The best AI for accounting isn't one tool

Someone spent a few months testing AI tools in their accounting workflow and summarized the conclusion in one sentence:

The best AI for accounting isn’t one tool — it’s how you use them together.

That matches what I’ve seen.

A cheap, fast model at the front filters incoming emails — newsletters out, invoices in. A more capable model extracts the structured fields. A code node applies the GL mapping rules. A threshold check decides what goes to approval and what goes straight through. Weekly, an anomaly agent compares recent transactions against historical patterns and flags anything unusual.

Each part of that chain is simple. The value is in having all of them connected.

What doesn’t work: one model trying to do everything. Ask a single AI to read an invoice, code it to the right account, check it for duplicates, and decide whether it needs approval — and it can do all of those things, but not reliably enough at scale and not cheaply enough to justify.

What also doesn’t work: expecting AI to handle the parts that require judgment and accountability. BAS submissions, tax filings, payment approvals — those still need a human. Not because AI can’t produce an output, but because someone has to own the result.

The practical path: pick one repetitive task, build something narrow that handles it reliably, and see what that actually saves. Then build the next piece.

The full pipeline emerges over time. It doesn’t arrive pre-assembled.