Inside the AI Workday of Thomson Reuters' CEO: How Leaders Drive Enterprise Transformation by Example

The Authenticity Key to Enterprise AI Transformation: Thomson Reuters CEO's Monday Morning Workflow

Amid the wave of enterprise AI transformation, the biggest challenge is often not the technology itself, but internal organizational adoption and cultural reshaping. Thomson Reuters CEO Steve Hasker recently shared his personal daily AI usage, offering business leaders a thought-provoking reference point on how to genuinely drive AI adoption.

Background on Thomson Reuters' AI strategy: Thomson Reuters is a global leader in information services and media, with core business lines spanning legal, tax, compliance, and news. As a company with annual revenue exceeding $6 billion, its AI transformation didn't start from scratch—back in 2023, the company announced it would invest over $100 million in generative AI development over the following two years, and launched CoCounsel (built on GPT-4), an AI assistant product for legal professionals. This context lends particular weight to Hasker's remarks: the "authenticity" he speaks of is grounded in actual product deployment and large-scale organizational change, rather than remaining at the conceptual level.

It's worth adding that CoCounsel was not developed in-house by Thomson Reuters, but acquired in 2023 through the roughly $650 million purchase of its original developer, Casetext. Founded in 2013, Casetext initially started as a legal research database, and after the release of GPT-4 quickly integrated its capabilities into its products, becoming one of the first companies to commercialize large language models in legal professional scenarios. By the time Thomson Reuters completed the acquisition, Casetext already had over 10,000 law firm customers—meaning Thomson Reuters purchased not just technology, but market-validated product-market fit and customer trust assets.

This acquisition decision reflects the "buy rather than build" strategy commonly adopted by large information service providers in the AI era. The underlying logic is that the competitive window for generative AI capabilities is extremely short: building an internal R&D system from scratch typically takes 3-5 years, while the market landscape could be reshaped within 18 months. At the same time, the core moat of legal AI products is not the model itself (foundation models can be accessed via API), but the data flywheel, workflow integration, and user trust built around specific professional scenarios—assets that can be acquired directly through M&A rather than cultivated internally from the ground up.

The competitive landscape of legal AI: Legal Tech is one of the most fiercely contested professional vertical markets for generative AI deployment. Emerging players like Harvey AI, Lexis+AI (owned by LexisNexis), and Ironclad are challenging traditional giants with standalone products, while Microsoft is also accelerating its penetration through legal-industry-customized versions of Copilot for Microsoft 365. The core challenge of legal AI lies in the demand for "high trustworthiness"—any factual error in legal documents could have serious legal consequences, making this field's requirements for AI accuracy far higher than in general enterprise scenarios. This context means Thomson Reuters' AI transformation carries the dual pressure of internal change and market defense: it must both drive thousands of employees to change how they work internally, and fend off AI-native competitors whose valuations are rapidly climbing.

The Key to Transformation Success: Authenticity

Hasker believes that the success or failure of large-scale AI transformations like the one Thomson Reuters is currently undertaking hinges on "authenticity." This means leaders can't merely advocate verbally—they must actually put the tools to use.

"Our biggest learning at Thomson Reuters is that we're all learning, we're all on this shared learning journey together," Hasker described the transformation. This statement itself is telling: even the CEO of a giant in content and information services admits that the entire organization is still in an exploratory phase. This top-down candor is precisely an important prerequisite for dispelling employee concerns and encouraging experimentation.

Why is "authenticity" so critical? In organizational behavior and change management, a leader's "role modeling" is one of the most powerful mechanisms for driving cultural change. McKinsey research shows that over 70% of failed digital transformation projects can be attributed to organizational and cultural resistance, rather than technical barriers. When leaders merely issue strategic statements without hands-on practice, employees often develop a "this is just another fad" mentality, adopting a wait-and-see or passive-compliance attitude. By making "authenticity" the core of AI transformation, Hasker is essentially responding to this deep organizational dynamics problem—behavioral consistency from leaders is the key lever for breaking organizational inertia.

This logic has deep theoretical roots in change management research. Albert Bandura, founder of Social Learning Theory, pointed out as early as the 1970s that organization members tend to calibrate their own actions by observing the behavior of authority figures—when a CEO showcases AI-generated summaries in meetings or personally demonstrates how to use tools, the adoption signal conveyed is far more persuasive than any policy document.

Notably, this role-modeling effect is especially important in promoting AI tools: unlike previous enterprise software, the effectiveness of generative AI tools largely depends on the user's level of "prompt engineering," and this skill can only be accumulated through actual use—it cannot be taught through training manuals. Prompt engineering refers to guiding a large language model to produce more accurate and useful outputs by carefully designing the structure, wording, and contextual information of input text—for the same task, prompts written by an experienced user versus a beginner can yield vastly different results. This means a leader's public practice is not just the transmission of a cultural signal, but an actual demonstration of the organization's "AI usage water line"—a leader practicing in public thus becomes the most effective path to lowering the learning threshold for employees.

The grounding point of authenticity is very concrete: every Monday morning, the first thing Hasker does is personally use these AI tools. When the leader himself becomes a power user of the tools, the signal he sends is far more persuasive than any internal email or all-hands meeting.

The CEO's Monday Morning AI Workflow

Hasker's approach to using AI is not mysterious—in fact, it's quite down-to-earth, revolving mainly around three daily work scenarios.

On the deployment logic of AI tools in executive workflows: The three types of AI usage scenarios Hasker describes—document analysis, calendar insights, and news summaries—correspond to the most mature application directions for enterprise-grade AI assistants today. Generative AI tools represented by Microsoft Copilot and Google Gemini for Workspace are already deeply integrated into email, calendar, and document systems. Research firm Gartner predicts that by 2026, over 50% of knowledge workers will use AI assistants daily to handle information-intensive tasks.

The deeper driving force behind this trend comes from the "information overload" dilemma that knowledge workers have long faced. Cognitive science research shows that modern executives spend an average of about 2.5 hours per day processing email, with substantial additional attention consumed by document reading and meeting preparation. The concept of information overload was first proposed by sociologist Alvin Toffler in his 1970 book Future Shock, but it wasn't until the popularization of the mobile internet and real-time communication tools that its erosion of executive decision-making efficiency was truly quantified by research. The core value of AI assistants in this scenario is not to replace human judgment, but to serve as an "information preprocessing layer"—compressing vast unstructured inputs into structured decision summaries, freeing up executives' scarce attention for the aspects that truly require human judgment. For executives who need to process large volumes of structured and unstructured information in a short time, AI's most direct value lies in "compressing the time chain from information to decision"—precisely the core logic demonstrated by Hasker's Monday morning workflow.

Document Analysis and Information Extraction

He uses AI tools to analyze the documents he needs to handle that day and throughout the week, quickly distilling key information. For an information company executive who faces massive volumes of reports, legal content, and news every day, this kind of document-processing capability directly affects decision-making efficiency.

From a technical implementation standpoint, this document analysis capability relies on the breakthrough of large language models (LLMs) in the "long context window." A token is the basic unit an LLM uses to process text, roughly corresponding to 0.75 English words or 1.5 Chinese characters—it can be understood as the basic granularity of the model's "reading memory," and the token limit determines how much content the model can "see" in a single processing pass. From GPT-3's 4,096-token limit, to GPT-4-turbo's 128,000 tokens, to Google Gemini 1.5 Pro's breakthrough support for 1 million tokens—roughly equivalent to the length of 750 ordinary novels—the leap in this technical metric directly determines the practical value ceiling of AI document analysis. A complex cross-border merger agreement, a complete litigation file, or a quarterly financial report can now be fed into the model in its entirety for analysis, without manual splitting or segmentation—this is the key turning point that has transformed AI document tools from "concept demos" to "everyday utility."

Calendar Management and Meeting Preparation

Hasker has AI sort through his calendar, extracting insights and preparing him for the week's work. "Going through my calendar, getting insights out, being prepared, whatever the task is—that's how I start my Monday mornings." This practice of embedding AI into personal workflows demonstrates the practical value of general-purpose AI assistants in executives' daily routines.

Calendar intelligence is one of the core battlegrounds in the current enterprise AI assistant competition. Microsoft Copilot can automatically generate agenda summaries and attendee backgrounds before meetings, while Google Gemini can identify scheduling conflicts and suggest priorities. A more advanced application is "post-meeting action item extraction"—AI automatically identifies to-do items from meeting recordings or transcripts and pushes them to the responsible parties. The core technology behind these features is a combination of Natural Language Understanding (NLU) and Named Entity Recognition (NER): the model needs to identify names, time points, task descriptions, and responsibility assignments from unstructured conversation text, then structure them into actionable to-do entries. The essence of these features is transforming information-integration work that once required an executive secretary or assistant into an AI-native capability available to everyone—which carries particular practical significance amid the trend toward organizational flattening.

Gaining a Global Perspective Through Podcasts

Notably, Hasker also listens to the Reuters News World podcast. Since the program is primarily recorded in the UK, by the time he wakes up on Monday mornings, the program is already ready, providing him with a global news overview. This clever use of the "time-zone difference" reflects his careful orchestration of the rhythm of information consumption.

As Thomson Reuters' core news brand, Reuters has over 200 news bureaus worldwide, producing an average of about 3.5 million words of news content daily. By leveraging the 5-8 hour time-zone difference between Europe and the US, one can obtain a summary of the previous day's global events early in the morning on the US East Coast. This "time-zone arbitrage" in media consumption is no accident—financial media like Bloomberg and the FT have long turned morning briefings into core subscription products, targeting precisely the executive population's need to establish a comprehensive information framework before decision-making begins.

From a more macro information-ecosystem perspective, podcasts as an information medium have unique advantages: they suit passive consumption in "fragmented scenarios" (such as commuting or morning exercise), and can convey tone, emotion, and in-depth analysis better than text content. From a media economics standpoint, podcasts also represent an efficient "attention capture" strategy—occupying a listener's time while they perform low-cognitive-load tasks, a window that text content struggles to reach. Data from both Spotify and Apple Podcasts show that the proportion of executive listeners for news podcasts is far higher than the general population average—which aligns closely with Hasker's usage habits. By incorporating Reuters' own podcast into his personal workflow, Hasker offers both a genuine endorsement of the product and a reflection of the media strategy logic by which information service providers deeply bind content products to executive audience behavior patterns.

Continuous Learning Both Inside and Outside the Organization

Beyond personal practice, Hasker also emphasizes an "outward-looking" learning mechanism. He continually asks his team several key questions:

• Who are the early movers?
• Who are the fastest movers?
• Who is making the most mistakes—and therefore accumulating the most lessons learned?

This perspective is particularly intriguing. Hasker does not treat "making mistakes" as a negative signal, but rather sees it as the best source of learning.

"Failing fast" as an organizational learning strategy: Hasker's special attention to "those who make mistakes" echoes the "fail fast" philosophy long advocated in innovation management. This concept originates from agile development and lean startup methodologies—Eric Ries systematized it in The Lean Startup as the "Build-Measure-Learn" loop—and its core logic is that in highly uncertain environments, small-scale, low-cost trial and error can accumulate effective knowledge faster than large-scale planning.

However, transplanting this methodology to the AI transformation of large enterprises comes with real-world constraints: enterprise-grade AI deployment involves upfront costs such as data governance, compliance review, and system integration, meaning the assumption of "low-cost trial and error" doesn't always hold. In regulation-sensitive industries like law and finance, certain "mistakes" may directly trigger compliance risks. A deeper constraint comes from the "asymmetry of AI errors"—in Thomson Reuters' legal AI scenarios, an incorrect legal citation caused by a single model hallucination could have consequences far exceeding the cost of fixing a software bug. It is precisely for this reason that Hasker emphasizes "learning from external mistake-makers"—essentially a smarter alternative strategy: gaining experience by observing the trial-and-error costs paid by others, building the organization's "AI usage intuition" at lower cost, while avoiding the potential price of "hands-on trial and error" in high-risk scenarios.

In a field as highly uncertain as AI application, rapid experimentation and rapid iteration often accumulate genuine organizational capability more effectively than cautious observation.

He also requires his team to continuously gather feedback from peers, competitors, customers, and partners, bringing external experience back inside to accelerate the company's overall transformation. Notably, in the context of AI transformation, the target of competitive intelligence has expanded from traditional product pricing and market share to "AI capability maturity benchmarking": tracking competitors' AI tool deployment depth, employee adoption rates, the cadence of customer product feature iterations, and publicly shared failure cases. For Thomson Reuters, the product roadmaps and customer feedback of Harvey AI and LexisNexis constitute the most strategically valuable signal inputs in its external scanning system.

This approach builds a two-way learning loop: on one hand, accumulating internal experience through the hands-on practice of the CEO and team; on the other, absorbing industry best practices through broad external information scanning.

The theoretical foundation of the two-way learning loop: The learning mechanism Hasker describes—combining internal practice with external scanning—is an organizational variant of what knowledge management calls "double-loop learning." This concept, proposed by management scholar Chris Argyris, emphasizes that organizations must not only learn from outcomes (single loop) but also continuously reflect on and revise their underlying assumptions (double loop). In the context of AI transformation, internal practice accumulates operational "know-how" about how to use AI, while external scanning brings back strategic "know-what" about what AI can do and where its boundaries lie.

This framework also aligns closely with the "competitive intelligence" practices of information science: systematically tracking competitors' AI adoption paths and failure cases essentially converts the trial and error scattered across the market into the organization's structured knowledge assets. Data from the McKinsey Global Institute shows that in digital transformation, companies that actively conduct external benchmarking have transformation success rates about 40% higher than those focused solely on internal improvement. Only through the synergy of both can enterprises avoid the pitfall of "only knowing how to execute head-down, without looking up to calibrate direction" during AI transformation.

Three Takeaways for Business Leaders

Hasker's final advice is simple yet powerful: "Get immersed, use the tools yourself. Make sure you keep an ongoing conversation with your team about the benefits, the potential pitfalls, and the lessons learned around the tools."

This passage distills his methodology for driving AI transformation, which can be broken down into three levels:

First, leaders must get hands-on. Only by actually using the tools can you understand their capability boundaries, and only then can you maintain authenticity while driving change. Leaders who only talk about AI strategy without taking action struggle to genuinely bring the organization up to speed.

Second, confront the tools' limitations. Hasker doesn't blindly tout AI's benefits—he explicitly mentions the "pitfalls." This balanced attitude helps prevent organizational over-optimism and reduces unnecessary setbacks during subsequent deployment. The main "pitfalls" of current generative AI include:

• Factual errors caused by model hallucination—large language models may generate citations or data that seem plausible but are actually fabricated. The root of hallucination lies in the LLM's autoregressive generation architecture: at each step, the model predicts the next word based on statistical patterns in its training data, rather than on explicit storage and retrieval of facts, making it prone to "creating" rather than "recalling" information in domains where training data is sparse. In legal scenarios, there are already real cases of lawyers being sanctioned by courts for submitting documents containing fabricated case citations. The industry currently mitigates this primarily through Retrieval-Augmented Generation (RAG) technology—RAG works by first retrieving relevant document fragments from an external knowledge base before the model generates its answer, injecting them as "reference materials" into the model's input context, so that the model's output is grounded in evidence rather than generated out of thin air. CoCounsel adopts precisely this architecture, combining the precise retrieval of the Westlaw database with GPT-4's reasoning capabilities to reduce the probability of legal citation errors.
• Data privacy and compliance risks—feeding sensitive business data into third-party AI services may trigger data leaks or regulatory compliance issues. Especially under regulatory frameworks like the GDPR (EU General Data Protection Regulation) and China's Personal Information Protection Law, enterprises need to clearly define which data can be uploaded to public AI services.
• Degradation of human judgment from over-reliance on AI—long-term outsourcing of tasks like document analysis and summary generation to AI may lead to the atrophy of employees' relevant professional skills, which is especially worth guarding against in legal and medical scenarios requiring deep domain expertise.
• Vast ROI differences across scenarios—AI excels at repetitive, standardized tasks, but in scenarios requiring high creativity or complex interpersonal judgment, its benefits often fall far short of expectations.

All of these require establishing corresponding usage norms and review mechanisms at the organizational level.

Third, establish a mechanism for ongoing dialogue. AI transformation is not a one-time tool deployment, but a dynamic process requiring continuous communication, feedback, and adjustment. Regular team dialogues centered on practical experience are key to making transformation truly take root. At the operational level, this dialogue mechanism can take many forms: setting up internal forums for sharing AI use cases, establishing cross-departmental AI communities of practice, or fixing an AI transformation progress segment into quarterly business review meetings. The concept of the Community of Practice was proposed by cognitive scientists Jean Lave and Etienne Wenger in 1991, referring to a knowledge-sharing group that spontaneously forms among people with a shared focus on a particular field—in the context of AI transformation, such cross-departmental communities can often break through the barriers of formal organizational structures, allowing AI practitioners from different business lines to directly exchange experiences and thereby accelerate the entire organization's learning curve. Research from MIT's Center for Digital Business found that among enterprises leading in AI adoption, about 85% have established some form of internal knowledge-sharing mechanism—precisely the key bridge for converting individual employees' AI exploration into organization-level capability accumulation.

Conclusion

This AI routine of the Thomson Reuters CEO is essentially an action guide on "how leaders drive technological change by example." Its value lies not in the use of any particularly sophisticated AI technology, but in demonstrating a pragmatic, sincere, and systematic attitude toward transformation—start by using it yourself, learn from mistakes, and advance through dialogue.

Examined from a more macro industry perspective, Hasker's practice also reflects the unique historical situation facing this generation of business leaders: they are the first batch of CEOs who need to complete the leap "from the pre-AI era to the AI-native era" within their tenure. They have neither a mature management textbook to reference nor immunity from the multiple time pressures coming from boards, investors, and competitors. Against this backdrop, "hands-on practice, maintaining candor, and continuous learning" is not only a methodology for organizational change, but also a survival wisdom for maintaining leadership relevance in an era of technological acceleration. For enterprises and managers standing at the crossroads of AI transformation, this line of thinking may be more valuable as a reference than any grand technological blueprint.

Key Takeaways

Key Takeaways