Automatic transaction matching is not the same thing as a completed bank reconciliation.
Matching pairs bank lines with book entries. A reconciliation establishes that the account balance is explained, outstanding items are correctly identified, adjustments are appropriate, and a reviewer can sign off that the result is supportable. High auto-match rates measure the first. They do not automatically deliver the second.
What current systems handle well
On clean, recurring activity—known vendors, consistent amounts, reliable bank feeds, and useful history—modern platforms match a large share of lines without intervention. Incumbent tools such as Xero’s JAX auto-reconcile only high-confidence items and leave the rest for review; secondary reviews have cited auto-match rates above 80% under those conditions, and Xero has reported high accuracy on the subset it auto-reconciles. Dedicated reconciliation platforms and AI-native ledgers commonly report auto-match rates in the 90s on favorable data. Practitioner testing on real client files shows lower hit rates when the books are complex.
These figures are auto-match rates, not reconciliation accuracy rates. They describe how many transactions the system pairs. They do not describe whether the account is fully and correctly reconciled.
Continuous matching improves the picture further. When feeds run throughout the period, most routine items are already matched before month-end. The backlog shrinks. The residual work does not disappear.
Where human investigation remains
| Stage | What systems handle well | Where human investigation or review remains |
|---|---|---|
| Routine matching | High-confidence 1:1 matches on clean, recurring items | Medium-confidence suggestions; threshold policy |
| Transfers | Recognition of some linked-account transfers | Misclassification as income or expense; related-entity moves |
| Timing differences | Flagging expected clearing patterns | Deposits in transit, outstanding items, clearing windows |
| Duplicates & missing items | Detection and flagging | Confirmation, correction, feed-completeness checks |
| Partial / batch payments | Improving fuzzy and many-to-many suggestions | Correct allocation and approval of proposed splits |
| Exception investigation | Context packaging and exception classification | Ownership of resolution and materiality |
| Adjustments | Low-risk recurring fees or interest in some tools | Appropriateness of any auto-created entry |
| Completion & sign-off | Match rate, outstanding list, status visibility | Whether the reconciliation is supportable |
The residual items are not a random remainder. Transfers, timing differences, partial and batch payments, and missing support concentrate judgment and error risk. A system that leaves an ambiguous item unmatched is often safer than one that forces a higher match rate by accepting a questionable pair. False-positive matches—correct amount, wrong counterpart—are more dangerous than unmatched lines because they can post cleanly and hide the problem until review or audit.
Continuous and “autonomous” reconciliation
Product language around continuous, AI, or autonomous reconciliation usually means continuous or high-volume matching plus an exception queue. Digits and similar platforms describe continuous matching that surfaces anomalies for review. Xero’s design is explicit: high confidence only; everything else stays manual. Enterprise and mid-market reconciliation tools follow the same pattern at higher match rates on cleaner data.
In other words, the dominant operating model is management by exception. The system clears what it can clear confidently. A person still owns the residuals and the sign-off.
This is the same distinction GraphAccount has drawn elsewhere. High transaction automation does not equal completion of the bookkeeping function. Automating work inside the month-end close does not automate the decision that the books are ready to close. Mechanical statement generation does not establish that the statements are ready to issue. Matching bank lines does not, by itself, complete a reconciliation.
The control question that matters
For a 5–30 person firm the useful question is not “What percentage can AI match?” Match rates will vary by client, feed quality, and history. The operational question is whether the firm’s exception process is strong enough that unmatched and ambiguous items are investigated and resolved before anyone treats the account as reconciled.
If the process treats a high auto-match rate as evidence that the account is finished, risk increases. If the process treats auto-match as a first pass and requires explicit resolution of residuals, timing items, transfers, and adjustments before sign-off, AI becomes a capacity tool rather than a source of false confidence.
AI can already match a large majority of clean bank transactions and can run that matching continuously. On a clean account, little intervention may be needed. Across typical client accounts, high auto-match rates do not demonstrate reliable end-to-end reconciliation without human exception handling and review. The reconciliation is complete when the residuals have been worked and a reviewer can stand behind the result—not when the match rate looks high.
