Ogletree Deakins Caps AI Deals at 12 Months to Avoid Vendor Lock-In

1,100 Attorneys, One Million+ Prompts: Building a 50+ Tool AI Stack at a Global Law Firm — Ai4

Ai4

Rigorous governance frameworks accelerate rather than delay enterprise AI deployment, while limiting commercial vendor contracts to 12 months prevents law firms from locking into obsolete software.

AI-assisted summary of 1,100 Attorneys, One Million+ Prompts: Building a 50+ Tool AI Stack at a Global Law Firm — original source. Check the original for full context.

Source published: 2026-08-16

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 16, 2026, this presentation features Tim Fox, Chief Data and AI Officer at Ogletree Deakins. The session details how the global firm grew from zero AI applications to a portfolio of over 50 licensed and bespoke tools, logged over one million prompts, earned an ISO 42001 certification, and successfully navigated buy-versus-build tradeoffs across 1,100 attorneys.

## TL;DR Tim Fox details Ogletree Deakins' deliberate 'fast follower' strategy, moving from 2023 legal skepticism and hallucination concerns to deploying over 50 AI tools with 80% attorney adoption and 250,000 monthly prompts. He highlights critical mistakes—such as understaffing change management and engaging in superficial 'governance theater'—and counters them with proven strategies like peer-led demonstrations and converting shadow AI into enterprise platforms. The firm maintains a strategic balance by buying frontier foundation models and specialized research platforms on strict 12-month contracts while reserving in-house engineering for unique, low-market niches like immigration.

## Atmosphere & style The presentation is pragmatic, transparent, and authoritative, striking a sharp contrast to standard corporate AI hype through wry humor and hard-won operational realities. Fox speaks with the grounded credibility of both a veteran litigator and a tech founder, openly sharing internal missteps alongside success metrics. It feels like an authentic insider briefing on enterprise change management, demystifying the intersection of cutting-edge technology, strict legal ethics, and human behavior.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=pu055hxdahs&t=4s) Ogletree Deakins adopted a deliberate 'fast follower' philosophy, allowing emerging AI tools to stabilize before implementing them to protect client confidentiality and comply with strict legal regulations. - [03:08](https://www.youtube.com/watch?v=pu055hxdahs&t=188s) The firm scaled its generative AI stack to over 50 tools, processing over one million cumulative prompts in under 18 months with 80% attorney adoption across 60 global offices. - [03:08](https://www.youtube.com/watch?v=pu055hxdahs&t=188s) Achieving ISO 42001 certification provided the foundational governance required to earn the trust of clients, attorneys, and leadership while rapidly expanding generative AI capabilities. - [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) The 'build it and they will come' mentality failed; meaningful adoption only accelerated after the firm staffed dedicated enablement personnel to guide practicing attorneys. - [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) Law firms must tailor AI tooling and enablement separately for practicing attorneys and administrative business professionals, as non-lawyers often avoid tools branded exclusively for legal work. - [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) Over-indexing on policy drafting and mandatory seminars leads to 'governance theater,' which creates delays without eliminating inherent AI risks; leadership must accept controlled risk and trust professional judgment. - [07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) Peer-led demonstrations by practicing partners drove significant, sustained adoption spikes compared to top-down instruction from central IT or innovation teams. - [07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) Systematic data collection and prompt classification enabled the AI team to uncover user intent, report actionable metrics to leadership, and secure multi-million-dollar licensing budgets. - [09:51](https://www.youtube.com/watch?v=pu055hxdahs&t=591s) Rather than banning unapproved 'shadow AI,' the firm identified unmet employee needs and negotiated enterprise licenses, creating built-in pilot groups with pre-validated use cases. - [10:33](https://www.youtube.com/watch?v=pu055hxdahs&t=633s) Governance functions as an operational accelerator rather than a roadblock, making organizations significantly more likely to scale extensive generative AI deployments successfully. - [11:52](https://www.youtube.com/watch?v=pu055hxdahs&t=712s) The firm buys frontier LLM capabilities, statutory legal research tools, and commodity utilities, reserving internal software development strictly for high-volume niche workflows like immigration that vendors do not support. - [14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) Weekend 'vibe coding' prototypes do not equal enterprise readiness, as translating prototypes into secure, production-grade tools still requires 6 to 9 months of engineering and governance reviews. - [16:06](https://www.youtube.com/watch?v=pu055hxdahs&t=966s) To prevent vendor lock-in and avoid the massive switching costs of migrating millions of prompts, the firm enforces a strict policy of capping AI vendor contracts at 12 months. - [17:41](https://www.youtube.com/watch?v=pu055hxdahs&t=1061s) Successful enterprise AI scaling requires viewing AI as a change management challenge rather than a technical deployment, focusing equally on human input, technology, and human review.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=pu055hxdahs&t=4s) Background and the Fast Follower Philosophy in Legal AI

Tim Fox introduces his background as the Chief Data and AI Officer at Ogletree Deakins, a global labor, employment, and immigration law firm comprising approximately 1,100 attorneys across 60 offices worldwide. A licensed attorney himself, Fox has worked at the intersection of law and artificial intelligence for over a decade, having co-founded the litigation-focused AI startup Legalmation in 2016. In addition to his leadership role at the firm, he teaches an AI fundamentals course at USC’s Gould School of Law and advises courts and judges as a member of the California Judicial Council’s Information Technology Advisory Committee. Opening with a touch of humor, Fox notes that he ran his slide deck and speaking notes through Claude and ChatGPT, which assured him this would be one of the top presentations at the conference—quipping that if it falls short, the audience can "blame AI for that."

Transitioning into the legal industry's structural constraints, Fox emphasizes that law firms face heavy regulatory and ethical burdens that make adopting technology on the bleeding edge nearly impossible. Because client confidentiality, ethical rules, and professional standards demand total reliability, legal organizations must wait for emerging technologies to stabilize before implementing them. Ogletree Deakins resolves this tension by adopting a deliberate "fast follower approach," purposefully choosing not to be the first to experiment with untested code, but moving decisively to adopt platforms the moment they are proven safe, robust, and compliant.

```mermaid flowchart LR accTitle: Fast Follower Adoption Framework in Legal Tech subgraph Market["Market Evolution"] A["Emerging Tech / Bleeding Edge"] --> B["Enterprise Stabilization"] end subgraph Firm["Ogletree Deakins Strategy"] B --> C["Safety & Compliance Vetting"] C --> D["Fast Follower Deployment"] D --> E["Firm-Wide Scale & Custom Workflows"] end ```

---

[02:07](https://www.youtube.com/watch?v=pu055hxdahs&t=127s) The 2023 Landscape: Skepticism, Sanctions, and Strategic Commitment

Looking back to early 2023, Fox characterizes the initial wave of generative AI in legal as fraught and chaotic. Hallucination incidents regularly made headlines, judges issued sanctions against lawyers who submitted briefs with fabricated case citations generated by tools like ChatGPT, and the models of that era offered questionable substantive value for rigorous legal drafting. Compounding the issue was deep professional anxiety: weekly studies oscillated wildly between claiming that AI would replace attorneys entirely and predicting a hiring boom to handle AI-generated complexity. Having heard warnings for fifteen years that technology was coming for their jobs, attorneys were understandably cynical. Fox acknowledges that in 2023, that skepticism was completely justified because the underlying technology was simply not ready for enterprise legal practice.

Despite the prevailing headwinds, Ogletree Deakins leadership identified generative AI as a transformative long-term priority rather than a passing fad. Rather than banning the tools or waiting passively on the sidelines, the firm established its foundational generative AI strategy in July 2023, laying the operational and governance groundwork for enterprise-level experimentation.

- **Industry Friction in 2023**: Frequent hallucination controversies and judicial sanctions for citing fake precedent. - **Lawyer Cynicism**: Decades of automated replacement narratives fostered natural resistance across the legal field. - **Strategic Mobilization**: Ogletree Deakins drafted and approved its initial generative AI strategy in July 2023.

---

[03:08](https://www.youtube.com/watch?v=pu055hxdahs&t=188s) From ODGPT to 1 Million Prompts: Scaling Tools and Enterprise Governance

The firm turned strategy into execution throughout 2024 by piloting major platforms across the industry, entering commercial licensing agreements with tools like Harvey and Lexis Protege, and engineering its own proprietary, firm-wide chatbot called ODGPT in October 2024. Fox marks the launch of ODGPT as the true turning point for firm-wide adoption. In its first month alone, ODGPT processed 9,000 prompts, initiating an exponential trajectory across the firm’s global footprint.

Within less than two years, the firm’s AI footprint expanded from a single internal prototype to an ecosystem of more than 50 deployed tools, split roughly evenly between commercial SaaS and homegrown applications. In May, the firm crossed the milestone of one million cumulative prompts in under 18 months from launch. Usage currently averages nearly 250,000 prompts per month, with projections on track to surpass two million prompts before the end of the calendar year. Beyond volume, adoption spans 80% of the firm's attorney population and accounts for more than 2,000 monthly active users (MAUs). Underpinning this infrastructure, Ogletree Deakins earned its ISO 42001 certification in June, becoming one of the earliest enterprises—and one of the first law firms globally—to certify its artificial intelligence management systems. This operational maturity has enabled the firm to co-develop bespoke, automated workflows directly with clients to lower costs and accelerate delivery while maintaining high-quality work product.

---

[05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) Pain Points and Failures: What Didn't Work

To provide a realistic picture of organizational transformation, Fox outlines the operational mistakes and dead ends Ogletree Deakins encountered during rollout:

1. **The "Build It and They Will Come" Fallacy**: The firm initially assumed that making tools accessible on the intranet would naturally drive engagement. In practice, only the top 10% to 20% of curious early adopters explored the tools independently. The vast majority of the firm, particularly practicing attorneys with heavy caseloads, required hands-on enablement and guided demonstrations before altering their established workflows. 2. **Understaffing Adoption Initiatives**: Early implementation was managed exclusively by Fox and a single additional attorney, who juggled vendor vetting, pilot management, technical rollouts, policy writing, and firm-wide training. This lean setup proved unsustainable. Adoption directly accelerated only when dedicated personnel were brought in to manage enablement full-time, a correlation clearly visible on the firm's monthly utilization curves: "It is a hire people and you will drive adoption. I can literally show you on our usage curves when somebody was hired at the firm to help lead our adoption initiatives." 3. **Treating the Firm as a Single Audience**: Law firms function as two parallel organizations: practicing attorneys and administrative business professionals (such as marketing, HR, finance, and knowledge management). Early efforts prioritized revenue-generating attorneys, assuming business staff would naturally follow along. However, business professionals refused to touch legal-branded platforms like Harvey because they perceived them as strictly for lawyers. Successful adoption required dedicated tooling and tailored training designed around non-legal operations. 4. **Governance Analysis Paralysis & "Governance Theater"**: Fearing headlines, reputational risk, and regulatory fallout before launching ODGPT, the team spent excessive time drafting exhaustive policy documents and running mandatory compliance seminars. Fox candidly labels this "governance theater"—activities that comforted leadership internally but did nothing to advance productive adoption while causing major release delays. Generative AI carries inherent, irreducible risks that no document can fully eliminate; true governance requires providing strong tools, clear parameters, and ultimately trusting qualified professionals to exercise sound judgment. Moving forward decisively while accepting controlled risk provides a distinct competitive advantage.

---

[07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) What Worked: Peer-Led Training, Metric Tracking, and Change Management

Turning to the core strategies that drove successful implementation, Fox identifies three foundational pillars:

First, peer-led education proved vastly more persuasive than top-down IT instruction. During a regional shareholder meeting in California, a practicing shareholder demonstrated live workflows showing precisely how he used AI to draft briefs, research arguments, and manage case files. The session resulted in a standing-room-only room of 100 partners and drove an immediate, permanent spike in regular platform usage across that cohort. Academic and industry studies consistently reinforce that professionals adopt complex technology fastest when observing peers and direct managers doing so, rather than listening to central innovation teams: "People are tired of me talking about AI at my firm. They want to hear from their peers."

```mermaid flowchart TD accTitle: Pillars of Effective Legal AI Adoption subgraph Execution["Core Adoption Drivers"] P1["Peer-Led Enablement\n(Partner-to-partner demos)"] P2["Telemetry & Prompt Classification\n(Demonstrate ROI & secure funding)"] P3["Change Management Mindset\n(Focus on human inputs & outputs)"] end Execution --> Outcome["Sustainable Firm-Wide AI Utilization"] ```

Second, rigorous data collection and prompt classification allowed the AI team to demonstrate concrete value. Every prompt is classified to uncover true user intent, and monthly summaries are distributed directly to executive leadership. This reporting creates a positive feedback loop: broader awareness of rising firm usage stokes organic peer interest, while transparent utilization metrics justify multi-million-dollar software licensing budgets to executive management.

Finally, Fox stresses that rolling out AI is fundamentally an exercise in change management rather than a technology deployment. Technology represents only one-third of the operational loop—sandwiched between human input on the front end and human review on the back end. Innovation teams that concentrate solely on the technology layer while neglecting the human components inevitably struggle with low engagement and missed business outcomes.

[09:51](https://www.youtube.com/watch?v=pu055hxdahs&t=591s) Turning Shadow AI into Enterprise-Licensed Solutions

Focusing on the broader cultural and operational aspects rather than just technical deployment is vital for driving AI adoption. At Ogletree Deakins, a particularly successful strategy involved closely monitoring "shadow AI"—the unapproved tools employees were already bringing in to do their jobs. Rather than penalizing users or outright banning these unauthorized platforms, the firm actively examined what unmet needs those tools fulfilled.

Once a clear pattern emerged, the AI team sought enterprise licensing agreements to fold those specific tools directly into the firm’s formal governance framework. This approach turned existing user behavior into a massive operational advantage: the firm gained an immediate, built-in pilot group of enthusiastic users with pre-validated legal use cases. Crucially, it prevented the widespread frustration and complaints that typically follow when IT abruptly restricts access to popular software, transforming an administrative friction point into an enterprise-enabled workflow.

- Rather than cracking down on shadow AI, evaluate organic usage to identify real operational needs. - Enterprise licensing brings unvetted tools under corporate compliance while retaining existing user momentum. - Utilizing built-in pilot groups dramatically cuts discovery and testing time for the AI team.

---

[10:33](https://www.youtube.com/watch?v=pu055hxdahs&t=633s) Why Rigorous Governance Is the True Accelerator of AI

While governance is traditionally viewed as the bureaucratic black hole "where you send projects to die," it acts as an indispensable accelerator in the AI era. Citing recent industry research, Tim Fox points out that organizations with mature governance frameworks are 75% more likely to drive successful AI adoption and nearly four times as likely to scale extensive generative AI deployments.

```mermaid flowchart TD accTitle: Governance as an AI Accelerator subgraph GOV["Governance Foundation"] A["Formal Committees & Policies"] --> B["ISO 42001 Certification"] end subgraph TRUST["Stakeholder Alignment"] B --> C["Win Reluctant Clients"] B --> D["Reassure Attorneys"] B --> E["Obtain General Counsel Approval"] end subgraph SCALE["Enterprise Velocity"] C & D & E --> F["Rapid AI Adoption & Scale"] end ```

Ogletree Deakins achieved its ISO 42001 certification and established dedicated oversight committees precisely because attorneys, clients, and general counsels refuse to engage with generative AI if they do not trust its safety and privacy controls. Solid governance is what ultimately gives leadership the comfort level required to approve deployments quickly, proving that "you cannot move fast without good governance" while ensuring everyone can sleep at night knowing systems operate responsibly.

- Good governance yields a 75% higher likelihood of driving day-to-day AI adoption. - Organizations with formal oversight are nearly four times more likely to run large-scale AI operations. - Earning the ISO 42001 certification validates compliance and builds trust across attorneys, clients, and risk officers.

---

[11:52](https://www.youtube.com/watch?v=pu055hxdahs&t=712s) Strategic Buy vs. Build Framework: Frontier Models to Niche Tools

The firm’s 50+ tool stack is split evenly between purchased software and internally engineered tools. Deciding whether to buy or build rests on a clear division of labor: Ogletree Deakins buys frontier capabilities, specialized legal knowledge bases, and hyper-efficient niche utilities, while strictly reserving internal development for workflows that third-party vendors cannot serve.

| Acquisition Category | Target Tools / Providers | Strategic Rationale | |:--- |:--- |:--- | | **Frontier Capabilities** | Claude (Anthropic), Microsoft Copilot, ChatGPT (OpenAI) | Unreplicable core LLM foundations; firms cannot compete with trillion-dollar AI labs. | | **Domain-Specific Legal Research** | Harvey, Lexis+ AI / Lexis Protégé | Deep integration with proprietary statutory and case-law datasets unavailable internally. | | **Cost-Effective Niche Utilities** | Third-party deposition summarizers | Commodity point solutions available for roughly $20 per run that make custom development uneconomical. | | **Bespoke In-House Builds** | High-volume arbitration workflows, operations, immigration | Highly customized proprietary processes and data that vendors cannot or will not build for small addressable markets. |

Bespoke in-house builds are deployed heavily across unique practice groups like immigration. With only five to ten peer firms operating immigration practices of comparable scale, third-party software vendors lack the financial incentive to create specialized platforms for such a narrow market. Building proprietary tools for these specialized groups has automated massive portions of low-stakes arbitration and administrative work, saving the firm thousands of hours.

---

[14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) The Hard Realities of In-House Engineering and "Vibe Coding"

Engineering AI solutions in-house requires extreme caution because "just because you can build it does not mean you should build it." Creating a piece of software is never a one-off effort; long-term ownership requires an ongoing infrastructure to handle support, maintenance, model updates, changing data schemas, and documentation when key engineers leave.

The advent of modern generative AI creates a deceptive dynamic around prototyping. An attorney can now "vibe code" an impressive proof of concept over a single weekend and enthusiastically bring it to the AI team on Monday morning expecting immediate rollout. However, bridging the gap between a weekend prototype and an enterprise-grade production tool requires rigorous code reviews, value assessments, security vetting, and deployment pipelines—a process that still takes 6 to 9 months of manual effort, often causing user frustration if expectations are not carefully managed.

- Every custom build requires an enduring support plan for weekend trial emergencies, schema shifts, and engineer turnover. - Rapid weekend prototyping does not equal production readiness; enterprise scale still requires months of manual stabilization. - Generative AI accelerates initial code authoring but dramatically increases the downstream architectural and governance review burden.

---

[16:06](https://www.youtube.com/watch?v=pu055hxdahs&t=966s) Managing High Switching Costs and Capping AI Contract Durations

When purchasing commercial AI platforms, organizations must factor switching costs directly into their initial evaluations. Once a firm logs over one million prompts, the accumulated institutional knowledge, prompt history, and user habits become virtually impossible to extract or migrate between competing platforms. Jumping ship risks significant data loss and wrecks hard-won change management gains, leaving many law firms trapped with lagging legacy vendors due to pure institutional inertia.

Simultaneously, the breakneck pace of model evolution means specialized point solutions are frequently subsumed by newer frontier models. Attorneys consistently voice a strong preference for staying inside a single unified frontier environment rather than toggling across five standalone point solutions for marginal quality gains. To avoid being locked into obsolete platforms, Ogletree Deakins maintains a strict policy: never commit to generative AI vendor contracts longer than 12 months.

- Migrating a million prompts of institutional knowledge across platforms is currently technically and culturally unfeasible. - Rapid advances in frontier models routinely render narrow single-purpose legal AI startups obsolete within months. - Limit all generative AI vendor agreements to a maximum 12-month contract duration to preserve operational flexibility.

---

[17:41](https://www.youtube.com/watch?v=pu055hxdahs&t=1061s) Key Principles for Scaling Enterprise AI

Closing out the presentation, Tim Fox highlights five hard-earned operational rules for driving enterprise-wide AI transformation:

- **Staff adoption as a dedicated function:** Tools do not distribute or evangelize themselves; dedicated team members are required to manage onboarding and ongoing support. - **Lead with peer-led conversations:** Enlist active internal power users and AI practitioners to train and guide their colleagues rather than relying purely on top-down IT messaging. - **Measure success and champion usage continuously:** Track metrics systematically and present utilization reports to firm leadership to maintain executive momentum. - **Buy where possible, build only when necessary:** Factor switching costs and workflow stickiness into every purchase while keeping contracts capped at 12 months. - **Build governance before scaling adoption:** Establish robust security, ISO standards, and review frameworks first, as trustworthy governance is the single most critical accelerator for AI speed and user confidence.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=pu055hxdahs&t=4s) blame AI for that — Tim Fox

> [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) It is a hire people and you will drive adoption. I can literally show you on our usage curves when somebody was hired at the firm to help lead our adoption initiatives. — Tim Fox

> [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) governance theater — Tim Fox

> [07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) People are tired of me talking about AI at my firm. They want to hear from their peers. — Tim Fox

> [10:33](https://www.youtube.com/watch?v=pu055hxdahs&t=633s) where you send projects to die — Tim Fox

> [10:33](https://www.youtube.com/watch?v=pu055hxdahs&t=633s) you cannot move fast without good governance — Tim Fox

> [14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) just because you can build it does not mean you should build it — Tim Fox

> [14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) vibe code — Tim Fox

## Mentioned Tim Fox, Ogletree Deakins, Legalmation, USC Gould School of Law, California Judicial Council Information Technology Advisory Committee, Claude, Anthropic, ChatGPT, OpenAI, Harvey, Lexis Protege, Lexis+ AI, Microsoft Copilot, ODGPT, ISO 42001

## Video description Tim Fox - Chief Data and AI Officer, Ogletree Deakins

In under 2 years, I helped take Ogletree Deakins from no AI tools to a stack of more than 50 licensed and homegrown tools, crossing well over one million prompts along the way and still climbing. We did it all while earning our ISO 42001 certification, ensuring enterprise governance for every decision. In this session, I'll cover the buy vs. build decisions behind that growth, including where Claude, Harvey, and Microsoft Copilot fit, and what worked and what didn't. I'll also share the use cases that actually drove adoption across the firm. You'll leave with lessons learned, practical use cases, and ideas on how to drive adoption at your own organization.

[Source](https://www.youtube.com/watch?v=pu055hxdahs)

中文简报

Ogletree Deakins将AI协议期限定为12个月以规避供应商锁定严谨的治理框架不仅不会拖延、反而会加速企业AI的部署，而将商业供应商合同限制在12个月内，可以防止律所被过时的软件所套牢。 Tim Fox详细介绍了Ogletree Deakins深思熟虑的“快速跟进者”策略：从2023年法律界对AI的质疑及对幻觉的担忧，转变为部署超过50种AI工具、实现80%的律师采用率以及每月25万次提示词使用量。他指出了若干关键失误——例如变革管理人员配备不足以及流于表面的“治理作秀”——并提出了行之有效的应对策略，如同行主导的演示以及将“影子AI”转化为企业级平台。该律所保持着战略平衡，在采购前沿基础模型和专业研究平台时严格执行12个月的合同期限，同时将内部工程研发保留给移民等独特的细分小众领域。全球劳工与移民法律事务所 Ogletree Deakins 采取“快速跟随者”策略，在不到 18 个月内将生成式 AI 堆栈扩展至 50 多款工具（采购与自研各半），累计提示词处理量突破 100 万次，全球律师月活采纳率达到 80%。 (3:08) “建好工具大家自然会用”的设想在实践中失败，推动全所普及的关键在于配备专职推广赋能人员、由一线合伙人进行同行实操演示，并将执业律师与行政管理人员的需求进行差异化工具定制。 (5:16) 通过 ISO 42001 人工智能管理体系认证为律所奠定了坚实的治理基础，成熟严密的安全与合规审查不仅没有阻碍创新，反而成为赢得客户与法务信任、快速规模化部署 AI 的关键加速器。 (10:33) 在买与建的权衡上，律所选择采购底层前沿模型（如 Claude、Copilot）与垂直法律工具（如 Harvey），自研则严格限定在供应商无法覆盖的小众自营场景（如移民与仲裁工作流），同时将所有商业合同上限设为 12 个月以防锁定与快速淘汰。 (11:52)

Reading summary

Tim Fox, Chief Data and AI Officer at Ogletree Deakins, argues that enterprise legal AI succeeds not through bleeding-edge experimentation, but through a disciplined 'fast follower' strategy supported by strict governance, peer-led training, and a pragmatic buy-versus-build framework. Scaling Enterprise Adoption Through Structured Governance (3:08) After establishing its initial generative AI strategy in July 2023, Ogletree Deakins launched an internal chatbot, ODGPT, in October 2024. The tool processed 9,000 prompts in its initial month and expanded across the firm's 60 offices, crossing one million cumulative prompts within 18 months. Monthly volume averages approximately 250,000 prompts, with attorney penetration reaching 80% and over 2,000 monthly active users across more than 50 deployed tools. Rather than viewing compliance as a barrier, Fox cites research indicating that mature governance frameworks make enterprises 75% more likely to drive adoption and nearly four times as likely to scale generative AI. In June, the firm obtained ISO 42001 certification for its artificial intelligence management systems. Fox argues that formal certification and oversight committees provide the assurance required by clients, attorneys, and general counsels to move quickly without succumbing to 'governance theater' or excessive policy drafting that delays deployment. Overcoming Adoption Roadblocks With Peer Enablement (5:16) Early implementation revealed several operational dead ends. Placing applications on the firm intranet attracted only the top 10% to 20% of curious users, while maintaining an initial implementation team of just two people choked firm-wide scaling. Furthermore, attempting to address attorneys and administrative staff as a single audience failed because business professionals avoided legal-branded software like Harvey. Significant adoption curves materialized only after the firm hired dedicated enablement staff and separated training for business operations. To drive sustained engagement, the firm shifted from central IT instruction to peer-led education, demonstrated by a California partner meeting where a practicing shareholder's live workflow demonstration generated immediate usage spikes across 100 partners. Additionally, rather than banning unauthorized 'shadow AI,' the AI team tracked organically adopted tools and negotiated enterprise licensing agreements. This converted shadow usage into compliant workflows with pre-existing, enthusiastic user cohorts. A Disciplined Framework for Buying Versus Building (11:52) The firm divides its 50+ tool ecosystem equally between third-party software and proprietary internal builds. Commercial licenses are reserved for frontier foundational models (such as Claude, ChatGPT, and Microsoft Copilot), established legal research platforms (Harvey and Lexis Protégé), and commodity point solutions like deposition summarizers priced at roughly $20 per run. In-house development is strictly restricted to workflows that commercial vendors neglect due to small addressable markets, such as bespoke arbitration and immigration processes. Fox warns against the deceptive simplicity of 'vibe coding,' where lawyers create weekend prototypes and expect immediate enterprise rollouts. Transitioning an unvetted prototype to production requires security vetting, code reviews, schema maintenance, and long-term support infrastructure that typically demands six to nine months of manual effort. He asserts that organizations should build internally only when commercial options are non-existent. Managing Switching Costs and Vendor Obsolescence (16:06) Accumulated usage data creates substantial operational friction when attempting to change software providers. Once an organization logs over one million prompts, user habits and institutional prompt histories become exceedingly difficult to export, threatening to erase hard-won change management gains and trapping firms in legacy contracts. At the same time, rapid frontier model advancements frequently absorb narrow, single-purpose legal AI utilities, while attorneys resist switching across multiple standalone tools for marginal gains. To mitigate vendor lock-in and retain the ability to pivot as frontier models advance, Ogletree Deakins enforces a mandatory policy limiting all generative AI vendor agreements to a maximum duration of 12 months. Ogletree Deakins' model illustrates that sustainable AI deployment depends on treating adoption as an ongoing change-management effort rather than a purely technical rollout. Success requires balancing controlled risk acceptance with firm-wide peer training, while enforcing rigid limits on vendor lock-in.

精炼阅读摘要

Ogletree Deakins首席数据与人工智能官Tim Fox阐明，受制于法律行业的强合规与职业伦理约束，律所应采取“快速跟随者”策略，通过扎实的治理架构、专职推广团队与同行现场实操示范，实现生成式AI的大规模平稳落地。从谨慎起步到百万级提示词规模化 (3:08) Fox指出，法律行业对客户保密和职业伦理的严苛要求，使律所难以冒险采用尚未成熟的前沿技术。Ogletree Deakins确立了“快速跟随者”路线，即不当首批试验未经测试代码的先行者，但在技术被证实安全合规后迅速采纳。律所于2023年7月正式确立生成式AI战略，随后于2024年10月推出自研聊天机器人ODGPT。该工具首月即处理9,000次提示词交互，推动应用规模呈指数级扩张。截至当年5月，该所在上线不到18个月内累计提示词处理量突破100万次，月均处理量接近25万次，按此增速年底前有望突破200万次。目前全所部署了50余款工具，商业采买与内部自研各占一半，覆盖全球60个办公室约80%的执业律师以及2,000多名月活跃用户。此外，律所于6月获得ISO 42001人工智能管理体系认证。Fox援引行业研究指出，具备成熟治理框架的机构推进AI日常应用的概率高出75%，实现大规模部署的几率接近四倍。破除“治理表演”与早期推广误区 (5:16) 回顾初期推广历程，Fox坦承律所曾遭遇四大执行失误。首先是“建好自然有人来”的被动设想：仅在内网上线工具只能调动10%至20%的早期极客，大部分承担繁重业务指标的律师若无针对性演示绝不会改变既有工作流。其次是专职推广人员不足：初期仅靠Fox与另一名律师兼职协调采购、测试和培训，直至正式引入专职团队负责推广赋能，月度使用曲线上才出现显著的增长拐点。再次是受众群体的单一化误区：初期资源过度集中于创收律师，导致行政、财务等运营人员对带有浓厚法律属性的工具产生疏离感，促使团队随后针对非业务条线定制辅助工具。最后，团队一度陷入过度追求合规的“治理表演”：因担忧声誉受损，在ODGPT上线前耗费大量时间起草冗长制度并组织强制性讲座，导致上线拖延。Fox强调，生成式AI存在不可消除的固有风险，治理的核心在于设定明确边界并提供坚实工具，而非靠合规文件消弭一切风险，律所最终必须信任专业人员的独立判断。以同行示范与收编“影子AI”驱动采纳 (7:50) 在推动全员使用的战术层面，Fox强调同行示范的效力远高于IT部门自上而下的行政宣导。例如在加州区域合伙人会议上，一名执业合伙人现场演示了如何运用AI起草辩护状、检索案例并整理卷宗，吸引了100名合伙人到场，并直接带动该群体日常使用量的永久性攀升。同时，团队建立了严格的提示词分类与遥测系统，通过洞察用户真实意图并每月向最高管理层汇报，以透明的数据指标证明实际投资回报，为数百万美元的软件采购预算提供支撑。另一项关键策略是将“影子AI”转化为企业正式资产。面对员工私自引入的未授权工具，团队并未采取惩罚或封禁措施，而是深入分析其背后的业务痛点。一旦发现普遍需求，便通过采购企业级许可将其纳入统一治理体系。这种做法保留了自带高黏性与已验证场景的种子用户群，大幅缩短了技术测试周期，避免了IT强制干预造成的抵触情绪。买建边界、原型陷阱与合同期限管控 (11:52) Ogletree Deakins在50余款工具中维持着清晰的“买建分工”：前沿基础大模型（如Claude、ChatGPT）、垂直法律知识库（如Harvey、Lexis Protege）及单次运行约20美元的证词摘要等低成本通用工具一律对外采购；自主研发则严格聚焦于外部厂商无法覆盖的业务场景。例如在其庞大的移民与仲裁业务中，由于具备同等规模的同行仅有5至10家，商业软件公司缺乏研发动力，律所通过自研定制工具自动化了大量低风险行政流程，累计节省数千小时。在自研过程中，Fox警示了“氛围编程”（vibe coding）带来的虚假预期：律师即便能在周末拼凑出功能惊艳的原型，但将其转化为企业级应用仍需经过代码审查、安全审计与持续运维，这一过程通常需要6至9个月的人工介入。最后，针对商业AI的高昂迁移成本，Fox指出跨平台迁移百万级提示词与组织习惯几乎不可行。为防止前沿大模型的快速演进令单点工具迅速过时，律所严格执行商业合同上限为12个月的纪律，以规避技术锁定风险。 Fox总结强调，法律机构的人工智能落地本质上是组织变革管理，技术本身仅占业务闭环的三分之一，其前后端高度依赖人类律师的精准输入与严格审核。尽管严密的制度化治理与为期一年的合同约束能有效规避锁定并构筑客户信任，但在前沿大模型持续吸收垂直工具功能、律师自研原型与工程化维护脱节的背景下，如何在保障敏捷度的同时维持自研体系的长期生命力，仍是规模化推广中需要持续平衡的现实挑战。

Detailed notes — zh

## Background 本视频由 Ai4 于 2026 年 8 月 16 日发布在 YouTube 平台。内容由全球知名律所 Ogletree Deakins 的首席数据与人工智能官 Tim Fox 主讲，根据简介，他分享了如何在不到两年内带领 1,100 名律师从零构建包含 50 余款工具的 AI 矩阵并突破百万次 Prompt 调用，同时阐述了自研与外采决策、落地用例及 ISO 42001 认证等企业级治理经验。 ## TL;DR 面对法律行业高监管与律师群体早期极度怀疑的挑战，Ogletree Deakins 确立了“快速跟随”策略，在不到两年内实现了全所 80% 律师常态化使用 AI 与超百万次 Prompt 交互。演讲坦诚复盘了早期陷入“建好平台自然有人来”、落地人手匮乏、忽视行政支持人员以及“治理剧场”导致的分析瘫痪等四大弯路，并提出了同行示范培训、量化数据看板和变革管理等成功实践。在技术选型上，律所坚守前沿模型与细分工具“能买则买、单次签约不超 12 个月”，仅在小众高容量业务中谨慎自研，同时将治理视为推动组织放胆创新的加速器。 ## Atmosphere & style 演讲整体呈现出极其务实、坦诚且极具行业洞见的清醒风格，Tim Fox 以资深律师兼技术创业者的双重身份，用幽默自嘲的开场迅速破冰。叙事没有停留在空泛的 AI 宏大叙事上，而是直面“周末随手写代码留下一堆运维烂摊子”和繁琐官僚流程等真实痛点，将血泪教训与翔实数据毫无保留地摊在台前。整场分享节奏紧凑有力、金句频出，既有高层治理的全局战略感，又散发着扎根业务一线的实战烟火气，为面临 AI 变革阵痛的组织管理者提供了极强的说服力与行动底气。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=pu055hxdahs&t=4s) 法律行业面临强监管与高合规要求，Ogletree Deakins 确立了“快速跟随者”（fast follower）策略：待技术被验证安全可靠后再迅速规模化落地。 - [02:07](https://www.youtube.com/watch?v=pu055hxdahs&t=127s) 律所从 2024 年底上线自研聊天机器人 ODGPT 到突破 100 万条 Prompt 耗时不到 18 个月，构建了 50 多款工具矩阵并斩获 ISO 42001 认证。 - [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) 警惕“系统上线自然有人用”的误区，仅 10%-20% 的早期尝鲜者会自发使用，必须配备专职 Adoption 运营人员手把手辅导才能带动用量上升。 - [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) 执业律师与行政运营人员需求迥异，不可视作单一受众，需针对商业支持团队定制专属工具以打破“仅律师可用”的心理壁垒。 - [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) 摆脱过度规避风险的“治理剧场”（Governance Theater），生成式 AI 不存在绝对零风险，组织须承担适度风险并建立对员工的信任以抢占先机。 - [07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) “同行驱动培训”（Peer-led Training）成效远超技术部门单向宣讲，业务骨干和合伙人的实操示范能极大激发全员使用意愿。 - [07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) 精细化度量 Prompt 与使用场景并通过公开看板汇报，既能营造从众效应推动普及，也为争取数百万美元预算提供了强力数据支撑。 - [09:51](https://www.youtube.com/watch?v=pu055hxdahs&t=591s) 正向收编“影子 AI”（Shadow AI），将员工私下使用的工具升级为企业级许可，不仅满足真实业务痛点，还直接获得现成的验证场景与先锋测试群体。 - [09:51](https://www.youtube.com/watch?v=pu055hxdahs&t=591s) 健全的治理是 AI 推广的“加速器”而非绊脚石，数据表明完善治理能使日常 AI 普及率提升 75%，并使大规模部署成功率提升近 4 倍。 - [11:52](https://www.youtube.com/watch?v=pu055hxdahs&t=712s) 践行“能买则买、不买才建”的原则：基础模型、法律专属数据库及廉价细分工具坚决外采，仅在市场空白且业务高度定制的场景（如移民业务）下自研。 - [14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) 慎防轻率自研带来的长期维护负债，AI 降低了原型开发门槛但未自动化代码审查、全员部署与合规运维等繁重工作。 - [14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) 采购工具严禁签署超过 12 个月的长期合同，以防基础模型快速迭代吞没单点工具，同时警惕高昂的转换成本与数据锁定风险。 - [17:41](https://www.youtube.com/watch?v=pu055hxdahs&t=1061s) 全所推进 AI 规模化的核心在于将落地设为专职岗位、依托同行互助、量化宣传价值、保持采购灵活，并在追求速度前筑牢治理底座。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=pu055hxdahs&t=4s) 法律行业的 AI 限制与“快速跟随者”策略 Tim Fox 担任全球知名律所 Ogletree Deakins 的首席数据与人工智能官（Chief Data and AI Officer）。作为一家深耕劳动、用工与移民法领域的全球律所，Ogletree Deakins 在全球设有 60 个办公室，拥有约 1,100 名执业律师。Tim Fox 本人既是一名执业律师，也是法律科技领域的资深创业者与学者——他早在 2016 年就联合创办了首家专注于诉讼领域的 AI 公司 Legalmation，在法律与 AI 交叉领域拥有超过 10 年的实战经验；此外，他还在南加州大学古尔德法学院（USC Gould School of Law）教授《AI 基础》课程，并任职于加州司法委员会（California Judicial Council）的信息技术顾问委员会，为法庭和法官提供 AI 相关的决策建议。在开场时，Tim 打趣道，自己在演讲前把 PPT 和讲稿扔进了 Claude 和 ChatGPT 进行评估，模型预测这将是整场大会最精彩的演讲之一，“所以如果讲得不好，别怪我，去怪 AI 吧”。随后，Tim Fox 深入剖析了法律行业在技术拥抱层面的特殊性。法律是一个受到高度监管的专业行业，面临着极其严格的行业规制与合规红线，这导致律所无法像其他科技企业那样一味追求最前沿的技术。“受制于这些限制，我们永远无法直接站在技术的最锋利尖端，必须等待技术逐渐成熟且被证明安全”，因此 Ogletree Deakins 确立了所谓的“快速跟随者”（fast follower）策略：一旦某项技术在律所场景中被验证为安全可靠，便以最快速度跟进并落地。 ```mermaid flowchart LR accTitle: Fast Follower Strategy Framework subgraph S1["行业特征与约束"] A["高监管与严格合规限制"] --> B["不可直接试水未经检验的尖端技术"] end subgraph S2["落地执行原则"] B --> C["等待技术成熟与安全性验证"] C --> D["快速跟随策略 (Fast Follower)"] D --> E["安全合规前提下实现规模化应用"] end ``` ### [02:07](https://www.youtube.com/watch?v=pu055hxdahs&t=127s) 从 2023 年的行业怀疑到全所爆发式落地回溯生成式 AI 在法律行业起步的 2023 年，当时的整体环境极为艰难。大模型幻觉（hallucination）的新闻层出不穷，经常有律师因 ChatGPT 编造虚假案例判例而面临法院制裁；当时的通用模型在法律实务中的直接应用价值饱受质疑，整个律师群体深陷“AI 究竟是实用工具还是职业威胁”的迷茫与分歧中。每周都有相互矛盾的研究报告出炉，一边声称 AI 将彻底取代律师，另一边又宣称 AI 将催生更多律师需求。叠加过去 15 年间技术替代论对律师群体的反复冲击，律师本能的怀疑主义在 2023 年达到了顶峰——Tim 坦言，站在当年的技术成熟度来看，律师们的怀疑是完全合理的。然而，Ogletree Deakins 的管理层选择在早期捕捉其战略潜力，并将 AI 列为全所核心战略： - 2023 年 7 月：制定首份全所生成式 AI 落地战略； - 2024 年上半年：全面启动各大主流平台试点，先后与 Harvey、LexisNexis 的 Protégé 等行业前沿平台签署许可协议； - 2024 年 10 月：正式上线面向全所的自研聊天机器人 **ODGPT**，首月即产生 9,000 条 Prompt，标志着律所规模化 AI 应用的正式起跑。在随后的发展中，律所的 AI 规模呈现出爆炸式增长。全所可用的 AI 工具矩阵从单一的 ODGPT 扩展至 50 余款（自研与采购比例约为 1:1）。从首条 Prompt 到突破第 100 万条 Prompt，整个过程耗时不到 18 个月；全所月均 Prompt 调用量已接近 25 万条且持续攀升，预计年底前累计 Prompt 数量将跨越 200 万大关。目前，约 80% 的执业律师每月固定使用 AI 工具，全所各平台合计月活跃用户（MAU）超过 2,000 人。2024 年 6 月，该所顺利通过 ISO 42001（人工智能管理体系）认证，成为全球首批、也是最早斩获该认证的律所之一。同时，他们与企业客户深度共创定制化 AI 工作流，在保障服务交付品质的前提下大幅提升效率并降低成本。 ### [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) AI 推广踩坑实录：四大惨痛教训在回顾光鲜战绩的同时，Tim Fox 坦率分享了推进过程中走过的弯路与踩过的四个核心“大坑”： 1. **“建好平台，他们自然会来”（Build it and they will come）的误区**：系统上线后并不会自然引发全员使用。“拥有工具绝不等于产生用量”，组织内部仅有 10% 到 20% 的早期尝鲜者（early adopters）会主动摸索，而绝大多数人——尤其是工作繁忙的律师——需要大量的辅导与手把手教学。 2. **推动团队人员配置严重不足**：项目初期仅由 Tim Fox 与另一名律师两人包揽技术评估、试点、部署、培训等全部工作，导致推进极其艰难。Tim 指出，这绝非“先有鸡还是先有蛋”的悖论，而是“只要配置专职落地人员，使用量曲线就会立竿见影地上扬”。 3. **将律所视作单一受众群体**：律所内部实际上由两类截然不同的群体构成——“执业律师”与“商业专业人员/行政运营团队”（business professionals）。初期团队将全部精力放在作为创收核心的律师身上，误以为只要律师用起来，行政与支持人员自会跟进。但实际情况是，行政支持人员认为 Harvey 这类平台属于“律师专用工具”，即便被告知有权限也不会主动尝试，必须针对非律师群体定制专属工具与场景。 4. **治理层面的“分析瘫痪”（Governance Analysis Paralysis）**：在上线 ODGPT 前夕，管理层陷入了极度焦虑，担心律师滥用工具导致负面舆情甚至引发责任风险。团队因此耗费大量时间起草各类繁琐政策、组织合规培训，Tim 将其形容为自欺欺人的“治理剧场”（governance theater）——这些举措除了推迟产品发布上线，并未解决实际问题。团队最终领悟到：生成式 AI 根本不存在绝对的“零风险”，组织必须承受适度风险并建立对员工的信任，在提供必要工具与安全培训的同时放手让其规范操作；而敢于比同行更早承担风险并付诸行动的机构，才能确立自身的竞争优势。 | 踩坑场景 | 错误假设与做法 | 实际后果与教训 | |:--- |:--- |:--- | | **产品推广** | 工具上线后员工会自发探索使用 | 仅 10%-20% 主动尝试，其余律师必须有人手把手辅导 | | **团队建设** | 依赖 2 名核心人员兼顾评估到培训全流程 | 推广停滞；必须配置专职 Adoption 运营人员带动用量跃升 | | **受众定位** | 将全所视作单一整体，重点突破创收的律师群体 | 商业与行政支持人员对“律师工具”产生壁垒心理，需针对性设计产品 | | **风险治理** | 试图通过制定繁琐制度和反复培训彻底规避所有风险（治理剧场） | 上线严重拖延；AI 必然伴随风险，需建立信任并在快速行动中构建优势 | ### [07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) 推动 AI 落地的成功经验：同行培训、数据度量与变革管理在总结成功经验时，Tim 强调了三项在实践中极其见效的核心策略：首先是**“同行驱动培训”（Peer-led Training）**。让业务一线员工向同事传授经验，其转化效果远胜于技术专员的单向说教。Tim 分享了一个生动案例：在半年前的一场加州合伙人（Shareholder）内部会议上，一位合伙人登台亲自演示了自己在案件实务中如何使用 AI 工具，现场 100 多人座无虚席、全神贯注。这场单一分享对加州律师 AI 使用意愿的拉动，超越了 Tim 过去举办的所有培训宣讲总和。“所里的同事早就听腻了我讲 AI，他们真正想看的是身边的同事、领导在实际业务中怎么用”。 ```mermaid flowchart TD accTitle: Peer-led Adoption and Measurement Flywheel subgraph S1["同行示范效应"] A["一线业务标杆案例 (如合伙人现场实操)"] --> B["引发同行共鸣与模仿意愿"] B --> C["全所使用率与 Prompt 交互量激增"] end subgraph S2["数据驱动与资源反哺"] C --> D["Prompt 分类分析与月度数据看板"] D --> E["向领导层证明 ROI 并获取数百万美元预算"] E --> F["持续赋能 AI 工具体系与变革管理"] F --> A end ``` 其次是**全面的使用量度量与数据公开（Measuring Usage）**。团队精细化追踪所有平台的用户活跃度，对提交的 Prompt 进行深度分类与场景解析，并每月向最高管理层汇报数据看板。实操表明“成功孕育成功”（success begets success），公开透明的高使用率数据会在全所营造出强大的从众效应，促使更多犹豫观望的人加入使用。更关键的是，向管理层清晰展示详实的使用数据和业务价值，直接为每年数百万美元的 AI 采购与研发预算提供了强有力的合理性支撑。最后，团队明确将 AI 落地定位为**“变革管理”（Change Management）项目**，而非单纯的技术开发。Tim 强调，AI 只是夹在输入端（人）与输出端（人）之间的中间那三分之一技术环节，若忽视前后的两大人性与组织要素，AI 战略注定无法推行。具备知识管理（Knowledge Management）背景的团队往往能更敏锐地把握人员心智与流程重塑，从而保障落地成效。 ### [09:51](https://www.youtube.com/watch?v=pu055hxdahs&t=591s) 影子 AI 的正向收编与治理作为推广加速器在企业 AI 推广过程中，切忌将所有精力都压在技术本身上，真正决定成败的往往是另外三分之二的组织与流程要素。Ogletree Deakins 摸索出的一个极为成功的策略是主动拥抱“影子 AI”（Shadow AI），即员工私下使用、未经正式批准的工具。面对这种现象，律所没有采取简单粗暴的“一封了之”，而是将其视为发现真实业务痛点的风向标：既然员工在私下使用，就说明存在刚性需求。IT 与管理层要做的是去接洽厂商，将其升级为企业级许可（Enterprise License），从而纳入正规的治理框架中。这样做不仅免去了繁琐的需求调研，因为这批“偷跑”的员工本身就是现成的试点人群，连实际用例都已验证完毕；而且还避免了直接剥夺工具引发的抵触与抱怨，员工得到的体验是工作方式得到了更强大的赋能与支持。与传统认知大相径庭的是，治理并不是让项目陷入泥潭的官僚程序，而是 AI 推广过程中不可或缺的“加速器”。通常人们很少会把“治理”（Governance）和“加速器”（Accelerator）放在同一个句子里，因为治理往往被视作让创新项目走向死亡的坟场；但在生成式 AI 时代，逻辑恰恰相反。统计数据显示，具备健全治理体系的企业推动日常 AI 普及的可能性高出 75%，成功运转大规模 AI 应用的可能性更是接近普通企业的 4 倍。缺乏治理，企业根本快不起来——无法消除持怀疑态度的客户的顾虑，无法赢得谨慎合伙人与律师的信任，更拿不到总法律顾问（General Counsel）的审批绿灯。Ogletree Deakins 通过组建多个专业委员会、制定严密的政策流程，并成功通过 ISO 42001（人工智能管理体系）认证，建立起了让全员放心的安全感。只有当人们对技术感到安心时才会主动使用；良好的治理不仅是推广的催化剂，更是让管理层夜能安寝的基石。 - **影子 AI 转化为企业工具**：把违规暗中使用变为官方赋能，自带精准用例与先锋测试群体，节省调研成本。 - **数据佐证治理价值**：健全的治理让企业日常 AI 推广成功率提升 75%，部署大规模级 AI 的几率高出近 4 倍。 - **合规标杆**：全所通过 ISO 42001 认证，以制度化保障换取客户、律师及总法律顾问的全力支持。 --- ### [11:52](https://www.youtube.com/watch?v=pu055hxdahs&t=712s) 自研还是外采（Buy vs. Build）的决策边界与实践律所目前的 50 多款 AI 工具栈中，自研与外采基本平分秋色（各占约 50%）。决定哪些该买、哪些该自研的核心原则非常简单明了：“能买到的尽量买，买不到的才自己做”（buy where we can and build where we cannot）。首先，律所绝不重复造轮子去竞争前沿基础能力（Frontier Capabilities），直接采购 Claude、Microsoft Copilot 和 ChatGPT 等平台，“Anthropic 已经是万亿级企业，我们绝不可能自己训练基础大模型去和他们竞争”。其次，对于拥有海量法律数据库与判例法等特有资产的垂直领域平台（如 Harvey、Lexis+ AI / Lexis Protégé），律所也选择外采，因为所内根本无法复刻此类全量法条与判例数据。最后，对于那些便宜且好用的小切口细分工具（Niche Products），例如市场价仅需 20 美元一份的证词摘要（Deposition Summary）工具，采购的性价比远高于内部组建团队开发维护。相反，只有在特定内部业务量巨大、流程极具独特性、且外部厂商因市场过小而不愿涉足的场景下，律所才会选择自研。例如针对海量标准化争议、低标的仲裁或特定诉讼业务组合，所内团队开发了深度契合流程的自动化工具，将案卷处理的核心环节完全自动化，这既避免了将敏感客户数据交给第三方厂商的风险，也解决了厂商无法深度适配定制化流程的问题。此外，律所内部的业务运营（Operations）流程高度个性化，市面上根本没有现成软件能够匹配；另一个典型案例是律所庞大的移民业务组（Immigration Group），放眼全美乃至全球，达到该律所移民业务规模的同行不过五到十家，极度狭小的客户基数让软件厂商根本没有动力开发专属解决方案。律所为移民团队量身打造了内部生成式 AI 工具，直接为该业务部门累计节省了数千小时的高价值工时。 ```mermaid flowchart TD accTitle: 买还是做（Buy vs. Build）决策流程图 Start["评估 AI 需求与场景"] --> Dec1{"是否属于基础大模型或核心垂直数据？ （如 Claude、Copilot、Harvey、Lexis）"} Dec1 -- 是 --> BuyFrontier["采购（Buy） 利用前沿大厂能力与现成数据资产"] Dec1 -- 否 --> Dec2{"市场上是否有成熟廉价的细分工具？ （如 20 美元一次的证词摘要）"} Dec2 -- 是 --> BuyNiche["采购（Buy） 直接采买，避免与成熟工具竞争"] Dec2 -- 否 --> Dec3{"业务流程是否高度定制？ 市场过小外部厂商不愿开发？ （如大规模自动化、内部运营、移民团队）"} Dec3 -- 是 --> BuildInternal["自研（Build） 在专属高容量场景下自建工具"] Dec3 -- 否 --> Reevaluate["重新评估需求必要性"] ``` --- ### [14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) 自研陷阱与采购契约的残酷教训关于自研，演讲者给出的核心警示是“切勿轻率构建”（build carefully）——“仅仅因为你有能力做出来，并不意味着你就应该做”。AI 降低了初期的编码门槛，催生了大量靠灵感与轻度编程（Vibe Coded）在周末搭出来的概念验证原型（POC），但这往往是巨大幻象的开端。原型做出后，真正繁重的工作才刚刚开始：必须有人进行底层代码与合规审查，必须向业务层论证其实际商业价值，而将系统推广至全所数千人规模的部署与运维工作“根本没有被 AI 自动化”。很多律师在周末用 AI 攒了一个原型，周一兴冲冲跑来找技术部门要求全所上线，得到的答复却是“没问题，需要 6 到 9 个月准备部署”，这极容易在业务端造成强烈的心理落差与挫败感。更严重的是长期的维护负债：周五下午系统崩溃而周一开庭急需准备资料时谁来应急？底层大模型版本更新或数据架构（Data Schema）变更时谁来调整？原作者休假、升职或离职后谁来编写文档？自研绝非一劳永逸，后续的维护与支持成本往往远超最初想象。而在采购决策上，同样有两记沉痛的经验教训： 1. **警惕高昂的转换成本与数据锁定**：一旦全所累积了上百万条 Prompt 与交互历史，这些沉淀的机构知识在现阶段的大模型技术下是“物理上完全无法迁移”的；此外，法务团队经过长年累月的变革管理才习惯了某套工作流，强行更换工具如同“从律师脚下把地毯抽走”。许多律所正是因为不敢承受巨大的转换成本与组织惯性，最终被落后、退化的厂商死死套牢，错失使用最新前沿工具的良机。 2. **严禁签署长期合同，单次授权期限不超过 12 个月**：前沿基础模型演进极其迅猛，很多原本独立的垂直单点工具很快就会被基础模型的原生能力全面吞没。而且从律师的日常使用心理来看，他们强烈倾向于在同一个前沿工具中完成所有工作，即使某些离散细分工具的输出效果略胜一筹，他们也厌烦在五套独立平台间切来切去。如果律所签下了多年期死合同，就会面临“工具早已被淘汰却仍须每年付款”的窘境。律所明确规定：生成式 AI 采购合同的期限绝不能超过 12 个月。 --- ### [17:41](https://www.youtube.com/watch?v=pu055hxdahs&t=1061s) 总结：全所推动 AI 规模化的五大铁律在演讲的尾声，演讲者将不到两年内让 1,100 名律师跨越百万级 Prompt 的实战经验浓缩为五条直击要害的核心心法，供所有尝试在大型组织内推动 AI 变革的领导者参考： 1. **必须将“推广落地”（Adoption）设立为专职岗位**：“工具绝不会自己把自己推广出去”，再强大的 AI 也不会自发在律所中生根发芽，必须配置专属团队与专职人员来负责组织变革与推广落地。 2. **以同行互助与同行对话为先导**：不要纯靠 IT 团队搞自上而下的单向灌输，而要从一线发掘已熟练掌握技能的业务先锋（Practitioners），让他们用律师能听懂的业务语言去向身边的同事传道授业。 3. **量化成功指标，持续担任内部的宣传旗手**：没有人会主动替技术团队宣扬成效。推动者必须主动搜集数据、建立指标仪表盘，在管理层与合伙人面前大声汇报 AI 带来的实际价值与使用规模，主导所内的讨论节奏。 4. **能买则买，迫不得已才自建**：采购时务必从第一天起就将“未来迁移与退出的转换成本”核算进预算中，并坚持签署短期合同（不超 12 个月），防止被快速迭代的技术浪潮甩下。 5. **在追求推广速度之前，先筑牢治理底座**：这是全场分享中“最重要的一件事”。良好的治理不是绊脚石，而是让企业能够放心狂奔的抓地轮胎。只有在合规与责任机制完备的前提下，前沿 AI 才能在严谨的专业服务机构中真正扎根绽放。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=pu055hxdahs&t=4s) 所以如果讲得不好，别怪我，去怪 AI 吧 — Tim Fox > [00:04](https://www.youtube.com/watch?v=pu055hxdahs&t=4s) 受制于这些限制，我们永远无法直接站在技术的最锋利尖端，必须等待技术逐渐成熟且被证明安全 — Tim Fox > [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) 建好平台，他们自然会来 — Tim Fox > [05:16](https://www.youtube.com/watch?v=pu055hxdahs&t=316s) 拥有工具绝不等于产生用量 — Tim Fox > [07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) 所里的同事早就听腻了我讲 AI，他们真正想看的是身边的同事、领导在实际业务中怎么用 — Tim Fox > [07:50](https://www.youtube.com/watch?v=pu055hxdahs&t=470s) 成功孕育成功 — Tim Fox > [11:52](https://www.youtube.com/watch?v=pu055hxdahs&t=712s) 能买到的尽量买，买不到的才自己做 — Tim Fox > [11:52](https://www.youtube.com/watch?v=pu055hxdahs&t=712s) Anthropic 已经是万亿级企业，我们绝不可能自己训练基础大模型去和他们竞争 — Tim Fox > [14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) 切勿轻率构建 — Tim Fox > [14:21](https://www.youtube.com/watch?v=pu055hxdahs&t=861s) 仅仅因为你有能力做出来，并不意味着你就应该做 — Tim Fox > [17:41](https://www.youtube.com/watch?v=pu055hxdahs&t=1061s) 工具绝不会自己把自己推广出去 — Tim Fox ## Mentioned Tim Fox, Ogletree Deakins, Legalmation, 南加州大学古尔德法学院 (USC Gould School of Law), 加州司法委员会 (California Judicial Council), Claude, ChatGPT, Harvey, LexisNexis, Protégé / Lexis+ AI, ODGPT, ISO 42001, Microsoft Copilot, Anthropic ## Video description Tim Fox - Chief Data and AI Officer, Ogletree Deakins In under 2 years, I helped take Ogletree Deakins from no AI tools to a stack of more than 50 licensed and homegrown tools, crossing well over one million prompts along the way and still climbing. We did it all while earning our ISO 42001 certification, ensuring enterprise governance for every decision. In this session, I'll cover the buy vs. build decisions behind that growth, including where Claude, Harvey, and Microsoft Copilot fit, and what worked and what didn't. I'll also share the use cases that actually drove adoption across the firm. You'll leave with lessons learned, practical use cases, and ideas on how to drive adoption at your own organization. [Source](https://www.youtube.com/watch?v=pu055hxdahs)

Source quotes

[0:04] Tim Fox: 所以如果讲得不好，别怪我，去怪 AI 吧 [0:04] Tim Fox: 受制于这些限制，我们永远无法直接站在技术的最锋利尖端，必须等待技术逐渐成熟且被证明安全 [5:16] Tim Fox: 建好平台，他们自然会来 [5:16] Tim Fox: 拥有工具绝不等于产生用量 [7:50] Tim Fox: 所里的同事早就听腻了我讲 AI，他们真正想看的是身边的同事、领导在实际业务中怎么用 [7:50] Tim Fox: 成功孕育成功 [11:52] Tim Fox: 能买到的尽量买，买不到的才自己做 [11:52] Tim Fox: Anthropic 已经是万亿级企业，我们绝不可能自己训练基础大模型去和他们竞争 [14:21] Tim Fox: 切勿轻率构建 [14:21] Tim Fox: 仅仅因为你有能力做出来，并不意味着你就应该做 [17:41] Tim Fox: 工具绝不会自己把自己推广出去

Source quotes

[0:04] Tim Fox: blame AI for that [5:16] Tim Fox: It is a hire people and you will drive adoption. I can literally show you on our usage curves when somebody was hired at the firm to help lead our adoption initiatives. [5:16] Tim Fox: governance theater [7:50] Tim Fox: People are tired of me talking about AI at my firm. They want to hear from their peers. [10:33] Tim Fox: where you send projects to die [10:33] Tim Fox: you cannot move fast without good governance [14:21] Tim Fox: just because you can build it does not mean you should build it [14:21] Tim Fox: vibe code

Original transcript

[0:04] Um, start off, I'll reintroduce myself. I'm Tim Fox. I'm the Chief Data and AI [0:09] Officer at Ogletree Deakins. For those of you not familiar with Ogletree Deakins, we are a global [0:14] law firm. We have about 1,100 attorneys across 60 offices throughout the world. We focus primarily [0:19] on employment, labor, and immigration law. Uh, a little bit more about my background. [0:25] I am also a licensed attorney and I helped found the first litigation-focused AI company back in [0:32] 2016 called Legalmation, which means I've been working in legal and AI for over 10 years now. [0:40] I also teach a course at USC's Gould School of Law fundamental fundamentals of AI to help [0:46] educate aspiring lawyers on AI and technology and I serve on the uh California Judicial Council's [0:54] uh Information Technology Advisory Committee where I also advise courts and judges on AI. [0:59] And so what I want to talk to you about today is our AI journey and how we drove [1:04] AI adoption at a global law firm. Um, before I get into that, I want to let you know I took [1:09] this deck, I took my notes and I put them in Claude and ChatGPT and you're all lucky [1:14] because it said this was going to be one of the best presentations at this conference. [1:17] Um, so if it's not, don't don't blame me. You can uh blame AI for that. So before I jump in, [1:24] I'd really want to talk briefly about what makes legal different. Legal is a a highly regulated [1:28] industry and so for those of you that are in the audience that are in the legal industry, you're [1:32] familiar with some of the challenges that we face as a law firm. For those of you that aren't, [1:38] uh the main thing I just want to highlight for you is that there are limitations, there are [1:42] restrictions that we have in legal that prevent us from adopting technology quickly. We can't ever be [1:48] on the cutting edge of technology because of these limitations that we have. We have to wait for [1:54] technology to mature before we can jump on it. Um, in my firm we take what we call a fast follower [1:59] approach where we aim to adopt technology as soon as it's proven safe for uh law firms to use it. [2:07] So, let me talk go way, way back in time um to the old days of generative AI and going back to 2023 [2:14] uh where we started our journey. And from a legal industry perspective, 2023 [2:19] was not a great time for generative AI. Uh there were hallucination stories every week. [2:24] There were lawyers getting sanctioned because the ChatGPT or other models were outputting made-up [2:28] case citations. Um and honestly the models were of questionable value for legal work at the time. [2:34] Uh and there was also genuine uncertainty among attorneys about whether AI was going to be a tool [2:38] or a threat to the practice of law. And every week there's a new study that came out. One would say [2:44] AI is going to eliminate attorneys and another that would say we're going to need more attorneys [2:47] than ever because of AI. And compounding all of that, uh you know, lawyers have been told for the [2:52] last 15 years that AI was going to replace them, that they were going to be replaced by technology. [2:57] And so attorneys are naturally skeptical. This made them even more skeptical. And in 2023 they [3:02] honestly weren't wrong to be skeptical because the technology was not ready for prime time when [3:07] it comes to attorneys. Um that said, my firm we were lucky. We saw the potential early on and [3:13] our leadership treated AI as a strategic priority starting in 2023. We created our first generative [3:19] AI strategy in July of that year. In 2024 we started putting that strategy to work. We piloted [3:25] every major platform we could get our hands on. We signed license agreements with Harvey, [3:29] Lexis Protege, and a bunch of other platforms. Um and we launched our first firm-wide homegrown [3:35] chatbot which we called ODGPT in October of 2024. And the reason I'm highlighting that is because [3:41] I think of that is really the start of our AI adoption journey when we launched this firm-wide [3:46] tool uh at the firm. In our first month we had 9,000 prompts in ODGPT. Jumping ahead 2 years, um [3:54] we've had explosive growth of AI at the firm. We've grown from just one ODGPT tool to we have 50 [4:00] plus AI tools available at the firm split roughly between what we bought and what we've built. [4:07] We crossed 1 million cumulative prompts in May of this year. So we went from our first prompt to [4:12] our millionth prompt in under 18 months. We're averaging close to 250,000 prompts per month [4:18] through our AI platform and that number is still climbing. We don't know what the upper limit is, [4:23] but we will pass 2 million prompts before the end of the year and expect that to keep increasing. [4:27] Roughly 80% of our attorneys use our AI tools monthly. We have over 2,000 monthly active users [4:33] across our platforms at the firm each month. We also recently obtained our ISO 42001 certification [4:40] in June making us one of the first companies in the world, but also one of the first law firms [4:44] in the world to have that certification. One other thing we've been able to do [4:49] with our AI adoption is partner with our clients to develop custom workflows, AI enabled processes [4:55] that let us work faster, cheaper while still maintaining the same quality of output. I will [5:01] say most law firms don't publicly share statistics like this. So I can't necessarily benchmark us to [5:06] other firms, but when I do talk to firms and when I do talk to clients, the general consensus is [5:10] that we are ahead of the market when it comes to our AI adoption and the impact on the practice. [5:16] So I've talked about our successes and where we are now. I want to talk about some of the stuff [5:20] that didn't work to share some of the pain that we experienced first hand. And I've got a few [5:24] listed here. The first is, you know, taking the build it and they will come approach. When we [5:29] launched these tools, we assumed people would find them. They did not. That approach does not work. [5:35] Having tools does not equal usage. You have maybe 10 to 20% of your early adopters at your firm or [5:40] your company, they will find the tools, they will come, but everyone else and attorneys especially, [5:45] they need hand holding. They need somebody to show them how to use that technology. Uh another [5:51] thing that didn't work for us was not being staffed for adoption. Early on, it was myself, [5:55] and I had one other attorney working with me, uh and we did everything from vetting to piloting [6:00] to deploying to training, uh and we realized it it doesn't work uh with that few people. [6:05] If you want this to work, you have to staff it. It's not really a chicken and the egg problem. [6:11] It is a hire people and you will drive adoption. I can literally show you on our usage curves when [6:16] somebody was hired at the firm to help lead our adoption initiatives. Something else that did not [6:21] work for us was treating our firm as one audience. Uh law firm is really two distinct organizations. [6:26] You have your attorneys, and you have your business professionals, and they both work in [6:30] completely different worlds. Our early focus was almost exclusively on attorneys. They're the ones [6:35] who generate the revenue. We assumed if we got them using AI, the business professionals would [6:40] follow. Um however, what we found is business professionals aren't going to use a platform like [6:44] Harvey because they see Harvey as an attorney-only tool. So, even if you tell them you can use it, [6:49] they're not going to use it. And what we learned is we need to develop tools specifically for them. [6:54] The last thing that we learned um is what I call governance analysis paralysis. Uh early [6:59] on when we're getting ready to launch OTGPT, we were terrified of what our attorneys were going [7:03] to do with this platform. Were we going to get in the headlines? Was I going to get fired? What [7:06] was going to happen with that? So, we focused on drafting policies, conducting trainings, [7:10] and doing all of these things that in hindsight I call governance theater because they made us feel [7:14] a little bit better, but they didn't actually move the needle, and all they did was delay our launch. [7:19] Um and so, the key takeaway that we had was you there's no way to mitigate all risk with [7:24] generative AI. If you want to use it, you have to accept that there is some level of risk in [7:29] the platform, and ultimately you have to trust your users to do the right thing. [7:33] You have to give your users all the right tools, all the right training to use it responsibly, [7:37] but ultimately it's in their hands, and you have to trust them to do it right. [7:40] And one other thing that comes from that is if you're willing to move earlier and accept some [7:44] of that risk before others, that actually gives you a competitive advantage in your industry. [7:50] So that was what did not work. Now let me talk about some of the things that did work really [7:53] well for us during our adoption. Um the first is peer-led training. Uh letting peers talk to peers [8:00] is one of the things that worked best for us. An example of this was uh 6 months ago we had a [8:05] California-only shareholder meeting. I had one of our shareholders up talk about how he used [8:09] AI in his practice. It was standing room only. We had 100 people in that room. Everybody was paying [8:14] attention. Um and that one session moved forward adoption with those users more than any webinar, [8:20] any training series, anything else that I could have done. People are tired of me talking about [8:24] AI at my firm. They want to hear from their peers. And this is actually borne out by studies. [8:29] I have some stats down on the bottom here. But in general, the key takeaway is that employees learn [8:34] better from other employees than they do from trainers, than they do from other people. They [8:38] want to see their coworkers, their managers use the tools and tell them how they're using them. [8:43] Another thing we did well uh was measuring our usage. So we track our uses across all [8:47] our platforms. We're classifying our prompts to see how people are actually using them. [8:51] And we report to leadership monthly on those usage stats. And what we found is that success [8:57] begets success. Um the more usage we show, the more people want to use the platforms. [9:01] The more they hear about them and see everybody else using them, the more they want to use them. [9:05] So making that data available has helped us immensely in driving more adoption. [9:10] The other thing that sharing these metrics widely has helped us with is funding. These tools are [9:14] expensive. We spend millions of dollars on AI tools. And to help justify that cost, [9:19] we can show leadership the amount of usage that we have and how valuable they are at the firm. [9:25] Uh one other thing that we did well was recognizing this is change management. Um [9:29] at its heart, that's all that AI adoption is. And so those of us coming out of knowledge management [9:33] or similar departments, we're generally better at change management. Um treating AI as just a [9:38] technology project is not going to help you with adoption. If you think about it, you have people [9:43] on the input, you have people on the output. AI is just that one middle piece. So, if you focus just [9:47] on that one-third that's technology and ignore those two-thirds, you're not going to have any [9:51] success in your adoption. You need to focus on the other two-thirds in that equation. The other thing [9:56] that we did well was looking at our shadow AI. So, looking at the unapproved tools at the firm that [10:01] people were using. By looking at those tools, we figured out where there was a need. So, instead of [10:06] going out and banning those tools, what we would do is go figure out how we can get them on an [10:09] enterprise license so that we could then pull them into our governance framework. And that worked [10:13] well because people have already shown there's a need for those tools. We already have a built-in [10:17] pilot group, and we already have use cases. And that saves us a lot of time for our jobs cuz [10:22] we can just go out and get them these tools and set them up. It also avoids complaints that you [10:26] get when you take away access to platforms because we're just enabling them to use them in a new way. [10:33] One other thing that worked well for us that I want to highlight is the importance of governance. [10:38] You very rarely hear the term governance and accelerator in the same sentence. Governance [10:44] is typically where you send projects to die. But with AI, it's the opposite. Good governance is a [10:49] key to AI adoption. And it's not just me saying that, I've got statistics up here that you can [10:55] see as well. The one study from earlier this year found that with good governance, you're 75% more [11:00] likely to to drive a day AI adoption, and you're nearly four times more likely to have a gigantic [11:06] AI running if you have good governance. And I can tell you we see this internally. You cannot move [11:11] fast without good governance. Good governance is how you're going to win over your reluctant [11:15] clients. It's how you win over your attorneys. It's how you win over your GC to get approval [11:19] to use tools in your firm. For us, we've formed multiple committees, we have policies, we have a [11:25] whole bunch of stuff in place. As I mentioned, we recently got our ISO 42001 certification. And that [11:30] governance is what makes people comfortable with AI. If they're not comfortable with AI, [11:35] they're just they're not going to use it. Um, and so if you want to move quickly with AI, [11:40] you have to start with governance. It will pay dividends. The other thing it will help you do is [11:45] sleep at night knowing that you've done all you can to make sure AI is being used responsibly. [11:52] The last topic that I want to cover with my time here is talking about our buy versus build [11:57] decisions here. I mentioned earlier that our tech stack is split pretty evenly between tools that [12:02] we bought versus tools that we have built. And so I want to talk about how we decide whether [12:07] to buy versus build and some of our lessons learned. So how do we decide what we buy? [12:12] Well, we're going to buy the frontier capability. We buy the Clauds, we buy the Copilots, we buy the [12:17] ChatGPTs. Those are things we cannot build and things we should not try to build. Anthropic is [12:22] a trillion-dollar company for a reason. We're not going to build our own model to compete with that. [12:27] We're also going to buy the domain-specific platforms. We're going to buy the Harveys, we're [12:31] going to buy the Lexus Proteges. Uh, those are platforms that have unique knowledge that we're [12:35] not capable of replicating internally, either. We don't have access to all of the case law and [12:40] statutes that Lexus has. Uh, we're also going to buy where there's a a niche product that works [12:45] well and is very cost-effective that we don't want to necessarily compete with. There are a number of [12:50] very specific tools in the legal market that do a good job and they're not super expensive. Um, [12:55] deposition summaries, we're not going to build a deposition summary tool because we can go [12:58] buy one for $20 a pop. It just makes more sense to do that. On the other hand, [13:04] uh, we're going to build where we cannot buy. Where we have internal volume that's high enough, [13:08] we have specific processes, it's a unique space where a vendor just isn't going to spend the time [13:13] to build a tool there. So in short, we buy where we can and we try to build where we cannot buy. [13:20] So just talking briefly about some of examples of what we've built, uh, we've done workflow [13:25] automation. I mentioned this earlier for large portfolios, large portions of matters where we can [13:30] take them, low stakes arbitrations, other types of litigation, and automate big portions of that [13:35] using our own internal tools. No vendor's going to do that for us. We're not going to trust a vendor [13:39] with our data to do that even if they would. Uh we also do a lot of uh automation for our internal [13:44] operations, for our business professionals. Um there's simply no vendor, no tool that's going [13:48] to do that. Our internal workflows are too customized. Uh no vendor has an off-the-shelf [13:52] solution that's going to fix that. Another area we do a lot of building is for our immigration group. [13:58] Um we need very specific tools for immigration practice group. We have a very large group. [14:03] There's immigration group that maybe five, 10 other firms that are as large as us when [14:06] it comes to immigration, and there's no vendor that's going to serve a market of just five to 10 [14:11] um clients. So that's an area where we find that we build a lot of tools to help save [14:15] time. We've saved thousands of hours for our immigration group in how we leverage AI for them. [14:21] Continuing on uh on buy versus build, I do want to share some lessons that we've learned [14:26] on on that as well. Um the first one being build carefully. Just because you can build it does [14:34] not mean you should build it. Building is never a one-and-done thing. When you're building a tool, [14:39] you need to actually think beyond the tool itself. You need to think about who's going to support [14:44] it when it breaks on a Friday afternoon and an attorney has a trial on Monday and they need to [14:48] use that tool over the weekend prepare for trial. You need to think about who's going to update it [14:53] when the latest model comes out or when the data schema changes and you need to address that. [14:57] Uh who's going to document its functions when the engineer who built it goes on leave or gets [15:01] promoted or leaves the company altogether? Um the other thing you need to think about with [15:07] building is it's gotten both easier but also harder with the advent of AI. You can now do [15:13] a proof of concept in a weekend, but deploying at an enterprise scale is still a lot of work, [15:19] and generative AI has made that much more complex than less. Somebody now has to review that vibe [15:24] coded application that you built. Somebody has to take it to the business to figure out is this [15:28] actually valuable for our firm? Should our firm actually spend the time to deploy this tool? [15:34] And deployment itself has not been automated by AI. That part still takes a lot of manual work. [15:40] So it really is a a double-edged sword. You can build that proof of concept now in the weekend and [15:45] a lot of our attorneys do and then they come to us on Monday and ask us to deploy it and we come back [15:50] and say sure we'll have it up in 6 to 9 months. Uh and that leads to a lot of frustration among [15:56] users. And so it's something to be cautious of when you start encouraging the build process [16:01] is thinking through long-term those steps and the impact on it. For buy I've I've [16:06] got two lessons here for you to think about. The first is to think about the switching costs. Um [16:11] a million prompts of institutional knowledge that's not easy to move. If we had to switch [16:15] vendors that would be very difficult for us to do and it would be challenging from the [16:20] data perspective because it's largely impossible. There's no way right now to transfer between one [16:25] uh AI platform and another but also from a change management and relationship standpoint, right? We [16:29] spent years telling attorneys to use these tools. We can't pull the rug out from underneath them. [16:34] So when you're looking at investing in a large vendor for a large scale deployment, think about [16:39] what this looks like long-term um because if you have to migrate off that vendor it's going to be [16:43] difficult. And this is what leaves a lot of law firms stuck with vendors that have fallen behind. [16:49] Um they have this institutional inertia they can't overcome which prevents their attorneys from being [16:54] able to use the latest and the greatest and the best products that we have in the legal space. [16:59] The other side of that is not over committing to long-term contracts. We've actually gotten rid of [17:04] several platforms because they've been subsumed by what the frontier models can do or by what [17:09] other providers can do. Uh and our attorneys would also much rather use one platform that does things [17:17] uh one frontier platform as opposed to five discrete platforms to do that same task. Even [17:21] if those five discrete platforms give them slightly better output. They all just want [17:26] to stay in the same platform there. So, if we'd signed, you know, multi-year deals with these [17:30] companies that we no longer need, we'd be stuck with them. So, our go-to rule is generally not to [17:36] have longer than a 12-month contract when we're signing anything for generative AI. [17:41] So, with my last minute here, I want to talk just briefly about what to take away from the session [17:48] here for you. And from all of our our hard earned lessons that we had. So, first, make sure you [17:55] staff adoption as its own function. The tools will not distribute themselves. They will not drive [18:01] adoption. You need people to do that for you. The second is lead with peer-led conversations. Find [18:08] your AI practitioners and let them teach their peers. Let them lead those conversations for you. [18:14] Third, measure your success and champion your usage regularly. Nobody is going to be that [18:20] champion for you. So, be the one showing your organization how much these tools are being [18:25] used. Be the one to report into leadership and drive that discussion forward. Fourth, [18:30] buy when you can and build when you cannot. And when you're thinking about buying, make sure you [18:36] price switching costs into the original decision and try to keep your contract length short. [18:41] And the last one, which you take nothing else from my session, [18:45] is make sure you build governance before adoption. Governance is going to make you faster. Governance [18:50] is going to make your job easier. I could spend an entire hour talking about the importance of [18:54] governance and what we've done at my firm. But from my perspective, that is the most important [18:59] thing that anyone who's looking to drive AI adoption can do. And with that, I know my time is [19:04] up. So, thank you all very much. My email address, my contact information is up here. Please feel [19:08] free to reach out if you have any questions. I'm always happy to chat about this. Thank you all.