Thomson Reuters launches rebuilt tax AI that produces work for human review
CoCounsel can plan assignments, build spreadsheets and explain calculations. Its practical test is whether checking that work takes less effort than doing it manually.
Loading page…
CoCounsel can plan assignments, build spreadsheets and explain calculations. Its practical test is whether checking that work takes less effort than doing it manually.
Listen to this story
Thomson Reuters’ rebuilt CoCounsel Tax, Audit & Accounting turns complex requests into multi-step work products, such as spreadsheets and memos, rather than stopping at conversational answers. Its outputs include citations, assumptions and calculations for professionals to inspect and approve; the company has not reported measured results across customer assignments. The October 22 launch is available to all U.S. customers, with access controls and approved-source limits that constrain what the system can use.
In a company demonstration, CoCounsel built a $5 million loan amortization workbook with changing rates, a lump-sum payment, a balloon payment and live formulas.
The demonstration also showed Python-based recalculation, but executives said the practical test is whether verification takes less effort than manual preparation.
Checkpoint contains more than 12 million primary-source and analytical documents, while shared client context can retain uploaded files and completed work across conversations.
Thomson Reuters launched a rebuilt CoCounsel Tax, Audit & Accounting platform on October 22, moving beyond answers to producing work professionals can inspect and approve. The company says it is available to all U.S. customers. Its pitch pairs broader automation with citations, assumptions and calculations—not an end to human review.
The predecessor was built around individual tasks, Chief Product Officer David Wong explained during a briefing. The new version is designed to take a complex request, divide it into steps and choose the tools and information needed to carry it out. That changes the assignment from answering a question to executing a sequence of work.
Thomson Reuters says the platform combines AI agents—software that selects and carries out steps—with client information, tax research and code execution. Assignments include building loan repayment schedules, preparing spreadsheets and drafting memos. The intended output is a document or calculation ready for professional review, rather than just a conversational response.
The distinction is important in tax work: a calculation can be mathematically correct while applying the wrong rule or missing a client's circumstances. CoCounsel supplies citations, assumptions and calculations so professionals can assess its conclusions before approving them. They remain responsible for reviewing and verifying the results.
Its research source is Checkpoint, Thomson Reuters' collection of more than 12 million primary-source and analytical documents, maintained by hundreds of experts. A shared client context layer also preserves uploaded documents and completed work across conversations, reducing the need to supply the same background repeatedly.
In a demonstration covered by SiliconANGLE, direct tax and audit product head David Ragones used a commercial loan agreement and a firm's spreadsheet template to generate an amortization schedule. The example involved a $5 million loan with changing interest rates, a lump-sum payment and a balloon payment due at maturity.
The resulting workbook included live formulas, supporting assumptions and references. Users could select individual cells and ask CoCounsel to explain the calculations and their sources. That made the demonstration about inspecting the spreadsheet's construction as well as receiving the finished file.
The demonstration also showed CoCounsel recalculating spreadsheet results with Python routines as an accuracy check. These were company-demonstrated capabilities, not a measured result across customer assignments. Executives acknowledged that the practical hurdle is whether checking the output requires 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
Lucas Adams, Thomson Reuters vice president of engineering, during the briefing covered by SiliconANGLE
The system's ability to select tools does not mean unrestricted access to files or the web. Thomson Reuters described limits on what agents can reach:
The platform supports multiple AI models and integrates with Thomson Reuters' UltraTax, Engagement Manager and Guided Assurance. It also supports Model Context Protocol, a way to connect AI tools with other systems. Those connections allow cited tax and accounting answers to be accessed through third-party assistants, including Anthropic's Claude and Microsoft's Copilot.
Wong linked demand to staffing shortages, saying firms want to manage workloads without increasing seasonal hiring. Thomson Reuters also sees an opportunity to shift professionals toward advisory work. That is the company's intended payoff for automating the preparation of work that still needs an expert's approval.
Loading discussion...
Join the conversation
Explain which approach would make mistakes easier to catch.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.