Keeping the Human in the Loop: Failure Points of Generative AI in Small-Business Accounting and Tax Compliance

Resumo

Generative AI is becoming part of everyday accounting and tax work in small firms and small businesses, and its productivity gains are real. This paper argues that those gains hold only when a qualified person stays in the loop. Drawing on the author's practice as founder of a U.S. tax firm and IRS Certifying Acceptance Agent, and on the professional and legal literature, it identifies four failure points that appear when people are removed or bypassed. The first is classification error: fluent output that assigns transactions to the wrong account or tax category. The second is advisory over-reliance: owners treat an LLM's answer as authoritative and act on it without seeing its assumptions, sometimes triggering IRS notices and penalties. The third is taxpayer data exposure: returns and statements containing Social Security numbers uploaded to general-purpose tools. The fourth is a compliance and accountability gap under IRC §7216, the FTC Safeguards Rule, IRS Publication 4557 and Circular 230. Cases from the author's practice, the abrupt closure of an online bookkeeping provider and an exploratory run of 26 questions on three free chatbots illustrate each point. In that single run per tool, given the same 60 real bank transactions, the chatbots produced a refusal, a fourfold profit range and a cash-flow total presented as profit; the demonstration shows how the failures arise and does not estimate error rates or rank the tools. The paper closes with human-in-the-loop principles: where review is indispensable, how to minimize the data a tool sees, and how to document who checked what.