AI Bookkeeping: What It Replaces and What It Doesn't

AI bookkeeping is the most over-marketed and under-delivered category in accounting software right now. Here is the honest cut.

What AI replaces in bookkeeping

• Receipt and invoice data extraction — solved
• Bank rule suggestions and transaction matching — solved
• Standard expense categorisation — mostly solved
• AP coding for repeat suppliers — solved
• Reconciliation prep — solved

What AI doesn't replace

• The judgment call on whether something is capex or opex
• Spotting that the client's "subscription" is actually a personal Netflix charge
• Knowing when to ask the client a question vs guess
• Year-end adjustments that require business context
• Anything where a defensible audit trail matters more than speed

What firms should install first

For a firm running Xero or MYOB across a client book, the highest-leverage AI bookkeeping install is an AP agent that:

1. Reads invoices from a shared inbox or SharePoint folder
2. Classifies them with awareness of the client's chart of accounts
3. Pushes coded transactions into Xero/MYOB as draft entries
4. Flags low-confidence calls for a human bookkeeper to review

That single install removes ~80% of the manual work the bookkeeper used to do, and leaves the bookkeeper doing the 20% that needs a brain.

The trap

Avoid any AI bookkeeping product that claims to do "the whole job end-to-end with no human in the loop." It will work for 30 transactions and then quietly mis-code the 31st in a way you only notice at year-end. The right architecture is always: agent does the work, human reviews exceptions.