Help Center / AI Features

AI-Assisted Reconciliation

What is this page about: This page explains how bank reconciliation works in Light: what the matching engine does on its own, where AI supports it, and how to set up rules, convert transactions, and clear exceptions.

On this page

  • Overview
  • How matching works
  • AI transaction metadata parsing
  • Creating rules with natural language
  • Convert to rule
  • Manual matching
  • Clearing open invoices
  • Exception handling
  • Best practices

Overview

Bank reconciliation matches transactions from your bank against entries in your accounting system. Light's matching engine handles this with a fixed set of system rules plus any custom rules you add. AI supports the engine in three specific ways:

  1. Parses bank transaction metadata. It extracts document identifiers, fees, original amounts, and end-to-end payment IDs from raw bank references, so the matching rules get clean data to work with.
  2. Creates custom rules from natural language. You describe a rule in plain language and AI converts it into structured matching conditions.
  3. Suggests rule descriptions. When you convert an unmatched transaction into a rule, AI suggests the description based on the transaction's details.

Bank reconciliation has two top-level tabs: Accounts, where you reconcile transactions, and Rules, where you manage matching rules. Within an account's reconcile view, there are four tabs: Unmatched, Matched, Excluded, and All transactions.

Matches the engine finds are applied directly. Matched transactions move straight to the Matched tab. There's no suggestion queue, no confidence score, and no accept or reject step. If a rule produces an incorrect match, undo it and match manually.

How matching works

Click Auto reconcile, or let it run after a bank import completes, and Light checks each unmatched bank transaction against your active rules, in order, until one produces a match. The built-in system rules are:

  • Match by document number — matches on invoice or document numbers found in the bank reference
  • Match by amount and description — matches where both amount and description align with a ledger entry
  • Register payment for open invoice receivables — matches incoming transactions to open customer invoices and registers the payment
  • Register payment for open invoice payables — matches outgoing transactions to open vendor bills and registers the payment
  • Create Journal Entry for card balance funding transactions — recognises transfers that fund a card balance account
  • Match by bank transaction end to end id — uses end-to-end payment references, common in SEPA and wire transfers
  • Match by amount and date — matches when exactly one ledger entry shares the same amount and date

These rules are deterministic. The same transaction and ledger data always produce the same result. The engine is conservative by design: amount-based rules only match when exactly one candidate entry qualifies. An ambiguous transaction stays unmatched for you to review rather than getting matched incorrectly.

See Automated bank reconciliation for the full reconciliation workflow.

AI transaction metadata parsing

Bank references are often messy. A single text field can contain an invoice number, a payment reference, a fee, and an original currency amount all at once. When transactions import, Light's AI parses each reference and extracts:

  • Identifiers — invoice numbers, document numbers, payment references
  • Fees — transaction fees embedded in the reference
  • Original amount — the amount in the original currency, for converted transactions
  • End-to-end ID — the payment's end-to-end reference

Light checks this extracted data with sanity checks when a match is applied. The matching rules then use the structured metadata. Match by document number, for example, looks up the extracted identifiers against your invoice and document numbers.

This is where AI adds value: it turns unstructured bank data into structured fields, while the matching decision stays rule-based and predictable.

Creating rules with natural language

Custom matching rules start as plain language, and Light's AI converts your description into structured conditions.

  1. Go to Accounting → Bank reconciliation and open the Rules tab (a top-level tab next to Accounts, not part of the reconcile view's tab strip)
  2. Click Create rule
  3. Enter the rule Name
  4. Under the When heading, fill in the Instructions field to describe when the rule should apply in plain language, for example "When the transaction reference contains 'RENT'"
  5. Select the Action the rule should perform when it matches
  6. Fill in the Details for the entry the rule creates: business partner, GL account, description, tax code, and so on
  7. Click Save & apply rule (this reads Validate & Save while you're reviewing an AI-parsed prompt)

Light parses your description into conditions on fields like amount, date, reference, and bank account. Review the parsed conditions before saving to confirm they match your intent, and refine the wording if they don't.

Once saved, the rule runs deterministically during auto reconcile. AI is only involved when you create or edit the rule, not each time it runs.

See Reconciliation automation rules for full detail on conditions and actions.

Convert to rule

When an unmatched transaction will recur, such as rent, payroll, or a subscription, turn it into a rule directly:

  1. On the Unmatched tab, select the bank transaction
  2. Click Convert to rule
  3. Light's AI suggests a rule description based on the transaction's details
  4. Review and adjust the suggestion, then click Save & apply rule

Similar future transactions are then handled automatically during auto reconcile.

Manual matching

For transactions the rules couldn't match:

  1. On the Unmatched tab, select one or more bank transactions in the left panel
  2. Select the matching ledger entry or entries in the right panel
  3. Check that the Difference indicator shows 0.00
  4. Click Match to confirm

Manual matches are recorded like any other match. Light doesn't learn from your manual matching decisions. To automate a recurring pattern, use Convert to rule or create a custom rule instead.

Clearing open invoices

The Register payment for open invoice rules handle bank transactions that pay outstanding invoices:

  1. During auto reconcile, the engine looks up open customer invoices or vendor bills using the identifiers parsed from the bank transaction
  2. When it finds a matching open invoice, Light registers the payment and reconciles it against the bank transaction in one step
  3. The invoice is marked as paid and the match appears on the Matched tab

Exception handling

Some transactions won't match. Here's what usually causes it and what to do:

  • Timing differences. The bank has cleared a transaction that isn't posted in accounting yet. Mark it for next month.
  • Bank fees. An unexpected bank charge. Create a journal entry to record it.
  • Currency differences. Multi-currency payments can carry rounding differences. Review and match manually.
  • Fraud or error. A transaction that doesn't belong. Investigate, and contact your bank if needed.

Best practices

  • Run Auto reconcile first. Let the rules engine handle the straightforward matches, then focus your time on the exceptions.
  • Use consistent invoice numbering. The document number rule works best when your invoice numbers appear in bank transaction references.
  • Review parsed rule conditions. Always check the conditions Light generates from your natural-language description before saving.
  • Reconcile on a schedule. Weekly or monthly reconciliation keeps the unmatched queue manageable.

Also searched as: AI bank reconciliation, automatic bank matching, bank rec AI, auto reconcile bank transactions, matching rules engine, unmatched transactions Light

Was this article helpful?

Book a demo