AI-assisted accounting: practical examples

A lot of accounting software is shipping AI features. Most of the marketing claims are unfalsifiable — "AI-powered insights," "intelligent automation," "machine learning at the heart of your books." None of it tells you what the AI actually does or where it fails.

Here are six concrete examples of where Ledgable's AI helps in day-to-day Australian SMB accounting, with the actual confidence levels we observe and the cases where it gets it wrong.

1. Bank transaction categorisation

What it does: When a transaction comes in from your bank feed, Ledge AI reads the description, amount, and recent context, then suggests an account from your chart of accounts plus a GST treatment. Above 0.95 confidence it's a quiet suggestion; below 0.95 it's flagged for review.

Where it works well: Recurring vendors with consistent patterns. "OFFICEWORKS 1234" → 700 Office Expenses → GST. The model has seen this pattern thousands of times across thousands of workspaces.

Where it fails: Generic descriptions ("PAYMENT - VISA"), one-off vendors with no historical context, and any transaction that should be split across multiple accounts. The model returns lower confidence and surfaces it for review — which is the right behaviour, but it means roughly 15% of transactions on a busy month still need human input.

Caveat: The model is biased toward the chart of accounts you have. If your CoA is missing an account that would be the correct classification (e.g. you have no "Subscription" account but the transaction is a software subscription), the AI will pick the closest available — usually "Office Expenses" or "Other Expenses." Spend 30 minutes setting up your CoA properly before relying on the AI.

2. GST treatment by vendor

What it does: The AI also picks GST/FRE/INP treatment based on the vendor type. It knows that bank fees are FRE, that most goods purchases are GST, and that wages are out of scope.

Where it works: ~90% accuracy on standard AU SMB transaction types.

Where it fails: Capital purchases that should split GST claims across multiple periods (e.g. depreciable assets), and any transaction that requires a partial GST claim (e.g. a vehicle used 60% for business). The AI picks the most-likely GST treatment for the full amount; the apportionment is your job.

3. AI-suggested journal-entry corrections

What it does: When an accountant overrides a categorisation, the model logs that as a correction. The next time the same vendor or pattern appears, the override has more weight than the original suggestion. After two or three overrides, the model effectively learns the rule for that workspace.

Where it works: Clear, repeating patterns. "Every payment to Vendor X in workspace Y is a Cost of Goods Sold."

Where it fails: Patterns that depend on context the AI doesn't have. "Payments to Vendor X are COGS unless they're for the new product line, in which case they're R&D." The AI can't see the second clause. This is what Money Rules are for.

4. Cashflow forecast smoothing

What it does: The cash flow forecast uses recurring transaction patterns (rent, payroll, subscription revenue, regular supplier invoices) to project forward 30/60/90 days. The AI distinguishes one-off blips from genuine recurring items.

Where it works: Mature, stable businesses with predictable cash cycles.

Where it fails: Seasonal businesses, businesses going through structural change (new product, lost a major customer), and businesses with very lumpy revenue. The forecast over-smooths in those cases. We surface a confidence interval rather than a single number, and we recommend treating the forecast as a planning tool rather than a hard prediction.

5. BAS prep anomaly detection

What it does: As you approach a BAS lodgement, the AI scans the period for anomalies — GST coding inconsistent with prior periods, transactions over a threshold without a matching invoice, unusual vendor patterns. It surfaces a "review before lodgement" list.

Where it works: Catching the kind of mistakes that would normally surface at the partner-review stage — typically saves 20-40 minutes per BAS.

Where it fails: It's a checker, not a substitute for review. It catches anomalies; it doesn't catch correct-looking entries with the wrong amount.

6. Conversational interface (Ledge AI)

What it does: A chat interface that can answer questions about your books — "what was my biggest expense category last month," "show me unpaid invoices over 60 days," "which vendor charged the most GST in Q3." Read-only. Cannot modify the books.

Where it works: Ad-hoc questions that would otherwise require running a custom report. Particularly useful for business owners who don't know how to navigate the reports menu.

Where it fails: Questions that require multi-step reasoning across data the AI doesn't have permissions to access (e.g. "compare my margins to industry benchmarks" — we don't have benchmark data). It says so when this happens, rather than hallucinating an answer.

What we don't do

A few things we explicitly avoid:

• Auto-posting. No transaction is recorded to the ledger by AI alone. Every classification requires an explicit human click. (This is a locked product decision; see Q4 in our spec docs.)

• AI-generated invoices to customers. We don't auto-draft invoices. The AI can pre-fill an invoice from a quote or a recurring template, but humans send.

• AI tax advice. The AI can flag patterns but does not give tax advice. That's a regulated activity and a job for your accountant.

Try it

Every Ledgable plan includes Ledge AI. The Starter plan caps usage at $10/month of AI compute (which is more than enough for a sole trader), Professional and Firm scale up.

Try it free at ledgable.co.