Reconciliation software rarely gets a headline outside the finance department. AutoRek’s acquisition of Grath in early August is worth one because it points to where the category is going: from configurable matching engines towards systems that claim to investigate exceptions as well.
AutoRek provides financial control and reconciliation software. Grath, founded in London in 2019 and operating in markets including Australia, has focused on AI-driven reconciliation and compliance work. The companies say the combination will bring automation to more of the controls process.
That sounds sensible. It also deserves a little scepticism, because reconciliation has always looked easier in a demo than it does on the morning after a messy settlement.
Matching is not the whole job
The clean version of reconciliation has one transaction in the payment platform, one entry at the bank and one invoice in the ledger. Amount, currency and reference all agree. Software has handled those rows for years.
The real queue contains partial settlements, netted fees, chargebacks, reversals, foreign-exchange differences, duplicated references, late files and transactions split across entities. One provider reports in UTC, another uses local settlement date, and a third quietly changed its file format on Tuesday.
These are not always pattern-recognition problems. Sometimes the data is incomplete. Sometimes two systems disagree about the economic event. Sometimes a human has to decide which source is authoritative and document why.
Where AI could genuinely help
An AI-assisted system can be useful before it is trusted to post anything. It can group exceptions that share a cause, suggest likely matches, read remittance text, summarise prior investigations and draft an explanation for review. It can notice that every unmatched item began after a particular gateway release or that a missing settlement file normally arrives two hours late.
That is meaningful work. It moves an analyst from opening twenty systems to testing a plausible explanation.
The danger is confusing a confident suggestion with a financial control. Reconciliation exists to prove that records agree—or to explain why they do not. Any AI recommendation needs evidence, approval rules and an audit trail showing what data it used. “The model matched it” will not satisfy an auditor, a regulator or a merchant missing $200,000.
Buyers should ask practical questions. Can the system reproduce a decision six months later? How does it handle a changed model? What happens when confidence is low? Can a reviewer see the original records? Does the product measure false matches separately from unmatched items?
The Payment Nerd view
Reconciliation is a good place for applied AI precisely because it is full of repetitive investigation and ugly data. It is also a bad place for magical thinking because the output sits inside financial statements, customer balances and regulatory controls.
AutoRek buying Grath is another sign that the market wants fewer static rules and more help resolving the breaks those rules produce. The winners will not be the platforms that make the exception queue disappear on a slide. They will be the ones that make each exception faster to understand, safer to resolve and easier to audit.
A 99% match rate sounds excellent until the remaining 1% contains all the money.