跳到主要內容

AI報稅代理評測:真實稅務工作僅過10.8% | AI Tax Agent Review: Real Tax Work Pass Rate 10.8%

By Kit 小克 | AI Tool Observer | 2026-10-30

🇹🇼 AI報稅代理評測:真實稅務工作僅過10.8%

AI報稅代理最近被官方評測狠狠打了臉:在Surge AI最新公布的「DAYJOB: Tax」評測中,表現最好的Claude Sonnet 5.5在50項真實稅務工作裡,也只拿下10.8%的通過率。這次評測不是考選擇題,而是把真正的報稅案件——合約、W-2、K-1、國稅局通知信——一股腦塞給AI代理,看它能不能像真人會計師一樣從頭做到交件。

AI報稅代理評測怎麼做的?

Surge AI找來稅務專家,設計了涵蓋8個執業領域的50項任務,包括個人報稅、小型企業、合夥事業、遺產規劃、跨國稅務與稅務爭議。每項任務平均附上30份文件,最多到123份,模擬真實客戶丟給會計師的那堆資料。評分標準嚴格:要達到95%的評分項目才算通過,涵蓋計算正確性、佐證文件、是否抓到關鍵問題,以及最終交付品質。

各家AI模型的報稅代理成績單

  • Claude Sonnet 5.5:10.8%(第一名)
  • Claude Fable 5.1:9.6%
  • Claude Opus 5.5:9.2%
  • GPT-6 Astra:4.8%
  • Gemini 4 Argon:3.2%
  • Kimi K3、Mistral Large 4、Qwen 3.8 Max:0%,一項都沒過

AI代理輸在哪裡?不是算錯,是沒判斷力

Surge AI的報告點出關鍵:這些AI報稅代理不是算術出錯,而是缺乏專業判斷力。例如有個案例要求判斷房屋出售時機,AI代理正確算出稅額,卻完全沒發現客戶如果拖到2027年8月之後賣房,會直接失去25萬美元的免稅額度——這題AI代理5次嘗試全部失敗。另一個案例是退休帳戶分配策略,AI代理沒去核對文件裡互相矛盾的婚姻狀態證據,就直接給建議。

換句話說,AI報稅代理擅長「照單全收執行」,但真實世界的稅務工作常常是「客戶沒問的問題才是真正的問題」——這恰好是資深會計師的價值所在。值得一提的是,另一個專攻企業稅務研究的「Tax Agent Bench」排行榜上,Qwen 3.8 Max反而拿下49.66%第一名,顯示不同評測測的是不同能力:研究型問答和端到端專業判斷是兩回事。

給一般人的建議

如果你是會計師,AI報稅代理目前還搶不走你的飯碗——它連自己的審查都通不過。如果你是想用AI省報稅費用的一般人,這份評測是個警訊:AI可以幫你草擬、算數字、整理文件,但千萬別把它當成唯一的稅務判斷來源,尤其涉及截止日期、資格認定這類魔鬼藏在細節裡的問題。

好不好用,試了才知道。


🇺🇸 AI Tax Agent Review: Real Tax Work Pass Rate 10.8%

A new benchmark just gave AI tax agents a reality check: in Surge AI's DAYJOB: Tax evaluation, the best-performing model, Claude Sonnet 5.5, passed only 10.8% of 50 real-world tax assignments. This isn't a multiple-choice quiz — it's actual client work, complete with contracts, W-2s, K-1s, and IRS notices, testing whether an AI tax agent can go from client intake to final deliverable like a real accountant.

How the AI Tax Agent Benchmark Works

Surge AI worked with tax professionals to build 50 assignments spanning eight practice areas: individual returns, small business, partnerships, estates, international tax, corporate transactions, and tax disputes. Each task ships with a median of 30 workspace files (up to 123) — PDFs, spreadsheets, emails, receipts, payroll records — mirroring the messy file pile a real client hands over. Grading is strict: an attempt only passes at a 95% rubric threshold, covering calculation accuracy, supporting evidence, catching consequential issues, and deliverable quality.

Tax Agent Benchmark Results by Model

  • Claude Sonnet 5.5: 10.8% (leader)
  • Claude Fable 5.1: 9.6%
  • Claude Opus 5.5: 9.2%
  • GPT-6 Astra: 4.8%
  • Gemini 4 Argon: 3.2%
  • Kimi K3, Mistral Large 4, Qwen 3.8 Max: 0% — zero tasks passed

Why AI Tax Agents Fail: It is Judgment, Not Math

Surge AI's report is blunt: these AI tax agents are not failing at arithmetic — they are failing at professional judgment. In one case, an agent correctly calculated the tax on a home sale but missed that selling after August 2027 would cost the client a ,000 exclusion — it failed that check in all 5 attempts. In another, an agent recommended a retirement-account split strategy without first resolving conflicting marital-status evidence buried in the client's files.

The pattern: AI tax agents are good at executing instructions, but real tax work is full of the question the client did not know to ask — exactly where experienced accountants earn their fee. Interestingly, a separate benchmark, Tax Agent Bench, which focuses on corporate tax research questions rather than end-to-end client work, ranks Qwen 3.8 Max first at 49.66% — a reminder that AI tax agent benchmarks measure very different skills depending on the task shape.

What This Means for You

If you are an accountant, your job is safe for now — these agents cannot even pass their own audit. If you are a taxpayer hoping AI can replace a professional, treat this as a warning: AI is fine for drafting and number-crunching, but do not trust it alone on deadline-sensitive or eligibility questions, where the real risk hides.

好不好用,試了才知道。

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code