If you let an AI analyze your books, there's a question almost nobody asks before trusting what it says: is the data you're feeding it actually complete? We assumed it was, until we checked one bank account transaction by transaction and found it was missing 492 out of 1,298 real ones. A year and a half of history with a hole in it, and nothing on screen gave it away.
We're telling it exactly as it happened, because the failure isn't exotic: it can hit any business that centralizes several bank accounts, with or without AI in the mix.
The problem: a "complete" report that wasn't
To bring transactions from each account into our centralized accounting, we used the "transaction history" report that most banks and aggregators offer. It's the one that shows up in every integration tutorial, and it works: you download, you import, it adds up. Or so it seemed.
That report is not a bank statement. It's a partial view built for quick lookups, and it can leave transactions out without warning that anything is missing. No error, no alert, it just doesn't include them. If your process trusts that "no error" means "complete," it inherits the flaw without ever knowing.
We uncovered it by comparing, account by account, the sum of what we'd imported against the real balance the bank declares in its official statement. In one personal Wise account, the gap was massive:
| Check | Result |
|---|---|
| Real transactions per the official statement | 1,298 |
| Transactions we had imported | 806 |
| Missing transactions | 492 |
| "Phantom" balances that appeared through the gap | 457 AUD, 1.65M Indonesian rupiah, 113 USD |
And it wasn't the only account affected: the same review found another 32 missing transactions spread across three more accounts (Wise business, Zen.com, and Revolut Business). In total, five accounts with incomplete data that had looked correct for a long time.
Why this is more dangerous with AI than without it
Without AI in the picture, incomplete data is already a problem: the books don't add up, you decide with partial information. But at least there's a human looking at the screen who, sooner or later, might notice something's off.
With an AI reading that same data, the risk changes shape. A model can't tell "this is everything there is" apart from "this is everything I was given." If you ask it to analyze cash flow on that account, flag anomalies, or summarize the financial situation, it will answer with full confidence about 62% of the real transactions, with no signal that the rest is missing. The answer sounds just as certain whether the data is complete or not. That's the exact same pattern we saw with a dashboard that calculated 1.6 million euros in net worth when the real figure was 1,150 euros: the failure was in the input data, not in the reasoning on top of it.
That's why at AutoBoost we always keep the same order: first the data gets sorted and verified, then AI gets applied on top. Never the other way around.
The decision rule: don't trust a report that doesn't declare a balance
The check that prevents this failure is simple and you can apply it today, with no new tooling required:
- If the report doesn't declare a balance, it's a partial view, not a statement. An official bank statement always carries an opening and a closing balance; a "transaction history" from an API or a dashboard usually doesn't.
- The sum of what you imported has to match the real balance exactly. Not "roughly": to the cent. Any difference, however small, is a sign that something got left out.
- Repeat the check account by account. A failure like this is almost never uniform: it can hit one account and not the others, exactly as it did for us.
- Automate the check before you automate the analysis. If you're going to let an AI reason over your accounts, make validating that the input data is complete the first automated task, not the last.
Checklist before you let AI analyze your accounts
- Does the report you use to import declare an opening and closing balance?
- Does the sum of imported transactions match the bank's real balance, to the cent?
- Have you verified this on every account separately, not just the main one?
- Is there a process that repeats this check every time you re-import, or was it done once and left alone?
- Do you know what happens if a transaction arrives duplicated or gets lost during sync?
The fix, in our case, was to stop trusting the transaction history report and verify each account against its official statement instead, re-importing until the sum matched exactly what the bank reported. It didn't take any new tooling, just dropping the assumption that "no error" means "complete."
It's the same principle we apply in the data and AI platform that unifies data across several locations for a pharmacy group: the AI only reasons over data after it's been verified and balanced to the cent, never before. That's what lets a number the AI gives you actually be defensible.
If you're thinking about putting AI on your accounts, your CRM, or any system fed by outside data, start here: verify that what you're about to give it is complete. If you'd like us to look at it together, get in touch.
