AI and the trial balance — finding what moved and why

Every accountant has scanned a trial balance looking for something wrong. It is a genuine skill, and it is also an unreliable one, because it depends on remembering what each of two hundred accounts normally does.

A trial balance balancing tells you the double entry held. It says nothing about whether the numbers are right. Every miscoded transaction, every accrual that stopped being true, every duplicate invoice — all balance perfectly.

The useful question is not "does it balance" but "what is behaving unusually, and why?"

What anomaly detection actually looks for

Movement against the account's own history. Not a fixed threshold, because a 40% swing in a volatile account is normal and a 6% swing in a stable one is not. The comparison that matters is each account against its own pattern.

Accounts that should move and did not. More easily missed than the reverse. Depreciation that did not post. A recurring accrual that failed. An account that has been identical for three months when it should vary — silence is a signal, and scanning does not catch it because nothing looks wrong.

Composition changes. The total is normal; what it is made of is not. Freight costs unchanged in total but now concentrated in one supplier. Revenue flat but shifted to a lower-margin product line.

Unusual entries. Round numbers, entries posted at odd times, manual journals in accounts that normally receive only automated postings, entries reversed and re-posted. None of these are wrong by themselves, and collectively they are where error and worse tend to appear.

Balances that will not decompose. A control account balance that no longer breaks into identifiable items. This is the strongest single indicator that something has gone wrong upstream.

Why this beats scanning

Not because it is cleverer, but because it is consistent and has perfect memory.

A person scanning a trial balance on day seven of a close, under time pressure, checks the accounts they are worried about and skims the rest. They do not remember what account 6340 did in March. Software compares every account against its full history every time, without fatigue and without preferring the accounts that were interesting last month.

The accountant's advantage is knowing why — which of the flagged items is a real problem and which is a business change that happened for a good reason. That is the division of labour: the system finds candidates, the accountant decides.

From flag to explanation

A flag is only half useful. "Account 6210 moved unusually" still requires investigation.

What changes the economics is being able to go from the flag to the transactions immediately — which entries caused the movement, from which documents, coded by whom or by which process. Investigating a variance used to mean running a report, exporting it, filtering, then opening documents. When that becomes a single step, variances get investigated instead of noted.

This is why explainability is practical rather than philosophical. A system that tells you something is unusual but cannot show you why is generating work rather than removing it.

The questions worth asking of every trial balance

A short standing list, whether or not software flags them:

• Which accounts moved more than their own history suggests they should?
• Which accounts did not move and should have?
• Do the control accounts decompose into items I can name?
• Are there manual journals in accounts that normally receive only automated entries?
• Is anything sitting in suspense, and how long has it been there?
• Do the accruals and prepayments still reflect things that are true?

Those six catch the large majority of what goes wrong. They are also the ones that get skipped when the close is under pressure — which is the argument for having them run automatically rather than depending on discipline.

What it does not tell you

Anomaly detection finds deviation from pattern. It does not find consistent error.

If a cost has been miscoded the same way for two years, it is the pattern, and nothing will flag it. If a supplier has been invoicing you 4% above the agreed price since the contract started, that is stable and therefore invisible.

Those require a different control — periodic testing against source documents and agreements rather than comparison against history. Anomaly detection is a strong net for change and useless against consistency.

Common questions

What can AI find in a trial balance that a person cannot?

It compares every account against its own history consistently, without fatigue and without the bias toward accounts that seemed interesting recently. That makes it good at spotting accounts that moved unusually relative to their own pattern, accounts that should have moved and did not, and changes in the composition of a balance whose total looks normal. A person under close deadline pressure checks the accounts they are already worried about and skims the rest.

Does a trial balance that balances mean the accounts are right?

No. Balancing confirms only that double entry held. Miscoded transactions, duplicate invoices, accruals that have stopped being true and entries posted to the wrong period all balance perfectly, so the useful question is which accounts are behaving unusually rather than whether the totals agree.

What is the strongest warning sign in a trial balance?

A control account balance that no longer decomposes into identifiable items you can name. It usually indicates something has gone wrong upstream — unreversed accruals, unmatched items absorbed into a balancing figure, or reconciliation differences written off to make things agree — and the balance has been accumulating for some time.

What can't anomaly detection catch?

Consistent error. It identifies deviation from pattern, so anything wrong in the same way over a long period becomes the pattern and never flags — a cost miscoded the same way for two years, or a supplier billing above the agreed rate since the contract began. Catching those requires periodic testing against source documents and agreements rather than comparison against history.

Related: explainability in accounting automation · the report that writes itself · closing the books with AI assistance