Documentation Index
Fetch the complete documentation index at: https://docs.dualentry.com/llms.txt
Use this file to discover all available pages before exploring further.
Bank Match AI
Bank Match AI is DualEntry’s ML-based engine for matching bank feed transactions to GL entries during reconciliation. It scores candidate pairs, auto-matches high-confidence results, and routes uncertain matches to a review queue so you only spend time on the items that need judgment.
How matching works
When you open a bank reconciliation, DualEntry runs rule-based matching first, then hands the remaining unmatched items to Bank Match AI. The AI evaluates every possible pairing of an unmatched bank transaction against an unmatched GL entry and assigns a confidence score from 0 to 100.
The model weighs multiple signals:
- Amount proximity. Exact or near-exact amount matches score highest. The model tolerates small differences (such as rounding on foreign-currency conversions) without penalizing the score heavily.
- Date proximity. Bank transactions and GL entries within a few days of each other score higher than pairs separated by weeks. The tolerance window is wider for items that historically post with a lag, such as credit card settlements.
- Description similarity. The model compares bank memo text against GL entry descriptions, vendor names, and reference numbers using text-similarity scoring. Abbreviations and truncated bank descriptions are normalized before comparison.
- Historical patterns. If you’ve previously matched similar pairs - same vendor, same amount range, same account - the model boosts confidence for the current pair.
Each signal contributes to the composite score. No single signal is sufficient on its own; a pair needs strength across multiple signals to reach the auto-match threshold.
Confidence scoring and thresholds
Bank Match AI uses a two-tier threshold system to separate high-confidence matches from items that need your review.
- Auto-match threshold (default: 85). Pairs scoring at or above this threshold are matched automatically and appear in the Matched tab of the reconciliation workspace with an AI badge.
- Review threshold (default: 50). Pairs scoring between 50 and 84 appear in the Review Queue - a dedicated section of the Unmatched tab where suggested matches are presented for your approval or rejection.
- Below review threshold. Pairs scoring below 50 are not surfaced as suggestions. They remain fully unmatched and require manual matching.
You adjust both thresholds under Settings → Cash Management → Bank Match AI. Raising the auto-match threshold reduces false positives but increases the number of items in the review queue. Lowering it does the opposite. Finding the right balance depends on your tolerance for manual review versus the risk of incorrect auto-matches - most organizations leave the defaults in place until they have a few months of reconciliation data to evaluate.
Threshold changes apply to future reconciliations only. Items already auto-matched in a completed reconciliation are not re-evaluated.
Learning from corrections
Bank Match AI improves over time by learning from your actions during reconciliation. Three feedback signals drive adaptation:
- Confirmed matches. When you complete a reconciliation, every matched pair - whether auto-matched or manually matched - becomes positive training data.
- Rejected suggestions. When you dismiss a suggested match from the review queue, the model records that pair as a negative signal.
- Manual overrides. When you unmatch an auto-matched pair and match the bank transaction to a different GL entry, the model treats the original pair as negative and the replacement pair as positive.
The model retrains periodically (not in real time). After retraining, you may notice that match suggestions shift - vendors that were previously sent to review start auto-matching, or patterns you’ve consistently rejected stop appearing as suggestions. The retraining cadence is automatic and does not require any action on your part. You can view the date of the most recent retrain under Settings → Cash Management → Bank Match AI → Performance.
Relationship to manual rules
Bank Match AI and rule-based matching are complementary. Rules run first and handle deterministic cases - for example, “match any bank transaction with description containing ‘PAYROLL’ to the payroll clearing account.” The AI runs second on whatever the rules didn’t catch.
If a rule and the AI would both match the same pair, the rule wins. This means you can use rules to enforce hard business logic (such as always matching a specific recurring charge to a specific GL entry) while letting the AI handle the long tail of variable transactions. Rules are also useful for edge cases where the AI consistently gets it wrong - creating a rule for that pattern ensures it’s handled correctly every time and frees the AI to focus elsewhere.
You configure rules under Settings → Cash Management → Matching Rules. For the reconciliation workflow that brings rules and AI together, see How to Reconcile Bank Accounts.
Accuracy over time
New DualEntry organizations start with a general-purpose model trained on anonymized, cross-tenant patterns. Accuracy at this stage is typically moderate - the model catches obvious matches but sends more items to review than it will after a few months of feedback.
After three to four completed reconciliation cycles, the model has enough organization-specific data to meaningfully improve. Most organizations see auto-match rates climb from roughly 60% to 85–90% within six months, though the exact trajectory depends on transaction volume and variety. High-volume organizations with repetitive vendor patterns see faster improvement, while organizations with highly variable transactions may take longer.
You can view accuracy metrics under Settings → Cash Management → Bank Match AI → Performance. The dashboard shows auto-match rate, review-queue acceptance rate, and override rate per reconciliation period. Use these metrics to decide whether to adjust thresholds or add manual rules for specific patterns.
Bank Match AI works best when reconciliations are completed regularly. If you skip several months and then reconcile a large backlog, the model has less context for date-proximity scoring and accuracy may dip temporarily.
