The accountant whose AI sent the wrong tax summary.

Robert runs a three-person CPA firm in Houston's Galleria area. Primarily individual and small business tax work, more than 200 active clients. Midway through a prior tax season, he adopted an AI document summarization tool to speed up client briefing packet prep.

The process

The AI pulled prior-year return data, generated a two-page briefing summary per client, and cut packet prep from 45 minutes to under seven. Staff loved it.

What almost went wrong

During a peak week in February, with more than 40 client meetings across three days, the AI tool auto-populated an email template and pulled the prior-year summary for a client named Michael Chen. The firm had two clients with that first name, different last names, different EINs. A recent template update had changed how the tool resolved name-to-file matching. The wrong Michael's summary attached to the outgoing email and went out before anyone caught it.

The fix

Two structural changes. First: every outbound document email now requires a human to manually confirm client name plus the last four digits of the EIN before any attachment can be sent. A 15-second step that cannot be bypassed. Second: the AI tool's file-matching logic was updated to use EIN as the primary key, not client name.

The AI still does all summarization. A human presses send.

What this engagement is, and is not

The Robert engagement is the worked example for the human-in-the-loop pattern across regulated verticals. The pattern generalizes wherever a wrong outbound message creates a compliance event: HIPAA-touching healthcare comms, financial advisor disclosures, immigration paperwork. The verification step is deliberately non-AI and deliberately short. The model drafts. The human confirms. Both signatures are required for the message to leave the building.