a "school official" under FERPA, or signing a Business Associate Agreement under HIPAA). | | **2** | **Data Residency** | The contract must specify geographic location; US-based companies may host or process data overseas, triggering shifting jurisdictional obligations. | | **3** | **No-Training Clause** | Language preventing customer data from training proprietary models must appear explicitly in binding contractual terms, not merely on marketing materials or a website FAQ. | | **4** | **Breach Notification Timeline** | Vague phrases like *"as soon as practical"* or *"as soon as possible"* are completely meaningless. Contracts require defined windows (e.g., **48 hours** at Sacred Heart University; **24 hours** in higher-security environments). | | **5** | **Encryption & Audit Verification** | Verification of encryption standards at rest (**AES-256**) and in transit (**TLS 1.2+**), validated via documented architecture, SOC 2 Type II audit reports testing controls over time, and third-party penetration testing completed within the last 12 months. |

Most vendors easily pass the first three questions but fail when pressed on strict breach timelines and verifiable encryption audits.

---

[19:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1144s) Preventing Mission Failures and Algorithmic Bias

AI systems that achieve flawless technical accuracy can still fail catastrophically against an organization's core institutional mission. Because artificial intelligence does not understand organizational values, human oversight must interpret and enforce those promises:

- **Healthcare Example:** A Catholic hospital system or nursing school deploying an AI model to predict patient readmission risk may find the model was trained on data that underrepresents low-income patients. While the algorithm functions accurately for commercial insured patients, it performs poorly for Medicare and Medicaid patients, contradicting the institution's core mission to "serve all patients with dignity." - **Financial Services Example:** A financial institution or university deploying AI models to score loan or merit-aid applications may produce an algorithm that approves suburban applicants at higher rates than urban applicants. While mathematically consistent, it directly undermines commitments to equal financial access.

---

[20:43](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1243s) Forming AI-Literate Graduates and Low-Cost Governance

Organizations are entering a new phase of talent acquisition: hiring college graduates who grew up using AI intuitively, just as previous generations adopted Google search without a formal framework. While virtually every graduate knows how to write a prompt and summarize a document, organizations desperately need employees who understand how to *govern* AI, question model outputs, and maintain a rigorous "human in the loop."

```mermaid flowchart LR accTitle: Core Components of Accessible AI Governance subgraph Inputs["Zero New Budget / Existing Tech"] A["Existing IT Infrastructure"] end subgraph Governance["Operational Governance Engine"] B["Automated Intake Process"] C["AI System Registry"] D["Human Committee Champions"] end subgraph Outcomes["Institutional Value"] E["Sound & Safe Decision-Making"] F["Mission & Regulatory Alignment"] end

A --> B B --> C C --> D D --> E D --> F ```

Institutions do not need multi-million dollar budgets, massive data science teams, or complex third-party software platforms to establish enterprise-grade AI governance. Success relies on leveraging existing organizational infrastructure, deploying an automated intake workflow, maintaining a centralized registry, and embedding human committee structures rather than relying on static paper policies.

## Notable quotes > [01:59](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=119s) The governance gap and the capacity gap really became two sides of the same coin. You can't govern what you don't know about, and you can't use what no one else has yet taught you. — Jason Makarow

> [05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) Eighty-eight percent of organizations now use AI in at least one function, and 95% of enterprise AI pilots fail without governance. — Jason Makarow

> [05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) I start negotiating it — Jason Makarow

> [13:42](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=822s) enablement over restriction — Jason Makarow

> [15:56](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=956s) as soon as practical — Jason Makarow

> [15:56](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=956s) as soon as possible — Jason Makarow

> [19:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1144s) serve all patients with dignity — Jason Makarow

> [20:43](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1243s) human in the loop — Jason Makarow

## Mentioned Jason Makarow, Sacred Heart University, NIST AI Risk Management Framework, Microsoft 365, OpenAI, Anthropic, McKinsey, MIT Project Nanda, John Kotter, Everett Rogers, Ronald Heifetz, Lee Bolman, Terrence Deal, Edgar Schein, Workday, Dayforce, Google

## Video description Jason Makarow, Chief of Staff to University President/Chief AI Officer, Sacred Heart University

Eighty-eight percent of organizations now use AI in at least one function, and 95% of enterprise AI pilots fail without governance. I will present a framework for AI governance that was built in 12 weeks on existing infrastructure at a mission-driven university, aligned with the NIST AI Risk Management Framework and grounded in organizational change theory. The talk uses four scenarios drawn from healthcare, financial services, and higher education to show how six modular components, including a risk classification system, an automated intake workflow, and a committee-as-champions model, detect and govern shadow AI before it creates liability. Attendees will leave with a governance model they can adapt to any organization, a vendor compliance checklist, and a new way to think about forming employees who can govern AI, not just use it. This talk bridges academic research and operational practice for anyone responsible for AI adoption at scale.

[Source](https://www.youtube.com/watch?v=v4I_pj_zaeY)

中文简报

专题讨论 8-3：治理先于自动化：AI采纳的实践框架 Jason Makarow 阐释了 Sacred Heart University 如何利用原生 Microsoft 365 工具替代昂贵的外部顾问，在 12 周内以 11.4 万美元开发出一套全校范围的 AI 治理框架。该举措解决了机构内近 400 个应用程序中存在的治理盲区及软件功能未充分利用的问题。该模型立足于组织变革理论与 NIST AI RMF，确立了“赋能重于限制”的方法，包含双轨联络机制、自动化准入流程以及严格的 5 项供应商合规检查清单。实际案例展示了人工监督如何捕捉供应商静默的功能蔓延、防范异常脚本引发的数据泄露，并避免算法偏离使命。圣心大学（Sacred Heart University）在现有 Microsoft 365 基础设施之上，仅耗时 12 周、花费 11.4 万美元，便成功部署了全校范围的 AI 治理框架，成本不足其年营收的 0.02%。 (0:04) 对年支出 2200 万美元的 396 款软件审计发现，传统采购审查忽略了嵌入式 AI 风险，导致未受监控的“治理缺口”与因未受培训而导致付费 AI 功能闲置的“能力缺口”并存。 (1:59) 针对 88% 机构采用 AI 但 95% 无治理试点以失败告终的现状，该校将六大模块化组件与 NIST AI 风险管理框架全面对齐，确立了“赋能重于限制”的变革理念。 (5:21) 影子 AI 主要通过核心软件静默更新与员工使用个人 API Key 编写本地脚本两条途径渗透且极难被传统安全系统监测，为此该校设立首席 AI 官推行双轨联络人模型，平衡合规与学术自主。 (10:33) 供应商合规审查聚焦五大基准，不仅要求明确的数据驻留与审计证明，更在合同中强制锁定禁止数据用于模型训练的条款以及 24 至 48 小时的数据泄露通知时限。 (15:56)

Reading summary

Drawing on an internal implementation at Sacred Heart University, Jason Makarow argues that organizations must establish operational AI governance before automating workflows, demonstrating that effective oversight can be built rapidly without expensive external tools or specialized data science teams. Legacy Software Audits Expose Hidden AI Risks (1:59) Standard software procurement protocols are poorly suited to identifying modern autonomous or generative features. At Sacred Heart University, an audit of a $22 million annual spend across 396 applications revealed that legacy checks focused purely on cost, utility, and basic compatibility. As a result, embedded AI components in contract management, campus security cameras, donor propensity tools, and student retention trackers had entered the institution without evaluation of external API calls, payload destinations, or vendor rights to institutional data. Low-Cost Implementation Aligned to NIST Standards (7:09) Rather than deploying costly third-party compliance platforms or hiring external consultants, the university constructed a governance structure in 12 weeks on a $114,000 budget using its existing Microsoft 365 environment. The project, initiated by the university president, aligns six operational components—including an automated intake workflow, risk classifications, and an inventory—to the four core functions of the NIST AI Risk Management Framework: Govern, Map, Measure, and Manage. Grounding the process in organizational change frameworks such as Kotter, Heifetz, and Rogers, Makarow emphasizes that governance should focus on enablement rather than pure restriction. Operational Realities of Shadow AI Deployments (10:33) Makarow outlines two primary ways ungoverned AI infiltrates everyday operations. In one scenario, enterprise vendors quietly roll out automated features during regular software updates—such as candidate screening tools in HR platforms like Workday or Dayforce—which staff enable without legal review, bias testing, or awareness of underlying training practices. In another scenario, employees write local scripts and use personal API keys to process internal enterprise data through external model endpoints. In both instances, standard security monitoring tools often miss the data exposure, as staff are simply attempting to solve operational problems in the absence of sanctioned institutional pathways. Vendor Due Diligence and Mission Alignment (15:56) To counter misleading vendor assurances regarding data safety and regulatory compliance, the framework mandates a five-point checklist: binding data handling agreements, defined geographic data residency, explicit contractual no-training clauses, strict breach notification windows (such as 48 hours), and verifiable encryption standards (AES-256 at rest, TLS 1.2+ in transit) backed by current SOC 2 Type II audits. Beyond legal safeguards, human oversight is required to catch algorithmic failures that conflict with institutional missions, such as readmission risk models that underrepresent low-income patients or lending algorithms that skew against urban applicants. The Two-Track Liaison Model (13:42) To resolve institutional impasses between operational leaders wanting functionality and compliance officers flagging risk, the framework introduces a neutral liaison function. Sacred Heart implemented a two-track governance model that separates administrative integration from faculty initiatives to preserve academic autonomy. Makarow notes that this structural separation prevents deadlocks across various sectors—such as balancing clinical autonomy against administrative oversight in healthcare, or programmatic goals against operational compliance in nonprofits. The Sacred Heart University model serves as an empirical case study rather than a controlled experiment, leaving open how effectively its lightweight, seven-proposition framework adapts across heavily regulated non-academic environments with differing scale and risk appetites.

精炼阅读摘要

圣心大学校长幕僚长兼首席AI集成官Jason Makarow指出，各类机构普遍低估了内部已存在的AI足迹；该校在未增加新人员编制与外部咨询的前提下，耗时12周、仅投入11.4万美元既有基础设施预算，成功构建起对齐NIST标准的治理架构，强调自动化落地前必须实行严格治理并贯彻“赋能优于限制”的原则。软件资产审计与治理能力双重缺口 (1:59) Makarow指出，多数组织并未意识到日常运营早已被人工智能渗透。圣心大学对年支出达2200万美元的396款在用软件进行审计后发现，传统采购审查仅核验成本与功能，完全未能审查现代生成式技术的数据流动。大量已购系统缺乏对外部API调用端点、敏感数据载荷流向、供应商训练衍生权利以及是否暗中调用第三方大模型等关键合规问题的把关。审计显示，合同管理、安防监控系统、捐赠人倾向分析及学生留存预测软件中均已内嵌AI模块。这种审查缺位同时引出了“治理缺口”与“能力缺口”。一方面，未被发现的嵌入式AI在缺乏监督的情况下运行；另一方面，员工虽然正在使用带有高阶AI功能的付费软件，却因从未接受指导而导致相关功能常年闲置。麦肯锡2025年调研显示88%的企业已在业务中部署AI，而MIT Project Nanda研究表明，缺乏有效治理的AI试点失败率高达95%。治理与能力互为表里：机构无法治理未知的工具，员工亦无法应用无人传授的技术。影子AI蔓延：静默功能与自制脚本 (10:33) Makarow列举了两类普遍存在于现代组织中的“影子AI”典型场景。第一类是核心企业平台的“静默功能蔓延”。例如机构长期使用的Workday或Dayforce等主流人力资源软件在常规版本更新时，静默上线了AI简历筛选与排序功能。HR人员随手开启开关后，该算法在未经过法律部门审核、未开展歧视性影响（disparate impact）审计、且未明确求职者简历是否会被供应商用于模型训练的情况下，在线上持续对候选人打分长达数月。第二类风险来自技术熟练员工的善意自发解决行为。例如财务或运营部门的新入职员工或实习生为完成复杂报表整理，在个人电脑配置本地Python环境并填入个人API密钥，直接从企业ERP系统批量导出机密数据传输至公网LLM端点处理。由于传统安全信息和事件管理（SIEM）与数据防泄漏（DLP）系统在缺乏细粒度端点检查时极易漏判此类常规出站API流量，敏感数据已实质性越出网络边界。这一行为并非出于恶意，而是管理层未能及时提供合规官方替代方案导致的后果。低成本架构落地与双轨中介机制 (7:09) 为消除企业级AI治理需要数百万美元及专职数据科学团队的误区，圣心大学设计了映射美国国家标准与技术研究院AI风险管理框架（NIST AI RMF）的六组件模型。该体系完全构建于机构既有的Microsoft 365办公套件内部，包含委员会章程、三级风险分类标准、影子AI盘点清单、进件数据登记与ROI追踪、供应商合规清单以及自动化进件工作流，全程未采购任何定制软件，总支出仅占该校年收入的不到万分之二。在协同治理层面，该方案采用“赋能优于限制”的变革思路，并设立双轨联络机制。当业务部门（如学院院长推进AI学生辅导）与合规部门（如合规官担忧隐私泄露）产生决策僵局时，首席AI集成官充当客观协调者：专业领导不越界签订合同，合规人员亦不强行干预课程设置。该校通过区分学术探索与行政管理两条治理轨道，在维护学术与专业自主权的同时，确保技术落地符合整体数据与安全红线。供应商五项合规核查与使命防偏 (15:56) 面对软件供应商在演示中普遍做出的合规宣称，该框架确立了合同签署前必须系统性核验的五项底线：其一，签订具有法律效力的数据处理协议并明确监管角色（如FERPA指定学校官员身份或HIPAA业务伙伴协议）；其二，明确数据托管与计算处理的实际地理驻留位置；其三，必须在合同正文中具备不可撤销的“禁止将客户数据用于专有模型训练”条款，拒绝仅停留在营销宣传页的口头承诺；其四，设定具体明确的数据泄露通知时限（如该校要求的48小时或敏感部门的24小时），排除“尽早通知”等模糊字眼；其五，核验静态AES-256与传输中TLS 1.2+加密标准，并调阅过去12个月内的第三方渗透测试与SOC 2 Type II实效报告。此外，Makarow强调技术精准度不等于机构使命一致性。例如医疗系统若采用未充分涵盖低收入群体的模型预测再入院率，算法虽在商保患者上表现优良，却会在低保人群中失效；金融或高校模型在数学打分一致的情况下可能系统性倾向郊区生源。如果缺乏“人在回路”的价值监督，技术逻辑便会侵蚀机构的核心公益与服务承诺。圣心大学的实践表明，企业级AI治理并非难以企及的庞大工程，依赖现有的基础软件、自动化工作流与跨职能委员会便可在低预算下实现闭环。随着具备天然AI工具使用直觉的毕业生逐步进入职场，如何促使员工从单纯的提示词使用者转变为具备算法审视意识、能够识别模型偏差并维护安全底线的复合型治理人才，仍是各行业面临的长期挑战。

Detailed notes — zh

## Background 该视频由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。页面简介显示，圣心大学校长办公室主任兼首席 AI 整合官 Jason Makarow 分享了该校如何在 12 周内依托既有基础设施，结合 NIST AI 风险管理框架与组织变革理论，搭建起低成本且高度可复用的 AI 治理体系，以应对影子 AI 风险并防止企业级 AI 试点失败。 ## TL;DR 演讲围绕“自动化之前先治理”（Governance Before Automation）这一核心理念，分享了圣心大学仅耗资 11.4 万美元在 12 周内基于 Microsoft 365 搭建轻量化 AI 治理架构的实战经验。内容深入剖析了组织中因缺乏可见性而导致的治理断层与能力断层，并通过 HR 隐形功能更新与员工个人脚本调用等典型场景揭示了影子 AI 的数据泄露隐患。框架融合了 NIST AI RMF 与经典组织变革理论，推行“赋能重于限制”与双轨协调机制，并提出了评估第三方供应商合规性的五大审查原则。此外，演讲强调防范算法偏见导致的“使命违背”，倡导组织将重点从单纯培养 AI 使用者转向培养具备质疑精神与治理思维的人才。 ## Atmosphere & style 整场演讲风格严谨务实、直击痛点，彻底摒弃了技术炒作，将高大上的 AI 议题拉回到具体的组织运营与合规现实中。演讲者条理清晰、叙事沉稳，将抽象的管理理论与生动切近的院校及医疗真实案例自然缝合。整段分享洋溢着制度理性与人文关怀并存的基调，既有严密的安全风控视角，又充满了对组织文化重塑与使命担当的深刻洞察。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=4s) 圣心大学在 12 周内仅花费 11.4 万美元，完全基于既有 Microsoft 365 架构成功搭建了企业级 AI 治理框架，证明 AI 治理无需高昂预算或庞大数据科学团队即可低成本落地。 - [01:59](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=119s) 组织在软件采购中普遍存在'治理断层'与'能力断层'：既有应用深度潜藏未经审查的第三方 API 与外部模型调用，同时大量已付费的内置 AI 功能因缺乏培训而严重闲置。 - [05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) 根据行业调研，88% 的组织已在业务中使用 AI，但若缺乏完善的治理机制，高达 95% 的企业级 AI 试点项目都将以失败告终。 - [05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) 圣心大学将 NIST AI 风险管理框架的治理、映射、度量、管理四大功能，与 Schein、Kotter 等经典组织变革理论深度融合，构建出赋能型的运营治理生态。 - [10:33](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=633s) 典型的影子 AI 风险往往源于非恶意的日常业务场景，如既有 HR 软件静默更新加入未经评估的算法筛选功能，或技术员工私自编写脚本调用个人 API Key 处理企业核心数据。 - [13:42](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=822s) AI 治理的核心理念是'赋能优于限制'（Enablement over restriction），通过设立各部门倡导者机制，将影子 AI 纳入合规轨道，而非一味封杀业务创新需求。 - [13:42](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=822s) 设立首席 AI 官（CAIO）与双轨协调机制，有助于化解业务/学术部门与法务/合规部门之间的管辖死锁，在保障合规的同时不干涉一线专业自主权。 - [15:56](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=956s) 供应商合规审查必须严格核实五大要点：具法律约束力的数据处理协议、数据驻留地理位置、免训练条款入约、明确的泄露通知时限（24-48小时）以及第三方安全审计实证。 - [19:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1144s) 技术指标完美的 AI 模型仍可能因样本代表性偏差导致'使命违背'（Mission Failure），必须依靠人工介入（Human-in-the-loop）来捍卫组织的价值观与公众承诺。 - [20:43](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1243s) 企业当前最匮乏的不是会用 AI 写 Prompt 的操作人员，而是懂得主动质疑模型输出、理解人类回路价值并具备 AI 治理思维的复合型人才。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=4s) 圣心大学的 AI 治理背景与低成本落地实践圣心大学（Sacred Heart University）是一所位于康涅狄格州的天主教大学，拥有约 10,000 名学生及 1,300 名教职员工，年收入在 5 亿至 6 亿美元之间。校长办公室主任兼新任首席 AI 整合官（Chief AI Integration Officer）Jason Makarow 指出，当今任何组织内部其实都已经在以某种形式使用 AI，关键在于管理层是否知情。为了避免盲目上线企业级 AI，大学决定在全面自动化之前优先建立治理机制（Governance Before Automation）。这项治理体系在短短 12 周内搭建完成，仅花费 11.4 万美元，完全基于现有的 Microsoft 365 基础设施，没有聘请外部咨询公司，也没有组建庞大的数据科学团队。对于一所年营收超 5 亿美元的机构而言，11.4 万美元的投入占比不到 0.023%，极具成本效益。这套框架不仅适用于高等教育，同样可直接复用到各类面临资源约束的企业与非营利机构中。 - **机构规模与财务**：约 10,000 名学生，1,300 名员工，年收入 5 亿至 6 亿美元。 - **治理启动成本**：耗时 12 周，总预算 114,000 美元，依托既有 Microsoft 365 架构。 --- ### [01:59](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=119s) 软件资产盘点与被忽视的 AI 盲区：治理与能力的双重断层在启动治理前，校方向 IT 副总裁了解全校软件采购情况，结果显示大学在 396 款应用程序上每年支出高达 2,200 万美元。尽管所有工具均通过了常规采购流程并评估了功能、成本与兼容性，但过去从未有人提出过针对 AI 的核心安全问题：哪些工具正在发起外部 API 调用？数据负载（Data Payload）被传输到了哪里？是否存在缺乏可见性的第三方端点？服务条款是否允许供应商使用大学数据进行衍生模型训练？营销页面标称“AI 驱动”的产品，底层是否悄悄调用了 OpenAI 或 Anthropic 的接口却未在合同中披露？深入排查后发现，从合同管理软件、校园安防摄像头系统，到捐赠者预测建模平台及学生追踪工具，几乎每个系统内部都嵌有 AI 组件。这暴露出两大核心矛盾： 1. **治理断层（Governance Gap）**：工具迭代的速度远快于采购与审查流程的更新步伐，导致未受监管的“影子 AI”随处可见。 2. **能力断层（Capacity Gap）**：组织已在为大量内置 AI 特性付费，但员工从未接受过培训，没有人探索如何用它们解决实际业务痛点，导致资产严重闲置。正如演讲者所言：“你无法治理你一无所知的事物，也无法使用无人传授的工具。”（"You can't govern what you don't know about and you can't use what no one else has yet taught you."） ```mermaid flowchart TD accTitle: AI Adoption Dual Challenge subgraph GapAnalysis["企业 AI 落地的双重困境"] A["现有软件资产 396 款应用 / 2200 万美元"] --> B["治理断层 (Governance Gap)"] A --> C["能力断层 (Capacity Gap)"] B --> D["隐蔽的外部 API 调用 第三方端点数据外泄 未经披露的模型训练"] C --> E["已付费功能未启用 员工缺乏操作培训 实际业务痛点未解决"] end ``` --- ### [05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) 行业数据与理论根基：将组织变革理论注入 NIST 框架根据麦肯锡（McKinsey）2025 年 AI 现状调查，88% 的组织已在至少一个业务职能中使用 AI；然而麻省理工学院（MIT）的 Project Nanda 报告指出，**如果没有完善的治理，95% 的企业级 AI 试点项目都将以失败告终**。这表明影子 AI 不是未来的潜在风险，而是当下的既成事实。圣心大学的框架将 NIST AI 风险管理框架（NIST AI RMF）的四大核心功能，与经典组织变革理论进行了深度融合： - **治理（Govern）**：基于组织章程与指导原则，结合 Schein 的组织文化理论，主张“赋能优于限制”（Enablement over restriction），避免治理部门成为阻碍创新的“说不部门”。 - **映射（Map）**：建立三级风险分类体系并清点影子 AI 资产，借助 Kotter 的变革八步法规范实施路径。 - **度量（Measure）**：通过资产登记库捕获数据并跟踪 ROI。 - **管理（Manage）**：实现供应商合规审查、自动化接入工作流与持续合规追踪。此外，该体系还融入了 Rogers 的创新扩散理论（构建委员会拥护者机制）、Heifetz 的适应性领导力框架（区分技术架构与人员评估习惯的适应性挑战），以及 Bolman & Deal 的四框架模型（从结构、政治、象征和人力资源四个维度协同推进）。该实践案例并不声称已彻底消除风险，而是提供了一套经过验证的运营模型与 7 项可供实证检验的研究命题。 --- ### [10:33](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=633s) 典型影子 AI 风险场景：隐形更新与个人脚本调用为了具象化治理盲区，演讲者分享了两个极具代表性的现实场景： 1. **HR 系统静默更新引发的合规与偏见风险**：组织使用的 HR 软件（如 Workday 或 Dayforce）在一次常规系统更新中悄悄加入了“AI 自动筛选与候选人排名”功能。HR 协调员觉得实用便直接开启，该功能随之对求职者打分筛选了数月。法务部门不知情、未对算法模型进行审查，更没有人评估过训练数据是否存在歧视或差异性影响（Disparate Impact），亦不清楚求职者隐私数据是否被供应商挪用于二次训练。这并非员工恶意违规，而是既有工具更新脱离了监管视线。 2. **业务端“技术能手”自写 Python 脚本导致的数据外泄**：财务或运营团队招募了刚毕业的大学生或实习生，遇到复杂繁琐的数据处理需求时，便让这些懂技术的年轻人“想想办法”。实习生在个人笔记本上配置 Python 环境，填入个人购买的 API Key，编写脚本直接从企业 ERP 中抓取敏感记录，发送给 GPT 模型做总结与分类，最后导出一份漂亮的 Excel 表格交差。在此过程中，企业核心数据已悄然流出内网并可能被用于大模型训练。由于多数机构未在终端深度检测应用层流量，传统的安全信息与事件管理（SIEM）和数据泄露防护（DLP）系统对此完全失效。年轻人出于好心解决业务问题，却在不知不觉中造成了严重的数据合规漏洞。 ```mermaid flowchart LR accTitle: Shadow AI Data Leak Scenario subgraph Scenario["非恶意影子 AI 泄漏路径"] direction LR A["ERP 核心数据库"] -->|提取业务数据| B["员工本地 Python 脚本"] B -->|调用个人 API Key| C["第三方大模型 (如 GPT)"] C -->|返回分类/摘要| D["业务 Excel 表格交付"] C -.->|数据可能被留存训练| E["外部未知风险"] end ``` ### [13:42](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=822s) 赋能重于限制：双轨协调与治理影子 AI 当部门内部缺乏统一的监管时，员工往往会在不知情的情况下使用不合规的工具或泄露敏感数据。这并非某个个人或部门的过错，而是缺乏治理体系的结果。在圣心大学（Sacred Heart University）落地的这套框架中，各部门的“治理倡导者”（Champion）能够及时发现这些影子 AI（Shadow AI）并将其上报，随后由 AI 负责人介入对接。一旦发现未经审批的工具或后台脚本，治理团队的目标并不是直接将其关停，而是将其纳入合规治理轨道，即践行“赋能重于限制”（Enablement over restriction）的理念：“我们的目的不是说‘不’，而是确保你清楚自己正在使用的是什么。” 在实际运行中，不同部门之间常会因权限与职责产生死锁。例如，某学院院长希望为学生部署一套 AI 辅导程序，但合规官因其涉及学生档案数据而予以阻拦；院长坚称这是教学学术决策，合规官则强调这是数据合规决策。两者各执一词且互无管辖权，传统模式下往往取决于谁在最高管理层面前更有话语权。而首席人工智能官（Chief AI Officer, CAIO）的设立则充当了两者的桥梁：院长无需处理繁琐的供应商数据条款，合规官也不干涉教学大纲制定。针对高校中行政与教研人员的天然壁垒，圣心大学设立了行政整合与教职人员专向的双轨对接机制，避免出现行政人员直接干预教职员工教学治理的摩擦。这种双轨协调架构同样适用于其他行业： - **医疗健康行业**：临床一线（Clinical）与行政管理（Administrative）之间的权限平衡。 - **非营利组织**：项目执行（Programmatic）与组织运营（Operational）之间的协调。 - **政府公共部门**：政策制定（Policy）与一线外勤（Field Operations）之间的衔接。 ```mermaid flowchart TD accTitle: 双轨 AI 治理与协调流程 accDescr: 展示首席 AI 官如何在业务需求与合规审查之间充当桥梁，实现赋能而非限制 subgraph Input["业务诉求与发现"] A["业务方发起 / 倡导者发现影子 AI"] end subgraph Coordination["双轨协调中枢 (CAIO)"] B{"Chief AI Officer\n跨部门协调"} C["业务/学术考量\n(如：教学价值、临床效用)"] D["合规/安全考量\n(如：数据合规、系统集成)"] end subgraph Output["落地结果"] E["纳管上线：赋能重于限制\n(签订条款、安全加固)"] end A --> B B --> C B --> D C --> E D --> E ``` --- ### [15:56](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=956s) 供应商合规性评估：必须提出的五大关键问题当供应商向管理层展示炫酷的 AI 演示文稿，并罗列出 SOC 2 认证、符合 FERPA（家庭教育权利与隐私法案）、兼容 HIPAA（健康保险可携性与责任法案）且声称“不用于模型训练”时，企业高管极易被说服并急于签约。然而，在正式签署合同前，必须严格审查以下五个核心问题。这套问题最初源于高校的 FERPA 合规需求，但完全适用于 HIPAA、PCI DSS、GDPR 以及各类州级隐私法案和企业内部机密制度： ```mermaid flowchart LR accTitle: 供应商 AI 合规审查五步法 accDescr: 评估第三方 AI 供应商必须核实的五个关键合规维度 subgraph CoreLegal["法律与数据约束"] Q1["1. 数据处理协议\n(具约束力的法定身份)"] Q2["2. 数据驻留权\n(明确存储地理位置)"] Q3["3. 免训练条款\n(写入正式合同而非FAQ)"] end subgraph TechSecurity["安全响应与加密标准"] Q4["4. 泄露通知时限\n(明确规定 24-48 小时)"] Q5["5. 加密与审计实证\n(AES-256 / TLS 1.2+ / SOC 2)"] end CoreLegal --> TechSecurity ``` 1. **数据处理协议（Data Handling Agreement）**：供应商是否在合同中承担法定义务？在高校场景中，供应商必须在合同中承诺以 FERPA 项下的“学校官员”（School Official）身份行事；在医疗场景中，则必须签署 HIPAA 商业伙伴协议（BAA）。 2. **数据驻留（Data Residency）**：明确数据存储的具体地理位置。即便与本土实体签约，其服务器与数据节点也可能分布在境外，进而引发跨境数据管辖与合规风险。 3. **免训练条款（No Training Clause）**：严禁供应商将客户数据用于基础模型训练。此项承诺必须白纸黑字写入正式法律合同，仅写在网站的常见问题解答（FAQ）页面上没有任何法律保障。 4. **数据泄露通知时限（Breach Notification）**：合同中绝不能接受“在可行情况下尽快”（as soon as practical / possible）这类毫无约束力的模糊措辞。必须规定明确的时间窗口，圣心大学的标准是 48 小时以内，部分高敏机构则要求 24 小时以内。 5. **加密标准与审计实证（Encryption & Audits）**：静态存储必须达到 AES-256 加密标准，传输过程中必须采用 TLS 1.2 或更高版本。供应商不仅要口头声明，还需提供架构文档、近 12 个月内的第三方渗透测试报告，以及证明控制措施持续有效的 SOC 2 Type II 审计报告。 - 许多供应商能满足前三项要求，但在最后两项（具体泄露通知时限与完备的安全审计证据）上往往无法达标。 --- ### [19:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1144s) 技术指标 vs. 组织使命：防范“使命违背”风险一个 AI 工具即使在所有技术性能指标上表现完美，仍可能导致严重的“使命违背”（Mission Failure）。AI 模型本身不具备理解价值观的能力，必须依靠人工治理来确保算法决策与组织承诺保持一致： - **天主教医院系统案例**：该系统引入预测患者再入院风险的 AI 模型。由于训练数据集中低收入与少数族裔样本代表性不足，模型对购买商业保险的患者预测极准，但对医保（Medicare）患者的预测效果极差。该工具虽然通过了纯技术层面的性能测试，却直接违背了医院及护理学院“以尊严服务所有患者”的核心使命。 - **金融服务与高校优异奖学金案例**：某金融机构（同类逻辑也出现在大学发放优秀奖学金的评分模型中）部署 AI 进行贷款资格审查，结果模型批准郊区申请人的比例显著高于城市申请人。算法在数学统计上是自洽且准确的，但直接击碎了组织章程中“提供公平金融机会”的承诺。 > “AI 根本不懂你的组织使命，必须由人来负责阐释。如果你的使命宣言对公众做出了承诺，AI 治理就必须成为捍卫这一承诺的防线。” --- ### [20:43](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1243s) 培养能治理 AI 的人才与轻量化落地实践当前企业招聘的应届毕业生，使用 AI 工具的方式就像上一代人使用 Google 搜索一样自然直观。如今几乎每位毕业生都会编写 Prompt 和提炼文档摘要，但企业真正匮乏的是“懂得如何治理 AI 的人才”——即在采取行动前主动质疑模型输出、深刻理解“人类在回路中”（Human-in-the-loop）价值的人。这种思维并非为了应付监管，而是帮助组织做出优质决策而非仅仅是快速决策的关键。圣心大学将 AI 治理理念贯穿学生从入学申请到毕业的全过程。在实践落地层面，建立 AI 治理体系并不需要高昂的预算投入、庞大的数据科学团队或采购全新的软件平台： - **依托现有环境**：盘点并利用企业现有的基础设施。 - **核心两大抓手**：建立清晰的工具登记册（Registry）与自动化进件审查流程（Intake Process）。 - **重构组织网络**：在组织架构中确立治理专人和倡导者机制，让治理扎根于具体流程而非停留在纸面文件。相关实践经验已总结为基于组织领导力与变革理论的学术论文，详细探讨了无专业数据科学团队的使命驱动型组织，如何在多学院、多部门和广大师生之间构建行之有效的 AI 治理生态。 ## Notable quotes > [01:59](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=119s) 你无法治理你一无所知的事物，也无法使用无人传授的工具。 — Jason Makarow > [05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) 如果没有完善的治理，95% 的企业级 AI 试点项目都将以失败告终。 — Jason Makarow > [05:21](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=321s) 赋能优于限制 — Jason Makarow > [13:42](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=822s) 我们的目的不是说‘不’，而是确保你清楚自己正在使用的是什么。 — Jason Makarow > [15:56](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=956s) 在可行情况下尽快 > [19:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1144s) 以尊严服务所有患者 > [19:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1144s) 提供公平金融机会 > [19:04](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1144s) AI 根本不懂你的组织使命，必须由人来负责阐释。如果你的使命宣言对公众做出了承诺，AI 治理就必须成为捍卫这一承诺的防线。 — Jason Makarow > [20:43](https://www.youtube.com/watch?v=v4I_pj_zaeY&t=1243s) 人类在回路中 — Jason Makarow ## Mentioned Jason Makarow, 圣心大学 (Sacred Heart University), Microsoft 365, OpenAI, Anthropic, 麦肯锡 (McKinsey), 麻省理工学院 (MIT), Project Nanda, NIST AI 风险管理框架 (NIST AI RMF), Edgar Schein, John Kotter, Everett Rogers, Ronald Heifetz, Lee Bolman, Terrence Deal, Workday, Dayforce, GPT, FERPA, HIPAA, PCI DSS, GDPR, SOC 2, AES-256, TLS 1.2, Medicare ## Video description Jason Makarow, Chief of Staff to University President/Chief AI Officer, Sacred Heart University Eighty-eight percent of organizations now use AI in at least one function, and 95% of enterprise AI pilots fail without governance. I will present a framework for AI governance that was built in 12 weeks on existing infrastructure at a mission-driven university, aligned with the NIST AI Risk Management Framework and grounded in organizational change theory. The talk uses four scenarios drawn from healthcare, financial services, and higher education to show how six modular components, including a risk classification system, an automated intake workflow, and a committee-as-champions model, detect and govern shadow AI before it creates liability. Attendees will leave with a governance model they can adapt to any organization, a vendor compliance checklist, and a new way to think about forming employees who can govern AI, not just use it. This talk bridges academic research and operational practice for anyone responsible for AI adoption at scale. [Source](https://www.youtube.com/watch?v=v4I_pj_zaeY)

Source quotes

[1:59] Jason Makarow: 你无法治理你一无所知的事物，也无法使用无人传授的工具。 [5:21] Jason Makarow: 如果没有完善的治理，95% 的企业级 AI 试点项目都将以失败告终。 [5:21] Jason Makarow: 赋能优于限制 [13:42] Jason Makarow: 我们的目的不是说‘不’，而是确保你清楚自己正在使用的是什么。 [15:56] 在可行情况下尽快 [19:04] 以尊严服务所有患者 [19:04] 提供公平金融机会 [19:04] Jason Makarow: AI 根本不懂你的组织使命，必须由人来负责阐释。如果你的使命宣言对公众做出了承诺，AI 治理就必须成为捍卫这一承诺的防线。 [20:43] Jason Makarow: 人类在回路中

Source quotes

[1:59] Jason Makarow: The governance gap and the capacity gap really became two sides of the same coin. You can't govern what you don't know about, and you can't use what no one else has yet taught you. [5:21] Jason Makarow: Eighty-eight percent of organizations now use AI in at least one function, and 95% of enterprise AI pilots fail without governance. [5:21] Jason Makarow: I start negotiating it [13:42] Jason Makarow: enablement over restriction [15:56] Jason Makarow: as soon as practical [15:56] Jason Makarow: as soon as possible [19:04] Jason Makarow: serve all patients with dignity [20:43] Jason Makarow: human in the loop

Original transcript

[0:04] So, good afternoon everyone. My name is Jason MRO and um [clears throat] excuse me, [0:09] I'm the chief of staff to the president of Sacred Heart University and now the new chief [0:15] AI integration officer at Sacred Heart University. Um, we're a Catholic university in Connecticut, [0:22] uh, with about 10,000 students and about anywhere now with, uh, a few less international students, [0:30] anywhere between half a billion and about $600 million in revenue per year. [0:36] Um, so I'm presenting to you some research that the president had asked me to do when we decided [0:44] to put in some governance, governance before automation uh, at Sacred Heart. And I think [0:51] that anyone here representing an organization in this room, you're already using AI in some form. [0:58] So the question is whether you know about it or not. at Sacred Heart University. Uh the reason [1:06] that I was appointed as the chief AI officer was to build a governance framework uh before [1:11] deploying any enterprise AI, all the things that you're going to come in contact with while you're [1:16] here this week. And we did all of this, we built this governance in 12 weeks and on a budget of [1:24] $114,000 using existing infrastructure. So if you're running a business and somebody [1:31] says this is going to cost a lot, um $114,000 on half a billion dollars in revenue is less than [1:40] um 2/10en of a percent. So the academic paper that I put together documents the framework [1:47] uh that we're using at Sacred Heart University, and this presentation is more practical to show [1:52] you how to adapt it for your own organizations regardless of what industry you might be in. [1:59] So let me just tell you about what we found at just at one university. And I want you to think [2:04] about your own organization as I go through this. Before we built any governance at all, [2:09] we asked our VP of IT a simple question. How much are we spending currently on software? [2:16] And the answer he came back with was $22 million across 396 applications. And each and every one of [2:26] those tools went through the normal procurement process. They were reviewed for functionality, [2:31] cost, and compatibility. But the real question was the questions that no one was asking. Which one of [2:39] these tools are making external API calls? Where's the data payload being routed? Is there a third [2:48] party? Is there an endpoint that we don't have visibility into? Does the terms of service give [2:55] the vendor a right to use our data and generate a derivation from it? And when a tool that has [3:04] been brought in in any department says AI powered on the marketing page, is it really calling open [3:11] AI or anthropic under the hood without disclosing it in the contract? Those were the questions that [3:19] did not need to be asked five plus years ago. they do now. And most procurement processes, [3:27] especially in academic institutions, nonprofit institutions, mission-based institutions, [3:34] and at some corporations have not yet caught up. So, what we found was inside of our AI contract [3:41] management software, camera systems, donor modeling platforms, student tracking tools, [3:49] all of these things. We found an AI component in each one of them. And not because people were not [3:55] doing their jobs, but because the tools changed faster than the process designed to evaluate them [4:02] could catch up. And there was a second problem. Even though the tools we knew about had an AI [4:10] capability to in them, people weren't using them to their full capacity. We were paying for AI [4:17] features inside of platforms. are people used every day. But no one had yet been trained on [4:24] any of those features. No one was directly looking at them and no one was asking whether or not we [4:30] could solve operational problems with them. So the governance gap and the capacity gap really became [4:38] two two sides of the same coin. So you can't govern what you don't know about and you can't [4:44] use what no one else has yet taught you. So now I ask you to think about your organization. We're [4:51] an organization of over 10,000 students, 1300 employees, and we found 22 million in software [4:59] with ungoverned and underused AI. Think about your own areas of expertise in your own businesses. How [5:08] many tools do you have running in your departments right now that your CIO didn't approve and that [5:14] the legal team may not have reviewed and some people are using but not to their full capability? [5:21] I bring that up because as you can see, McKenzie in their 2025 state of AI survey found that 88% [5:29] of organizations now use AI in at least one business function. And MIT's Project Nanda [5:36] reported um that 95% of enterprise AI pilots, everything you're going to encounter this week, [5:44] will fail without governance. And those numbers are not just about universities. They're about the [5:50] businesses that you're here to represent. So this is why the academic research on change management, [5:57] this paper, matters to every organization in this room. The main authors, Carter, Rogers, and Hyitz, [6:05] they didn't build these things in a vacuum just for academics to sit around and talk about. They [6:11] built them for corporate environments. And so the basis of these of this paper was built on their [6:18] work. So what you're about to see in the rest of these slides is the framework that applies real AI [6:24] governance implementation at a university. And the reasons that they should matter to you is that the [6:31] frameworks work. This framework that I'll present to you works in a resource constrained institution [6:39] with only $114,000 budget, no data science team, and no outside consultants. It'll work in your [6:48] organization. I promise. The question, the real research question here is whether is not whether [6:55] or not you're using them, but whether or not you're deploying AI without any governance and [7:00] just hoping for the best. So shadow AI is not just a future risk. It's really a present condition. [7:09] What we built at Sacred Heart, what we built in 12 weeks on a $114,000 budget using existing [7:18] Microsoft 365 in infrastructure. The university built a framework with six components and we [7:24] designed it. The architecture aligns with NIST AI risk management with the NIST AI risk management [7:30] framework across all four functions. So taking all of the NIST functions and applying them with a lot [7:40] of the literature that's in the paper, we govern through committee charter and guiding principles. [7:46] We map through a three- tiered risk classification and we shadow AI inventory. We measure through [7:55] the registry's data capture and ROI tracking methodology and we manage through vendor review [8:01] and automate intake and workflow and an ongoing compliance. The theoretical foundations drawn on [8:09] organizational leadership and change in this paper are Cotter's eight-step change model structured [8:16] on implementation sequences, Rogers diffusion of innovation. It explains the committee's champion [8:23] design. Hyitz his adaptive leadership framework distinguishes the technical architects from the [8:30] adaptive challenge of changing how people evaluate AI tools. Bowman and Deal. Their foreframe model [8:38] ensures the framework operates structurally, politically, one of the most important pieces, [8:45] symbolically, and through the human resource investment all simultaneously. And finally, [8:52] Shine's organizational culture works the grounds of enablement over restriction. There's a joke [8:58] that goes around the university of a picture of me. Um and in the background it says I start [9:04] negotiating it. No, that's not the case. We want enablement over restriction. So this particular [9:13] paper that I'm presenting to you, this is a practice-based case study, not an experiment. [9:19] The paper doesn't claim the framework has reduced risk or has incr has improved outcomes. [9:26] It documents an operating model that connects every design decision to literature and it the [9:33] paper itself offers seven testable propositions for future empirical research. This particular [9:41] setup requires no new hires, no new platforms and no incremental costs to AI to it. [9:50] So these six components are described in detail in the paper's framework and design. Each one mapped [9:59] to both a governance problem and an organizational behavioral function. Table one in the paper in the [10:06] actual paper shows how each component addresses a specific governance need and simultaneously [10:12] performs a cultural and behavioral function within the organization. Rather than walk you [10:19] through each one of these six abstractly and going deep into the academia, I'm just going to show you [10:26] how they work together in short scenarios so that you can recognize them in your own organization. [10:33] And if any of these components interest you in detail, I'm certainly happy to walk you through [10:38] part of the treatment in the paper. Two versions of the same problem. First, your organization runs [10:46] workday or day force in HR. The latest update quietly added the option to screen and rank [10:52] applicants using AI. Nobody in HR purchased an AI tool. Nobody submitted a request. It [11:00] simply showed up on the platform update and an HR coordinator turned it on because it looked useful. [11:07] And it has been scoring candidates now for months. Legal doesn't know it's active. It didn't review [11:14] the model and nobody asked what data the ranking algorithm was trained on. So whether has it been [11:21] tested for disparit impact or whether turning that on turning that feature on allowed your data to be [11:29] processed by the vendor. This is not shadow AI in the traditional sense. Nobody in HR went rogue. [11:38] This is a feature that appeared inside a tool you already owned. Someone used it and your governance [11:45] process had no mechanism to catch it. Second, and this hits me right at home because this is what we [11:53] train our students to do. This is what would keep you uh CISOs up at night. A new hireer fresh out [12:01] of college or some techsavvy intern that was hired to work in your finance or your operations team [12:09] might have a Python environment on their laptop and they might have their own personal API key. [12:16] They write a script because that's what finance or operations tells the smart kid out of college [12:24] to do. You figure it out. This looks techy. They write a script that pulls records from your ERP, [12:31] passes them through the GPT um for the summarization and the classification and [12:38] you get back a spreadsheet, but the data now has left your network and no one knows about it. And [12:46] depending on the API this particular intern or new student or uh new graduate that you hired [12:55] purchased your data might be retrain might be retained for training and they don't even know it. [13:01] Your SIM won't flag it. Your DLP won't catch it unless you're inspecting some traffic at the end [13:06] point and most organizations aren't. This young person is not being malicious. They're solving a [13:14] problem finance or operations told them and no one gave them a better tool for it. And they learned [13:20] to do this on their own or in college. And by the time you find out, the data was already processed [13:28] and it was never evaluated. And who catches it? Most organizations. Nobody. This happens to us. [13:36] This was happening to us at the university all the time. We gave them a Python environment and [13:42] they were doing things putting out data that they shouldn't have. So it's no person or no [13:49] department. There was just no oversight. This particular framework now catches it. The committee [13:56] uh champion in that department now flags it and AI is now the point of contact for it. [14:03] Someone hears about a tool or a script that's running and the tool doesn't get killed, [14:08] it gets governed. So that's the enablement over restriction. We're not here to say no. We're [14:15] here to make sure that you know what you're using. Another one, a dean wants to deploy an AI tutoring [14:24] program for students. A compliance officer blocks it because it processes student records. [14:30] The dean says it's an academic decision. The compliance officer says it's a data decision. [14:35] Both are right. Who decides? Neither overrides the other. So it usually is whoever has some leader's [14:46] ear at the end of the day wins. But when you have an AI officer, the domain doesn't sign the [14:54] domain leader here in this case the dean does not sign vendor contracts. and the compliance [15:00] officer does not make curriculum decisions. The chief AI officer is the liaison in between. At [15:07] Sacred Heart University, I'm AI integration and we have someone else for faculty because god [15:13] forbid someone from staff does something that the faculty doesn't like. I love the faculty, [15:19] but god forbid a staff person or an administrator tells a faculty member how to govern. [15:27] This is not a university specific solution. This particular setup could be used in health care. [15:34] This is the same problem you see all the time. It's clinical versus administrative. This twotrack [15:40] model helps in so many ways. In nonprofits, it's programmatic versus operational. In government, [15:47] it's policy versus field operations. And so for us, it was definitely needed. Something I think [15:56] you'll run into a lot here. A vendor demos an AI platform for your leadership. The slides that they [16:02] show you are SOCK 2 certified. For us, they're furpa compliant. They're HIPPA ready. There's [16:08] no training. A president or a CEO is a totally impressed. And they want to sign. But before [16:14] you do, there's five questions. And there's five questions that we offer throughout this [16:21] entire framework that they originated as FURPA requirements because we're at a university, [16:27] but the logic applies to any regulated data environment, HIPPA, PCIDSS, state privacy laws, [16:34] GDPR, or any internal confidentiality policies. First, the data handling agreement. Is the [16:41] vendor contractually obligated to treat your data according to a specific standard? At a university, [16:48] this means the vendor agrees to function as a school officer under furpa. In healthcare, [16:53] it means their business associate agreement under HIPPA. You have to have it in the contract. Data [16:59] handling. Second, data residency. This is becoming more and more. Each one of us knows you may be [17:07] dealing with one of these companies here in the United States. They may have a place here in the [17:11] United States, but the data is held elsewhere. So obligations may change on data if it's held in a [17:17] different country. The no training clause, this is also pretty big. Nobody wants their data to train, [17:25] but if it's not in writing, if it's just on the FAQ, you're going to get in a lot of trouble. The [17:32] last two are key. The last two are things that I think most most people fail on. It's breach [17:40] notification. Excuse me. What's the vendor's commitment if your data is compromised? The [17:46] phrase as soon as practically as soon as practical or as soon as possible is not it's meaningless [17:52] in a contract and it's not helpful to you. You need a defined timeline. The standard at Sacred [17:58] Heart University for any AI tool is 48 hours. In some organizations, they require 24. The [18:05] point is you have to be specific and you have to ask. If a vendor can't commit to a number, [18:11] ask why. And the last one and the most important is encryption. What standard are they using at [18:17] rest than in transit? You know, are they using AES 256 at rest and TLS 1.2 or higher in transit? Both [18:26] questions. These are questions. They just can't claim them. Can they document the architecture? [18:32] Have they completed a SOCK 2 type two audit? What test controls over time do they have? [18:37] and have they had a third-party penetration test in the last 12 months? These are real questions [18:44] that we have to ask as we go through this with our vendors. Again, we're not here to say no. [18:51] We're just here to govern it. And a lot of vendors pass three out of the five and they usually fail [18:57] on those last two. Two quick examples, one nonprofit and one for-profit. Same problem. [19:04] I come from the Catholic university world. So we talk about the Catholic hospital system and an AI [19:10] tool to predict patient re uh readmission risk. But when they trained the model, the training [19:16] data was on underrepresented and low-income patients. The tool works well for insured [19:22] patients and badly for Medicare patients. So the tool passes every technical performance measure, [19:29] but the hospital's mission or in our case we have a nursing school. The mission is we serve all [19:35] patients with dignity. Well, when our AI serves uninsured patients better than Medicare patients, [19:40] the technology contradicts the mission. And a financial services firm, it deploys an AI model [19:47] to score loan applications. This is something a lot of universities are starting to do now, [19:53] especially when it comes to merit. It doesn't necessarily have to be just loan applications [19:57] for a house or a car or for something else, but the model approves suburban applicants at a higher [20:04] rate than urban applicants. So, it's technically accurate, but when the mission statement says [20:10] equal access to financial opportunities, it doesn't happen unless there's someone at the [20:16] wheel to be able to make sure that there's no mission failures. Mission is key. Whatever your [20:23] mission is, AI doesn't understand it. Someone needs to interpret it. So, you don't have to [20:29] be a Catholic organization or a mission-driven nonprofit to ask these questions. If your mission [20:34] statement makes a promise to the people you serve, AI governance has to enforce the promise. [20:43] Everything I've described to you so far is how institutions govern AI. But there's a second [20:48] output that matters to everyone in this room. You're now hiring people, students out of college, [20:56] who use AI the way you learned how to use Google. Nobody had a governance framework for search. You [21:05] just started using it and you went on your way intuitively. Every graduate from a university [21:11] can write a prompt and summarize the document. What you actually need is someone who knows how [21:17] to govern AI. people who question the outputs before acting on them and people who understand [21:24] that there is a human in the loop. It's not a regulatory requirement, but it's the reason [21:29] organizations make good decisions instead of fast ones. And so that's what we produce [21:34] at Sacred Heart University. Our students come in contact with AI governance from the minute [21:40] they apply to the minute they graduate. And so if there's anything that you might take from [21:46] this presentation, I want you to know you don't need a large budget, a huge data science team, [21:52] or a new platform for AI governance. You just need what's already in your environment, a registry and [22:00] an intake process, and people in the structures, not just policies. And so I invite you for anyone [22:08] who would like to read it, I'd welcome you to read this paper. It's very practical. [22:13] It's based in organizational leadership and change. And the paper really does ask how a [22:19] mission-driven organization without a data science team designed AI governance that actually works [22:26] across multiple colleges, multiple departments, and for multiple students. So, my email is here on [22:33] the screen, and if there's anything I could do to help you in that, I'm happy to do so. Thank you.