Skip to main content
When transactions arrive in DualEntry through a bank feed, OCR upload, or bulk import, the system suggests a GL account and dimensions for each line. Suggestions draw from rules you define, patterns learned from your prior corrections, and the transaction description or memo. Nothing is applied automatically. Every suggestion waits for you to accept, edit, or exclude it, so no suggestion reaches the ledger without your say.

​
How the suggestion engine works

DualEntry works through an uncategorized bank transaction in this order:
  1. Rules - Rules you have written run first and handle deterministic cases. A matching rule builds a draft record and offers it as a suggestion without consulting the AI.
  2. Similar transactions - With no rule match, DualEntry looks for a transaction you previously matched that resembles this one, and infers the transaction type from it.
  3. AI detection - With no similar transaction to learn from, the transaction type is detected from the transaction itself. DualEntry then resolves the vendor, customer, or GL account the suggestion needs, depending on the type.
Suggestions to create a record carry a confidence score, shown as a badge beside the suggested record. It tells you how much corroboration the suggestion had. It does not decide anything on its own, and no score is high enough to apply a suggestion for you. This path applies to transactions arriving from a bank feed, which are the ones that reach the Bank Match page. Records you build elsewhere, such as a bill created from an uploaded document, get their account suggested from your own posting history for that vendor, company, and memo. The suggestion is offered on the record form as you fill it in rather than through a queue.

​
What you can and cannot configure

There is no categorization settings page, and there is no auto-apply toggle, no confidence threshold to tune, and no per-dimension suggestion setting. Those controls do not exist because suggestions are never applied on their own: each one sits on the Auto tab until you act on it. What you do control: Rules are the real lever. A rule runs before the AI and settles the cases you can state deterministically, which both improves accuracy and reduces what reaches the Auto tab. See Bank Match AI for how rules and the model interact.
Ask Copilot to categorize a batch of uncategorized bank transactions and it returns a preview of the records it would create rather than creating them. The request takes a minimum confidence, 0.85 by default and clamped to the range 0.70 to 0.95. It is a parameter on that one request rather than an organization setting, and it decides what enters the preview rather than what gets applied.

​
Reviewing suggestions

Suggestions the system generated for you appear on the Auto tab of the Bank Match page, under Close Management → Close Workflows → Bank Match. The page also carries To Match, For Approval, Matched, and Excluded views, so the Auto tab is where you check the work rather than where every uncategorized transaction sits. Review a suggestion before accepting it, and correct the ones that are wrong rather than approving in bulk without looking. A transaction you match by hand becomes the reference point for the next transaction that resembles it, so a correction made once pays off on every similar transaction after it. Sort or filter to put the least certain suggestions in front of you first. Working in that order means the transactions that most need a human decision get one while you are still paying attention, and the routine remainder can be confirmed quickly at the end.
Overrides train the model. When you correct a suggestion, DualEntry records the correction and weighs it heavily for future transactions with similar attributes. Over time, the engine becomes more accurate for your specific data patterns.

​
Working with bank feed rules and templates

Automated categorization runs alongside the matching rules you write on the Bank Match page. If a rule matches an incoming transaction, the rule settles it and the AI is not consulted. For transactions no rule matches, the AI suggestion path described above fills the gap. You therefore do not have to choose between rules and AI suggestions. Rules carry the cases you can state as a condition, and the AI covers the long tail. Because suggestions are derived from transactions you have already matched rather than from a trained model, adding a rule for a pattern does not discard anything: the history the suggestions draw on stays exactly as it was. If you later delete the rule, the AI path picks the pattern up again from that same history. Rules are worth writing for anything you would otherwise correct repeatedly. A pattern the AI gets wrong twice is a pattern worth stating as a rule, which settles it permanently and keeps it off the Auto tab entirely.
Last modified on August 27, 2026