AI in accounting, what's actually working

We have been running large-language-model tooling inside our practice for two years. Here is the honest version of what helps and what doesn't.

Where AI earns its keep

• Document extraction: invoices, bank statements, contracts, high accuracy with a human review step.
• Drafting first-pass commentary on management variances.
• Searching across years of policy memos and prior advice.
• Reconciling unstructured data into a clean ledger feed.

Where it still falls short

• Anything requiring legal or tax-position judgement.
• Sign-off, controls require an accountable human.
• Novel transactions where the right answer is not in the training data.