SUMM

Thompson Reuters' initial focus for AI assistants was on being helpful, referencing information accurately, and providing citations.
Recently, the priority has shifted from simple helpfulness to productivity, expecting AI to produce output, make judgments, and decisions on behalf of users.
High-stakes environments (law, tax, global trade, risk, and fraud) require higher accuracy due to the severe consequences of errors.
Thompson Reuters has a long history in these fields, with a strong base of domain expertise and proprietary content.

The company employs 4,500 domain experts, primarily lawyers, making it the largest employer of lawyers globally.
Possesses over 1.5 terabytes of proprietary domain content in legal, tax, compliance, audit, and risk.
Invested over $3 billion in acquisitions in recent years and spends over $200 million annually on AI product development.
Operates an applied research lab with over 200 scientists and engineers working with development teams.

There's a significant industry shift from building "agentic tools for law firms" to constructing "law firms of agents."
Agentic AI refers to systems with varying degrees of autonomy, moving beyond just being helpful to taking actions and making decisions.
Agency is viewed as a spectrum with adjustable "dials" like autonomy, context, memory, and coordination, depending on risk tolerance and use case.
Autonomy dial ranges from performing simple tasks to self-evolving workflows where the AI plans and replans its work dynamically.
Context dial evolved from using static model knowledge to integrating multiple and even dynamic knowledge sources.
Memory dial importance has grown, with systems needing persistent memory across complex workflows and sessions.
Coordination covers the evolution from single-task execution to delegation and multi-agent collaboration.

Evaluation (eval) of AI systems is complex and crucial for building trust, hindered by AI's non-deterministic nature.
Both users and internal teams struggle with inconsistency in evaluation; human experts’ assessments can vary by over 10% on identical cases.
Referencing reliable source material grows harder as agency increases; tracing decision drift and building robust guardrails requires deep domain expertise.
Rigorous evaluation rubrics are developed, but user preference remains a key north star in assessing if systems are improving.
Legacy applications, often seen as outdated, are now valuable as decomposable assets—agents can utilize their highly tuned domain logic as tools.
The concept of MVP (minimum viable product) can be limiting; building out the whole system first often yields better understanding and results.

Demonstration of a tax workflow where AI automates document ingestion, data extraction, mapping to tax engines, applying tax law rules, validating results, and generating returns.
Success in automation is due to legacy tools (tax engines, validation engines) which AI can leverage for calculations and error checking.
Legal research application: AI uses litigation research tools to search, validate, and compare legal documents, statutes, citations, and blogs.
The AI tracks its evidence, writes intermediate notes, and compiles a final report with explicit, traceable citations and risk flags.
These examples highlight the benefits of decomposing legacy systems and building full-fledged products that agents can operate within.

Start by considering the entire problem and solution space when building agentic AI systems, rather than focusing on minimal components.
Agency should be dynamically adjusted depending on the use case’s risk and the user's tolerance.
Leverage and break down old systems into components that agentic systems can utilize uniquely.
Maintain human-in-the-loop evaluation for trust, especially in high-stakes and specialized domains.
Differentiate products by capitalizing on unique assets, such as proprietary data and deep domain expertise.

Question addressed on how Thompson Reuters ensures security and compliance, especially for regulated or government-related clients.
Emphasis on compliance with government and industry standards (e.g., FedRAMP, ISO).
Security posture is described as adaptive, ensuring alignment with rapidly evolving standards.
Detailed technical documentation is available for further specifics on architecture and compliance.

From Copilot to Colleague: Building Trustworthy Productivity Agents for High-Stakes Work - Joel Hron