What an AI bookkeeper should never post automatically.

A bank description alone may not establish business purpose. New vendors, vague payment processors, owner-related activity, unusual cash withdrawals, and conflicting prior categories often need a document or question.

The agent should hold, explain the gap, or create a focused request rather than pick the most common account and move on.

Loans, equity contributions, distributions, asset purchases, inter-account transfers, card payments, tax payments, and clearing activity can affect more than an expense category. Incorrect treatment can duplicate cash movement or misstate liabilities and equity.

Automation may still help match or prepare these items, but the clean path needs strong structural evidence—such as both transfer legs, a liability account, or an approved policy.

Voiding, deleting, unposting, reopening periods, replacing ledger history, changing billing or integrations, sending company-owner messages, and bulk-changing many records have consequences beyond one category decision.

These actions should require explicit intent or confirmation and a current preview where the impact can drift.

A coded row with an active warning is not a clean ready row. A transaction inside a locked period should not be posted because the model believes the category is obvious. Pending bank activity should not be treated as settled.

Hard state validation must outrank conversational urgency.

Materiality, capitalization thresholds, unusual accruals, estimates, contingencies, related-party treatment, suspected fraud, tax positions, and final close decisions require firm judgment and often more evidence than the bookkeeping system contains.

An AI bookkeeper can assemble facts, identify alternatives, and prepare a draft. The accountable professional should decide and approve the treatment.