Reviewing classification batches

When you have many uncategorised transactions, the Batch Review screen lets you process them in one view rather than one at a time on the Transactions list.

Opening Batch Review

From Transactions → Batch review, or Reports → Attention queue → Uncategorised transactions.

The screen lists every transaction needing categorisation in a single table:

DateDescriptionAmountAI SuggestionConfidenceAction

What each row shows

AI Suggestion — the primary account code + name suggested by Ledge AI.

Confidence — colour-coded badge: green (≥0.95), amber (0.80–0.94), red (<0.80).

Action — three options: Accept the suggestion, Override (pick a different account from a dropdown), or Skip (leave for later).

Accept-vs-override behaviour

Accept — records the suggestion to the journal entry. Logs an entry to classification_feedback with accountantOverrode: false.

Override — opens a dropdown of all active CoA accounts. Pick the correct one. The transaction is posted with that account; logs to classification_feedback with accountantOverrode: true and the original suggestion alongside the chosen account, so the AI learns the correction for next time.

Skip — leaves the transaction in the uncategorised queue. No record written.

What the screen does NOT do

No "Accept All" button. This is intentional. Locked product policy: every classification requires an explicit per-row click. No bulk-accept under any circumstance.

No auto-post on high confidence. Even at 1.0 confidence, you must click Accept.

Performance note

The Batch Review screen fetches an AI suggestion for every visible row when the screen loads. For workspaces with hundreds of unclassified rows this can take 10-30 seconds and use meaningful AI budget. Tips:

Filter to a single date range first (current month, last week) to reduce the visible set.

If the budget cap fires mid-load, partially-loaded rows still work; affected rows show "budget cap reached" — accept manually.

Run Batch Review at a lower-traffic time so caching from sibling screens hasn't already filled.

Recommended workflow

Connect bank feed (one-off).

Let the bank feed pull a week of transactions.

Open Batch Review.

Accept the high-confidence (green) rows quickly.

Spend more time on amber and red rows — these are the ones the AI is genuinely unsure about.

Set Money Rules for any pattern you find yourself overriding repeatedly (e.g. "every payment to Telstra is Phone & Internet").

After three or four weekly batches, the AI learns your patterns and the proportion of green-confidence rows climbs.