Preparing statements ≠ reviewing them

I took a course on monthly accounting review a while back, and one case has stayed with me since. The trial balance handed over for review looked spotless. Every balance posted, bank reconciliation clean, nothing left hanging. And then there was one line that read "trade payables" with an empty amount column.

A preparer scrolls past that line without a second thought, because there is no number to check. A reviewer stops right there and asks: the company traded all month, bought things, paid people, ran up costs every day. Is it really possible that it owed nobody a single baht at month end?

That question did not come from a vague feeling that something was off. It came from knowing that "no number" does not mean "nothing to check". And that is the whole difference between someone who can prepare statements and someone who can review them.

Part 1Preparer and reviewer ask different questions

Both roles sit in front of the same stack of documents with different goals in their heads.

A preparer asks: is everything recorded, is it in the right account, did we close on time. Those are answerable against the paperwork on the desk. A reviewer asks a layer up: is this balance right and reasonable for a business like this one, which is a question the document in front of you cannot answer. You have to know how the business actually makes money before you can answer it.

The four questions in a reviewer's head

When a balance catches the eye, a reviewer does not guess at what went wrong. The same four questions get fired every time.

1. Where did this number come from?
2. What document supports it?
3. Does it make sense?
4. How does it compare with the prior period?

They look unremarkable, but this is what turns suspicion into work that can be handed over and re-checked, rather than a hunch you cannot explain to anyone else. There is a fifth question people tend to forget: what will the client do with this set of statements next? A bank loan application and a filing with a government agency put the scrutiny in completely different places.

Part 2Skepticism is trained, not inborn

The word skepticism makes people read it as temperament. Some are born suspicious, some take things at face value, end of story. In reality, people who review well are not suspicious of everything. They have a clear order in which they look, and that order can be taught.

Two cases from the same workshop show why. In the first, someone booked last year's office rent as this year's expense. Nobody catches that by reading the receipt, because the receipt is genuine. But compare this period's rent against the prior period, see 600,000 against 17,000, and the anomaly surfaces on its own.

In the second, gross profit jumped from 32% to 39% in a single month. Nothing about that number breaks a rule, and every entry had its documents. Digging in showed accrued revenue had been booked without the matching accrued cost, so revenue rose with nothing beneath it.

Notice that in both cases the anomaly is not in any document, it is in the relationship between numbers. You need a second thing placed beside the first before it becomes visible. That is exactly why professional skepticism is trainable: it is not sitting around waiting to feel that something is odd, it is picking things up and comparing them in a set order.

Part 3Where AI still cannot take over

Now the part I ran into myself. Reviewing financial statements is not my day job, but the shape of my work is identical: checking what a machine produced. Three things I got wrong in the past month line up with the first three reviewer questions exactly.

The answer key can be wrong too

While measuring how accurately a model reads documents, I pulled out 46 cases where the model's document number did not match the answer key. Trust the key and it ends there: the model got 46 wrong. Instead I had two unrelated models read them again as witnesses, and found the answer key itself was broken in at least 10 of the 46. Some entries had been truncated. Some had grabbed the 13-digit tax ID and recorded it as the document number. On some pages the person building the key had skipped Thai script entirely.

Had I not questioned the key, I would have "fixed" it into something worse with my own hands, and every accuracy figure reported afterwards would have been wrong across the board. This is reviewer question one, where did this number come from, aimed at the ruler you are measuring with rather than at the thing being measured.

Reading through a summary means inheriting its framing

On another job I had AI read a long legal document and summarise it. It came back with a confident heading: requirements that must be met. I repeated that onward, until someone asked me whether I was sure. Going back to the actual sentence word by word, the original said may, not must. In a legal document that single word is the entire matter, the difference between "you have to" and "you can if you want".

The failure was not that the AI summarised badly. It summarised exactly what I asked for. The failure was that I read through the summariser's eyes and took its framing as fact without going back to the source. This is question two, what document supports it, and "the AI summarised it for me" is not an acceptable answer.

Every line on a screen is a claim

The third one travelled further than the other two, because it left my hands. I copied a line off a screen into a document meant for someone else to read. The line said: every access is logged and auditable. It sounded great. I had never opened the code to check whether that was true. When I did, the system logged the command name to a file and nothing else. No record of who called it, no record of which company it belonged to, and no retention period at all.

In a reviewer's language, that is a balance with no supporting document. The difference is that in financial statements everyone knows a balance needs backup, while a line on a screen or a summary from an AI slips past because people forget it is a claim that needs backing just the same.

System defaults go wrong quietly

Back to accounting, two cases turn up often enough to be alarming. Accounting systems will switch on 7% VAT automatically for a customer whose name is in English, when a genuinely overseas customer should be at 0%. And the payee default is usually set to "juristic person", so payments to individuals land on the wrong tax form.

In neither case is the system broken. It is faithfully doing what it was configured to do, and nobody told it this case was different. All four stories collapse into one sentence: a tool answers the questions you ask it very fast, but it does not know what you forgot to ask. And that trade payables line sitting at zero never emitted a signal to anyone.

Part 4Running the four questions against AI output

The good news is that you do not need a new framework. The reviewer's question set applies to AI output directly.

1. Where did this number come from? Make it show provenance for every figure, not just the final answer. And if there is an answer key you are scoring against, question the key too.
2. What document supports it? Anywhere money or law is involved, open the original and read it yourself at least once. Do not read through a summary.
3. Does it make sense? Compare against what you know about this business, not against how confident the answer sounds.
4. How does it compare with the prior period? Variance and ratios are still the best instruments for catching odd things, and they are work AI genuinely does make faster.

Two more from last month's work. First, when deciding whether something is right or wrong, get at least two independent witnesses. Do not let a single one adjudicate. Second, check what is missing, not only what is there, because an entry that does not exist will not appear in any report.

Where to start

You do not need to rebuild the whole process. Start with a single period.

1. Take the three lines that moved most against the prior period and run the four questions on each.
2. Take one line sitting at zero or blank that should not be zero for this kind of business, and chase down why.
3. Write down what this round turned up, and start next period from the same list.

Repeat for two or three periods and it becomes a review sequence you carry, rather than a feeling you cannot explain to your team. The full seven-section checklist and the competency model used to train a whole team stay inside the training course, because they have to be tuned to each firm's kind of business anyway.

AI will keep getting better at answering the questions we ask it, faster than people, cheaper than people, and never tired. What stays ours is knowing what to ask, and having the nerve to stop at a line where everything already looks fine.

If this is your lane, two related pieces: vouching entries against source documents is question two in practice, and bank reconciliation that produces an auditable report shows what "it balances" still fails to prove.