AI Can Sort Your Transactions. It Can't Tell You Whose Money It Is.

So here's a refresher on the fundamentals, and an honest look at where automation genuinely helps — and where it doesn't, no matter what the software's marketing page promises.

The refresher: still just two buckets

Since 2018, under FASB's ASU 2016-14, nonprofit and church financial statements report net assets in exactly two categories: net assets without donor restrictions, and net assets with donor restrictions. That replaced the old three-tier system — unrestricted, temporarily restricted, permanently restricted — that a lot of longtime church treasurers still remember and sometimes still use out of habit.

A restriction comes from the donor, not from your board. If someone gives $10,000 and says it's for the youth mission trip, or a grant agreement limits the funds to a specific program, that's a donor restriction — legally binding, and it stays with donor restrictions on your books until the purpose is fulfilled or the time period passes. If your board sets aside $20,000 toward a future roof repair, that's a board designation, not a donor restriction — it still counts as unrestricted, because the board can reverse its own decision. A donor can't.

When a restriction is satisfied — the trip happens, the grant period ends, the building fund gets spent on the building — you record a release, moving that dollar amount from "with donor restrictions" to "without donor restrictions" on the statement of activities. For any nonprofit or chapter organization that files a Form 990, this same split shows up directly on Part X of the return. Getting the classification right during the year isn't just internal housekeeping — it's what ends up on a public tax filing.

Where the mistakes actually live

In my experience, the errors that cause real problems are rarely dramatic. They're a memorial gift that gets deposited into general checking with no fund tag attached. They're a mission trip fund that never gets released after the trip happens, so it sits on the books looking restricted for another eighteen months. They're a building fund and a benevolence fund sharing one QuickBooks class because nobody set up a second one when the second fund started.

None of that is fraud. It's the ordinary erosion that happens when tracking restricted and unrestricted dollars separately isn't clearly anyone's job — and it's exactly the kind of erosion a board tends to discover at the worst possible time: audit season, or the moment a major donor asks where their gift actually went.

What AI is genuinely good at right now

I want to be fair here, because the tools have gotten better. Automated bank feed matching now reconciles routine transactions in seconds instead of the hour or two it used to take a bookkeeper. OCR receipt and invoice capture pulls line-item data straight into the ledger, cutting out manual entry. Donation-platform sync between giving software and the general ledger means a gift recorded at the door doesn't have to be typed in twice. For larger multi-site ministries, some platforms now offer real consolidated dashboards across campuses — reporting that used to take a controller days to assemble by hand.

Those are real gains, and I use tools that do some of this in nearly every engagement. Hours that used to go into data entry can go into actually reading what the numbers say instead.

What AI still can't do

Here's the gap the marketing tends to skip over. Classifying a gift as restricted isn't a data-entry task — it's a judgment call. Software can learn to tag "Youth Mission Trip" the same way every time once a person tells it what that label means. It can't decide, the first time, whether a check with "for missions" written in the memo line creates a binding purpose restriction, a polite suggestion, or an unconditional gift. It can't read a grant agreement and apply the FASB test for whether a barrier exists that makes the funding conditional. Someone has to make that call — and then build the chart of accounts so the system applies it consistently afterward.

This isn't unique to church bookkeeping. Even in the broader accounting profession, standard-setters are still working out how to assign responsibility when an AI system misclassifies a transaction or drafts a workpaper that turns out to be wrong — there's currently no authoritative standard defining what adequate human oversight actually requires. AI systems can also produce confident, wrong answers: citing the wrong part of a regulation or misreading a condition in a document, not because the system is being deceptive, but because it's predicting what looks plausible rather than verifying what's true.

Before you get excited about the cost savings

Every AI accounting product's pitch leads with hours saved, and that part is often true. But hours saved on data entry doesn't mean hours saved on review — if anything, a tool importing and categorizing thousands of transactions overnight needs a more disciplined review process than the old one-line-at-a-time approach, not a lighter one. A misclassification that used to happen occasionally now happens at volume, and it happens consistently, because the tool will apply whatever rule it was given every single time, correct or not.

That risk is sharper for churches specifically. A general-purpose bookkeeping AI doesn't know that your congregation's designation letter restricts a fund to camp scholarships only, through a specific year, unless a person tells it — and keeps telling it correctly every time a new fund shows up. Automated categorization is only as good as the fund structure a human built underneath it. Get that structure wrong, and the tool won't second-guess you. It will just misfile everything, confidently and consistently.

What still falls to a person, no matter which tools you use

Where I land

I'm not against these tools. I use software that automates reconciliation and bank matching in most of my engagements, and it genuinely speeds up the part of bookkeeping that was never the interesting part anyway. But I don't let a dashboard decide whether a gift is restricted, and I don't let an automated report substitute for actually reading a client's financials before they go to their board. That's not a faith-based conviction, though it lines up with one I hold anyway — it's just an accounting standard. The numbers your board reads are only as good as the judgment behind them, and judgment isn't something you can subscribe to.

If you're evaluating a new AI bookkeeping tool for your church or nonprofit, that's a good thing to explore — just go in clear-eyed about what it will and won't take off your plate.