納税者のための AI

A tax practice's year follows a predictable cycle, and AI fits into different stages of it. Pre-season (late fall): firms send organizers and engagement letters, update their written information security plan, and set up tools. This is when to check how vendors handle client data. Intake (January to March): W-2s and most 1099s arrive early in the year, and partnership K-1s often arrive much later. Document AI classifies files, extracts fields and flags gaps against the prior year. It works best on standard forms and least well on scanned brokerage statements and K-1s with long footnotes. Preparation and review (February to April 15): review tools compare returns with the prior year and look for inconsistencies. Language models can draft client emails and explain notices in plain words. Extensions (April to October 15): the same tools help manage extended returns. The rules do not change because software did the work. The preparer signs the return and can face penalties under IRC §6694 for unreasonable positions. Circular 230 requires due diligence from practitioners. For the earned income, child and education credits and head-of-household status, Form 8867 requires the preparer to make and record reasonable inquiries, and a chatbot cannot do that on their behalf. Confidentiality is its own risk. IRC §7216 and its regulations restrict how preparers use and disclose return information. Some disclosures to service providers for return preparation are permitted, while others, including certain disclosures outside the United States, require client consent. The FTC Safeguards Rule requires tax preparers to keep a written information security plan, and IRS Publication 4557 gives safeguarding guidance. The most common misconception is that pasting client documents into a public chatbot is harmless. It can breach confidentiality rules and the firm's own security plan.