How AI suggestions work

When a transaction comes in from your bank feed, Ledgable's AI (Ledge AI) reads the description, amount, and recent context, then suggests an account from your chart of accounts plus a GST treatment.

The flow

Cache lookup. Identical descriptions in the same workspace are cached for 1 hour. If we've seen this exact description recently, the cached suggestion returns immediately at zero AI cost.

Consent check. If your workspace has opted out of AI classification (Settings → AI → Classification), no AI call is made and you'll see the "AI disabled" indicator next to each row.

Budget check. Each workspace has a monthly AI spend cap (default $10/month, configurable per workspace). If the cap is reached, the suggestion is skipped and a "budget cap reached" indicator appears.

LLM call. A request goes to Anthropic's Claude with your chart of accounts, recent corrections, and the transaction. The model returns a primary suggestion, two alternatives, and a confidence score (0.0 – 1.0).

Validation. The returned account code is checked against your active CoA. If the model hallucinates a code that doesn't exist in your workspace, the suggestion is replaced with a "needs review" placeholder and we log the failure.

The confidence score

ConfidenceWhat it means ------ 0.95 – 1.00High. No "needs review" badge shown. Accountant still must click Accept; no auto-posting. 0.80 – 0.94Medium. "Needs review" badge shown. Model is reasonably sure but flags it. 0.50 – 0.79Low. Suggestion shown but accountant should look closely. 0.00 – 0.49Very low. Default to Manual classification.

What we do NOT do

Auto-post. No transaction is recorded based on AI alone. Every classification requires an explicit human Accept click. This is locked product policy.

Auto-accept high-confidence suggestions. Even at confidence 1.00, you must click Accept. The high-confidence threshold only suppresses the "needs review" badge.

Send your transaction descriptions to third parties beyond Anthropic. Anthropic's enterprise terms apply. We don't train on your data.

Improving accuracy

The AI improves over time as your accountant overrides incorrect suggestions. Each override is logged in the classification_feedback table and surfaced to the model in the "recent corrections" section of the prompt for future requests.

For patterns that should always go to a specific account (e.g. "every payment to Vendor X is COGS"), use Money Rules instead of relying on AI corrections — Money Rules are deterministic and override the AI suggestion.

Cost

Each AI suggestion costs approximately $0.001 – $0.005 depending on chart of accounts size. Cached suggestions cost nothing. The default $10/month cap covers ~2,000-10,000 unique transactions per workspace per month.