Bank Reconciliation With AI: From 8 Hours to 23 Minutes

The problem: eight hours nobody wanted

IN SHORT At this large fund, the check took a full day, once a cycle. My job was to change that.

The fund is a US real-estate private-equity firm with roughly $3B in assets under management. The engagement is confidential, so the fund stays unnamed.

I was the forward-deployed AI engineer, meaning I worked inside the fund’s own team. I built the reconciliation agent as part of a wider AI platform. An agent is an AI program that works through a task’s steps itself.

This is a first-party resultFirst-party result A figure reported by the people who did the work, here me, without claiming independent verification. See the glossary from this one engagement, and human review stayed part of the design. It doesn’t show how long each stage took, or guarantee that every future match will be right. Another team would need its own baseline.

IN SHORT Bank reconciliation is proving that your books and the bank tell the same story about your money.

You have two records, your ledgerLedger Your own record of every payment, fee and transfer you believe happened. People also call it the books. See the glossary and the bank statementBank statement The bank’s list of what actually went in and out of the account. See the glossary, and they should agree.

For a fund, that explaining step is where the control sits. It catches a duplicated payment, a fee nobody approved, or a wire that went somewhere it shouldn’t have.

ADAM

ADAM says: It’s like checking your card statement against your pile of receipts. Except your receipts fit in one drawer, and the fund’s evidence was spread over 13 sources and 13 years.

It’s like checking your card statement against your pile of receipts. Except your receipts fit in one drawer, and the fund’s evidence was spread over 13 sources and 13 years.

▼

IN PLAIN WORDS

At the end, you can say with confidence that the money did what the books claim.

?Coming up next: So why did this fund need a whole day for it?

IN SHORT The eight hours went on searching for evidence, not on matching lines.

The documents that explain an odd line sat in sources built up over more than a decade. Their formats didn’t agree, and whoever saved a file named it.

When I mapped it, the scale explained the day: about 2 TB and 1.4 million files, across 13 years and 13 sources. Explaining one odd line could mean opening five systems to find the one PDF behind it.

ITEM GET

Find where the hours go

Find out where a slow check’s hours go before you automate it. Speed alone doesn’t show useful work or a process for mistakes, and faster work helps only if your team can inspect the result.