Thomson Reuters推出可自行编写并执行代码的CoCounsel税务、审计与会计AI平台
Thomson Reuters debuts new CoCounsel for tax, audit and accounting pros
Thomson Reuters已推出CoCounsel for Tax, Audit and Accounting AI平台,可由代理自行编写并执行代码,完成复杂多步骤工作流。平台可处理折旧、摊销等计算,生成带公式的Excel工作簿、Word和PDF成果,并按事务所工作底稿模板更新文件。系统还能通过Python重算并核对数值,该平台现已面向美国全部客户开放。
平台把复杂计算、工作底稿生成与多步骤流程交给代理执行,其代码重算校验方式为专业服务自动化边界提供了具体参照。
- Key insight: Programming accounting software to write and run its own code for complex calculations.
- What's at stake: Shifting human accounting labor from performing calculations to reviewing automated spreadsheet outputs.
- Expert quote: "Building AI for professions is much harder than trying to build for general purpose productivity or helping you to summarize emails or helping you book concert tickets. AI for tax and audit and accounting is hard." — David Wong, Thomson Reuters
Thomson Reuters released its CoCounsel for Tax, Audit and Accounting AI platform, which sports enhanced agentic capacities to write and run its own computer code to execute complex, multistep workflows for professionals.
Processing Content
At a high level, it can perform calculations including depreciation, amortization, required minimum distributions, qualified business income, threshold checks and multi-scenario comparisons; gather client documents and communications in one place so everyone can gain a full context of the client relationship; build its own multi-tab Excel workbooks with live formulas and a documented assumptions tab; draft memos and other deliverables in Word and PDF; and update the spreadsheets and documents uploaded by human professionals following the structure of a firm's own workpaper templates.
It also has native Model Context Protocol connectivity, making it possible for CoCounsel to operate within a range of other applications and systems, and to incorporate other systems into its own context. It also boasts expanded integrations with UltraTax, Engagement Manager, Guided Assurance, Ready to Review and Ready to Advise.
David Wong, chief product officer, said in a briefing on Wednesday that while previous versions had been capable of performing simple tasks, it struggled whenever work was more complex, or people needed multiple tasks strung together, or things deviated just slightly from what it was taught. What was needed, he said, was a new AI capable of complicated problem-solving and independent tool use so it could do things like interpret complex tax forms, produce workpapers or handle tough audit questions. Considering the profession's low tolerance for error, complex layered rules, and reliance on outside context, he admitted this was no easy feat.
"Building AI for professions is much harder than trying to build for general purpose productivity or helping you to summarize emails or helping you book concert tickets. AI for tax and audit and accounting is hard," he said.
Key to developing this new functionality was giving the AI the ability to write and execute its own code. Lucas Adams, vice president of engineering, said that agents before would generally read and write text and then users would connect that text into the code themselves, which was how "you can sort of simulate an agent doing these more complex tasks." However, he said it is fundamentally different when an agent actually writes code itself, whether snippets or whole files, and then executes it. The AI is not just, for example, taking a CSV output or text extraction from a document, it is looking at the information natively just how a computer program would.
"Why is this so important in a domain where we're building agents to do tax and accounting work? It's not software engineering. But it's really a core abstraction that allows you to do a lot more than you can without it," he said.
He said it is like giving an AI agent its own computer, which it can then use like any human might. This helps especially when it needs "ad hoc deterministic things" like running and verifying calculations.
During a demo, head of product Dave Ragones showed the platform building a full amortization table for a $5 million commercial real estate loan after loading in the loan agreement and template. In very little time, it returned with a filled in Excel spreadsheet with a loan summary, a complete amortization table, a balloon payment calculation and further notes for follow-up, with everything sourced and cited.
"When we're calculating the balloon payment as well as the interest payment in the spreadsheet, you don't see this in the product, but behind the scenes we recalculated that in Python code," said Adams. "It's just another way for us to verify that the numbers in the calculation are correct, so that's all enabled by the ability to write and run code."
He noted that the AI built everything itself — the template, the workbook, the Excel files — using code it wrote and executed "rather than just something like a CSV." Looking deeper into the AI's actions, he also noted that it checks and corrects itself in real time.
"At the end, you can see that there was actually some failure in the way the schedule was built. It fixed that failure, it rebuilt, and then it verified the values," he added. "We're actually taking the logic of the formulas from Excel, we're rewriting them in a different system — in this case in Python code — and rerunning it, and tying those numbers together to verify that they match."
This ability allows for a great deal of automation. Ragones noted it can create finished work products as well as provide explanations on how its work was done. Overall, it is designed to take on a great deal of tasks, both simple and complex, for accounting, audit and tax professionals.
"Instead of just asking CoCounsel and getting an answer back, we want to empower CoCounsel to do much, if not all, of the work," he said. "And the professional role in that process is, of course, to be the human in the loop to review it, to verify, to make sure that the work is correct. But we want to empower co-counsel as much as possible to be able to do that tax work, the audit work, the accounting work that all these firms do,"
Chief technology officer Joel Hron noted that the code execution capacities, while impressive, are backed by Thomson Reuters' vast library of content, allowing it to do the work to the standards demanded by professionals. It does not just perform tasks; it exercises judgment.
"We didn't just build a tool that does math. We built one that knows why the math is right," said Hron in a statement. "The code execution piece is table stakes now; every company in this space has it. What's hard is making sure the system is reasoning from the right facts and the right authority before it ever runs a calculation. We built that judgment in, grounded in content that's kept current and tested by the people who actually do this work, so a professional can check it themselves rather than take our word for it."
The new CoCounsel Tax, Audit & Accounting platform is available now for all customers in the United States, with additional features and capabilities to come.
Introductory bullet points created by AI with editorial review.
Technology Editor
Chris Gaetano is the technology editor for Accounting Today. He brings with him more than a decade of experience covering the accounting profession... Read full bio
来源:Accounting Today(美国会计行业媒体) · accountingtoday.com