汤森路透为税务与审计软件 CoCounsel 加入 AI 智能体

Thomson Reuters Corp. today is launching a rebuilt version of its CoCounsel Tax, Audit & Accounting platform designed to move beyond answering questions to planning assignments, performing calculations and producing documents for professional review.

The enhanced platform combines artificial intelligence agents with client information, tax research and code execution to complete complex tasks such as building amortization schedules, preparing spreadsheets and drafting memos. The company said the new platform is available to all U.S. customers.

The release targets a challenge in applying generative AI to professional services: A mathematically correct answer can still be wrong if the system uses the wrong tax rule, overlooks client circumstances or fails to explain its reasoning.

“They expect close to perfection with any tool that they use,” Chief Product Officer David Wong said of tax accountants during a briefing.

Unlike its predecessor, which was designed around individual tasks, the rebuilt system can break complex requests into steps and select tools and information needed to execute them, Wong said. Professionals still must review and verify the results. The system provides citations, assumptions and calculations to help them assess its conclusions before approving work.

The research database is Checkpoint, Thomson Reuters’ collection of more than 12 million primary-source and analytical documents, maintained by hundreds of experts. The platform supports multiple AI models.

In a demonstration, David Ragones, head of product for direct tax and audit, showed how CoCounsel could use a commercial loan agreement and a firm’s spreadsheet template to produce an amortization schedule for a $5 million loan. The example included changing interest rates, a lump-sum payment and a balloon payment at maturity.

The resulting workbook contained live formulas, supporting assumptions and references. Users can select individual cells and ask CoCounsel to explain the calculations and their sources.

The demo also showed the system recalculating spreadsheet results using Python routines as a check on their accuracy. “A lot of times showing the work is actually doing the work,” said Lucas Adams, Thomson Reuters’ vice president of engineering.

Security controls limit agents to files permitted by the user’s access and the relevant workspace. Adams said those restrictions extend into the underlying infrastructure. Internet access is limited to approved government websites alongside Checkpoint.

A shared client context layer preserves uploaded documents and completed work across conversations, reducing the need to repeatedly supply background information. Integrations include Thomson Reuters’ UltraTax, Engagement Manager and Guided Assurance.

Support for Model Context Protocol allows connections with other systems and access to cited tax and accounting answers through third-party AI tools, including Anthropic PBC’s Claude and Microsoft Corp.’s Copilot.

Wong said staffing shortages are helping drive demand as firms seek to handle workloads without expanding seasonal hiring. Thomson Reuters also sees an opportunity to free professionals for advisory work.

Executives acknowledged that the practical test is whether reviewing AI output takes less effort than producing it manually. “If it takes just as much time to verify the work than it takes to do the work, AI is not really helpful,” Adams said.

Photo: Unsplash