In almost every company we meet, bank reconciliation works the same way. At month end someone downloads the statements, opens the accounting system and matches transactions to invoices by hand. Transfers that pay three invoices at once, bank fees, payments with the wrong reference, customers paying under a name that isn't the one on file. It takes days, and for those days the books are behind.
It's one of the first processes we automate, because it has three perfect traits: it's repetitive, it follows clear rules and every mistake shows up straight away.
What the AI agent does, step by step
At night, when nobody is using the systems, the agent reads the day's bank transactions through a read-only open banking service. It can't move money. It sees the transactions and nothing else.
- It reads every transaction: amount, date, reference, payer.
- It looks for matching open invoices in the accounting system, even when one transfer pays several of them or arrives slightly short because of a fee.
- It recognises the customer even when they pay under a different name, because it learns from the matches already made.
- It prepares the match and queues it for approval, with the reason written in plain words.
In the morning, whoever runs the books finds the list ready. They approve the safe matches in one click and only look closely at the exceptions.
What stays with your team
The decisions. A partial payment to close with a write-off, a customer who paid the wrong invoice, a transaction nobody recognises. The agent doesn't finalise anything until a person approves it, and every exception comes with an explanation: why it couldn't match it and what it suggests.
Every time someone corrects it, the rule is saved. After a few weeks fewer cases end up as exceptions, because the agent has learned how that specific company works.
What if the accounting system has no API?
It's the most common situation for small and mid-sized companies, and it's ours too: the cloud accounting system we use has no API. In that case a robot on a dedicated server uses the software through its interface, like a person would, at times when it's free. It records the approved match and logs every action. We cover this in detail in this article.
How much time you get back
It depends on volume. With a few hundred transactions a month, the work goes from days at month end to 15-20 minutes of approvals a day. The bigger change isn't the time, though. It's that the books are up to date every morning instead of once a month. You see unpaid invoices right away, chase them earlier and keep cash under control.
Where to start
You need three things: read-only access to your accounts through open banking, a dedicated user on the accounting system and one week in which your team records a screen video of how they reconcile today. From there we build a fixed-price pilot on one account or one company in the group, measure the result and decide whether to extend it.