Agentic AI Systems for Legal, Tax, and Compliance Workflows

thomson_reuters · Legal · 2024 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Legal, tax, and compliance

Thomson Reuters needed agentic systems that could make judgments and produce work in high-stakes expert workflows.

Constraints:

Architecture

Thomson Reuters frames agentic systems as adjustable agency dials rather than a binary architecture. Autonomy ranges from atomic tasks to self-evolving workflows; context ranges from model parametric knowledge to RAG and multiple controlled/web sources; memory ranges from stateless retrieval to persistent shared workflow memory; coordination ranges from single tasks to tool delegation and multi-agent collaboration. Those dials sit on top of decomposed legacy applications: tax engines, validation engines, source-document extraction, proprietary legal content, search/fetch/citation tools, and evaluation rubrics from domain experts.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  agency-dial-controller[Agency dial controller] --> autonomy-dial[Autonomy dial]
  agency-dial-controller --> context-dial[Context dial]
  agency-dial-controller --> memory-dial[Memory dial]
  agency-dial-controller --> coordination-dial[Coordination dial]
  coordination-dial --> legacy-application-tools[Legacy application tools]
  legacy-application-tools --> tax-engine[Tax engine]
  legacy-application-tools --> legal-research-tools[Legal research tools]
  context-dial --> proprietary-content[Proprietary content]
  legal-research-tools --> proprietary-content
  evaluation-rubrics[Evaluation rubrics] --> agency-dial-controller
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
Agency dial controller Use-case decision layer setting autonomy, context, memory, and coordination levels. Autonomy dial, Context dial, Memory dial, Coordination dial
Autonomy dial Controls whether work is atomic or self-planning/replanning. Agency dial controller
Context dial Controls parametric, RAG, controlled, web, and evolving data-source context. Agency dial controller, Proprietary content
Memory dial Controls stateless, workflow-shared, or persistent cross-session memory. Agency dial controller
Coordination dial Controls single-task execution, tool delegation, or multi-agent collaboration. Agency dial controller, Legacy application tools
Legacy application tools Decomposed tax, legal, compliance, and research systems exposed to agents. Coordination dial, Tax engine, Legal research tools
Tax engine Calculation and validation engine used by tax-return agents. Legacy application tools
Legal research tools Search, fetch, citation comparison, and citation validation tools. Legacy application tools, Proprietary content
Proprietary content 1.5-plus terabytes of Thomson Reuters legal/tax/compliance content. Context dial, Legal research tools
Evaluation rubrics Domain-expert evaluation criteria for variable expert workflows. Agency dial controller

Data flow

A tax or legal workflow sets the agency dials appropriate to risk. For tax, source documents are parsed, fields are mapped to a tax engine, rules are applied, validation errors feed back into more document retrieval, and the return is completed. For legal research, the assistant uses litigation-research tools to search, fetch, compare, and validate citations over proprietary content before producing an answer.

3 supporting quotes
“Uh I think we've been talking about this a little I think we we like to define it more as a spectrum.”transcript, implied
“But what agents have allowed us to do is to really decompose these legacy applications and decompose the components of them as tools that agents can now use.”transcript, stated
“Uh we also have a validation engine that's built into that tax engine that the the AI system can use to validate the work that it's doing can inspect the errors. It can go look for more information from the documents when it needs it and and resolve to finish the the workflow.”transcript, stated

Why this architecture

Thomson Reuters uses dials because risk tolerance varies across legal, tax, and compliance tasks. Some workflows can tolerate high agency and exploration, while high-precision tasks need constrained autonomy, controlled context, and validation. Decomposing legacy systems as tools preserves decades of domain logic instead of rebuilding everything from scratch.

3 supporting quotes
“And so we view these things as levers that we are able to kind of move up and down depending on what our users are willing to tolerate in terms of the the risk of the situation that they they may be dealing with.”transcript, stated
“We were leaving all that behind us and building something new somewhat from scratch. But what agents have allowed us to do is to really decompose these legacy applications and decompose the components of them as tools that agents can now use.”transcript, stated
“we're finding new ways to leverage a lot of these legacy applications and infrastructure that um you know previously we might have thought of as baggage but I think are really unique assets for us uh to build on going forward.”transcript, stated

Models used: LLMs

Scale

Alternatives considered (roads not taken)

Binary agent/non-agent framing

Rejected because: The source argues agency is a tunable spectrum by use case and risk.

“Uh it is not that this system is agentic or it is not but in fact um these are dials that can be used to uh to to sort of tune what the experience and how much agency the experience has for the user depending on the use case.”transcript, stated

Rebuild AI workflows from scratch

Rejected because: Legacy systems contain tuned domain logic and validation engines that agents can use as tools.

“But we've have 100 plus years of building software systems that have highly tuned domain logic. Uh and and our users expect this sort of logic in the way that they work.”transcript, stated

Patterns

tool-use-orchestration (expert-workflow)

Legacy applications are decomposed into tools that agents can use.

“They developed a framework treating agency as adjustable dials (autonomy, context, memory, coordination) rather than binary states, enabling them to decompose legacy applications into tools that AI agents can leverage.”summary, stated

agent-memory (workflow-state)

Memory is treated as a configurable dial in the agency model.

“They developed a framework treating agency as adjustable dials (autonomy, context, memory, coordination) rather than binary states, enabling them to decompose legacy applications into tools that AI agents can leverage.”summary, stated

Tradeoffs

How to define agentic behavior.

Chose adjustable dials over binary agent/non-agent classification.

Rationale: The summary says Thomson Reuters treats agency as dials rather than a binary state.

“They developed a framework treating agency as adjustable dials (autonomy, context, memory, coordination) rather than binary states, enabling them to decompose legacy applications into tools that AI agents can leverage.”summary, stated

How to set agency in expert workflows.

Chose risk-calibrated autonomy/context/memory/coordination dials over one uniform agent behavior.

Rationale: Legal and tax workflows vary in precision requirements and user tolerance for autonomous exploration.

“you may not want to dial the agency up in those situations”transcript, stated

Evaluation

Thomson Reuters emphasizes rigorous evaluation for expert judgment variability.

Metrics: expert judgment variability · workflow quality

“5+ terabytes of proprietary content, with rigorous evaluation processes to handle the inherent variability in expert human judgment.”summary, stated

Outcomes

The transcript references 1.5+ TB of proprietary content in the summary.

“return generation from source documents and comprehensive legal research systems that utilize their 1.5+ terabytes of proprietary content, with rigorous evaluation processes to handle the inherent variability in expert human judgment.”summary, stated