Thomson Reuters Cuts Case Law Search to Minutes in Anthropic Agent Report

AI agents are software systems that use frontier models to plan and execute multi-step workflows with limited human supervision, typically connected to enterprise tools and data sources.

Anthropic's 2026 State of AI Agents Report, based on a survey of more than 500 technical leaders conducted in late 2025 with research firm Material, found that 80% of organizations report their AI agent investments are already delivering measurable economic returns. The finding signals that enterprise AI agents have moved from pilot projects to production deployments that are being measured against business outcomes. The report pairs that aggregate statistic with two named customer deployments, Thomson Reuters and eSentire, that quantify time savings in legal research and cybersecurity analysis.

In legal, Thomson Reuters uses Claude to power CoCounsel, its AI legal assistant. Lawyers who once spent hours manually searching through documents can now access 150 years of case law and 3,000 domain experts in minutes, according to the report. In cybersecurity, eSentire compressed expert threat analysis from 5 hours to 7 minutes, with 95% alignment to senior security experts. Both figures describe work performed in production, not benchmark demonstrations.

For enterprise executives, the implication is that AI agents are no longer an experimentation line item. The survey indicates most organizations have moved past single-task assistants toward agents that execute multi-stage workflows, and a majority now tie those deployments to measurable economic returns. The practical question for chief executives and chief technology officers is no longer whether agents work, but which workflows should be converted first and how fast the organization can absorb the change.

The report's aggregate ROI figure and the two customer case studies provide a template for how boards should evaluate agent investments: choose a high-frequency, high-judgment process, measure the time-to-outcome before and after deployment, and validate output quality against senior staff.

80% of organizations report their AI agent investments are already delivering measurable economic returns.Anthropic, The 2026 State of AI Agents Report

Metric	Before	After	Source
Thomson Reuters legal research	Hours of manual document searching	Minutes to access 150 years of case law and 3,000 domain experts	Anthropic report
eSentire threat analysis	5 hours of expert analysis	7 minutes with 95% alignment to senior security experts	Anthropic report
Enterprise agent ROI	No baseline reported	80% of organizations report measurable economic returns	Anthropic report

What does the 2026 State of AI Agents Report show?

The report is the latest in Anthropic's annual series on enterprise AI adoption. It is based on a survey of more than 500 technical leaders, conducted in partnership with research firm Material in late 2025. The report defines the current state of agents by workflow scope, use case maturity, and reported barriers, and it distinguishes between agents that handle discrete tasks and agents that coordinate multi-stage processes.

By workflow scope, 57% of organizations deploy agents for multi-stage workflows, while 16% deploy them for cross-functional processes spanning multiple teams. That distribution suggests the center of gravity has shifted from single-turn assistants to agents that can move across steps within a function. Cross-functional agents remain the exception, which is consistent with the integration and data challenges the report identifies.

Coding remains the most mature agent use case. Nearly 90% of organizations use AI to assist with coding, and 86% deploy agents for production code. Respondents reported similar time savings across the development lifecycle: planning and ideation (58%), code generation (59%), documentation (59%), and code review and testing (59%). These numbers are close enough to suggest the benefit is spread across the workflow rather than concentrated in code generation alone.

Outside software development, data analysis and report generation was the highest-impact non-coding use case, cited by 60% of organizations, followed by internal process automation at 48%. For executives in functions such as finance, operations, and compliance, those use cases are where agent ROI is most likely to appear outside the engineering organization.

Why are agents moving from pilots to production?

The survey numbers point to maturation in how enterprises evaluate agents. Rather than measuring model capability in isolation, technical leaders are reporting on time savings, output quality, and alignment with senior experts. The eSentire result, in which automated analysis aligns with senior security experts 95% of the time, is an example of a quality threshold that makes speed gains usable in production.

Adoption is not frictionless. The report's biggest challenges are system integration (46%), data quality (42%), and change management (39%). These are organizational and architectural constraints, not model limitations. The same barriers explain why only 16% of organizations have reached cross-functional agent deployments.

What did Thomson Reuters achieve with Claude-powered CoCounsel?

Thomson Reuters has integrated Claude into CoCounsel, its legal research product, and the companies expanded the partnership in May 2026. According to Thomson Reuters, a new Model Context Protocol integration brings CoCounsel Legal into Claude workflows to help legal work meet fiduciary-grade standards, and the next generation of CoCounsel Legal is rebuilt on Anthropic's Claude Agent SDK. The MCP connection lets legal professionals invoke CoCounsel research inside Claude's agentic interface rather than switching between applications.

For legal teams, the operational change is access speed. The report says lawyers who once spent hours manually searching through documents can now access 150 years of case law and 3,000 domain experts in minutes. That compression changes how legal research is staffed and how quickly attorneys can validate precedent, although the report does not quantify the revenue or billing impact of the change.

How did eSentire compress threat analysis from 5 hours to 7 minutes?

eSentire's deployment is a cybersecurity example with a different character. Threat analysis is a time-sensitive, high-judgment process in which a five-hour delay can be the difference between containing an incident and responding after damage. The Anthropic report says eSentire compressed expert threat analysis from 5 hours to 7 minutes, with 95% alignment to senior security experts.

The alignment figure is important because it addresses the main objection to autonomous security analysis: that speed is worthless if output quality drops. A 95% alignment rate means the agent is producing conclusions that match senior analysts often enough to be used as a first-pass or triage mechanism, with exceptions escalated for human review.

What technical patterns are driving measurable ROI?

The report's aggregate data suggests ROI correlates with workflow scope and integration depth. Organizations deploying agents for multi-stage workflows, 57% of the sample, are more likely to capture time savings across an entire process than organizations using agents for isolated tasks. The 16% that have reached cross-functional deployment represent the frontier, and they face the hardest integration problems.

1. Deploy agents on multi-stage workflows, where 57% of organizations report adoption, rather than limiting them to single tasks.
2. Connect agents to production data and systems early, since 46% of organizations cite system integration as a top challenge.
3. Prioritize code generation, documentation, and review, where 59% of organizations report time savings across the development lifecycle.
4. Expand into data analysis and report generation, cited by 60% of organizations as the highest-impact non-coding use case.
5. Build change management into the rollout plan, because 39% of organizations identify it as a barrier.

The Thomson Reuters and eSentire deployments follow a similar pattern: the agent is connected to specialized enterprise data, its output is validated against expert judgment, and the time saving is measured before and after deployment. Those are the conditions under which the report's 80% ROI figure becomes plausible for a specific organization.

How should organizations measure agent ROI?

The report's ROI evidence is built on time-to-task comparisons rather than revenue attribution. Thomson Reuters' case law access and eSentire's threat analysis are both measured in elapsed time, which makes them easier to audit than productivity estimates. The 80% aggregate figure, by contrast, is based on self-reported assessments from technical leaders, a softer metric that executives should triangulate with internal pilots.

A practical measurement framework has three components: a baseline time and quality score for the current process, a target for the agent-assisted process, and a review mechanism that samples agent output against senior staff. In the eSentire case, that review mechanism produced the 95% alignment figure. In the Thomson Reuters case, the baseline was hours of manual document search.

What should C-suite executives take from the report?

The report gives chief financial officers a benchmark: 80% of organizations report their AI agent investments are already delivering measurable economic returns. The most defensible metrics are time savings, such as the legal research and threat analysis examples, because they can be measured in hours and minutes without relying on revenue attribution.

Chief information officers and chief technology officers should weigh the report's top barriers before scaling. System integration was cited by 46% of organizations, data quality by 42%, and change management by 39%. These findings imply that agent ROI is constrained less by model capability and more by the quality of enterprise data and the willingness of teams to change workflows.

The survey is self-reported and vendor-commissioned, so the percentages should be treated as directional rather than independent benchmarks. Anthropic conducted the survey with research firm Material, and the customer figures come from Anthropic's own report and a Thomson Reuters press release. That does not diminish the value of the time-to-task comparisons, but it argues for validating results internally before making investment decisions.

What do executives say about the Thomson Reuters-Anthropic partnership?

The Thomson Reuters press release frames the expanded partnership around reliability in high-stakes professional environments. In the release, Scott White, Head of Product, Enterprise at Anthropic, said the work reflects a deeper strategic partnership rather than a point integration.

Our work with Thomson Reuters reflects a deeper strategic partnership to deliver AI that can operate in high-stakes professional environments.Scott White, Head of Product, Enterprise at Anthropic

White's comment points to a broader enterprise pattern: agents are being embedded in products that carry professional responsibility, such as legal research and security operations. The combination of Claude's agentic capabilities and CoCounsel's domain data is the model for how Anthropic says it intends to serve regulated industries.

For boards, the report raises a governance question: when an agent performs work that previously required professional judgment, who is accountable for the output? Thomson Reuters' framing of fiduciary-grade standards and eSentire's 95% alignment benchmark are attempts to answer that question with product design and validation metrics. Executives should expect similar accountability requirements in regulated industries.

What happens next for AI agents in the enterprise?

The report's data points to expansion beyond software development. Data analysis and report generation, cited by 60% of organizations, and internal process automation, cited by 48%, are the next frontier. Cross-functional agents remain early at 16%, which suggests the largest gains will come when organizations solve integration and data quality enough to connect agents across departmental boundaries.

The pace of that expansion will depend on how quickly enterprises address system integration and data quality, the two most cited barriers. For executives, the actionable sequence is to identify a high-frequency workflow with measurable before-and-after time, connect the agent to the relevant data, validate output against senior staff, and then expand to adjacent processes. The Thomson Reuters and eSentire cases provide the template.