How Do You Rebuild a Year of Books From Bank Statements?

If you've got a year of statements but no books, you can rebuild the records. Start by listing every bank account, credit card and loan, then collect every month's statement, including scans and photographs. Test your method on a year that's already been filed before using it on the new year. Check transactions against the bank's daily balances, rather than relying on monthly totals. Write sorting rules for every spelling of a payee's name, especially after a bank switch. Keep money held for customers out of business income. Then give your accountant the finished reports and a short list of decisions. I'm not a bookkeeper by trade, and I never prepare or file tax returns. The accountant makes the tax decisions.

Key takeaways

• A year of books can be rebuilt from bank and card statements, starting with a complete list of accounts and every month's statement.
• Testing on an already-filed year checks whether your method reproduces known results before you use it on a new year.
• The daily balances a bank prints catch mistakes a monthly total can't, like two errors that cancel each other out.
• After a bank switch, recurring payments can show up under different wording, and a sorting rule written for one version can quietly miss another.
• The accountant should get the reports plus a short list of the decisions that are theirs to make.

---

What do you need before you start rebuilding the books?

You need a complete account list and every month's statements for each account. On this job, that meant four bank accounts at two banks, five credit cards and a vehicle loan. A client had a year of those records but no bookkeeping records for that year. In my experience, that's a normal place for a small business to be. The records can be rebuilt.

The accountant needed three reports:

• A profit and loss report, showing business income and expenses over the year.
• A balance sheet, showing what the business owns and owes on a particular date.
• A general ledger, the detailed record of entries behind those reports.

I used Claude in Claude Code to read, sort and build the records, with Python scripts creating the spreadsheets. I directed the work, checked it and talked to the client. Before AI, I spent about 20 years running online businesses.

The IRS recordkeeping page says, "You may choose any recordkeeping system suited to your business that clearly shows your income and expenses." So I focused on rebuilding clear records and checking them against the statements before preparing the reports the accountant requested.

How do you know your method works before you trust it?

Test it against a year with known results before asking it to build the new year. Before categorizing any new-year transactions, my system rebuilt the previous year's already-filed books from the same kind of statements. All 107 figures matched. It was a starting test, and I kept checking after it passed.

I ran that test again after every change. A change that helped the new year could still break something that had worked before. The known year gave me a way to see that happen.

Most of my time went into checking. Nearly every problem produced a confident wrong answer instead of an error message. A confident answer didn't tell me whether the result was correct. I had to check the records underneath it.

How do you check statements that were scanned or photographed?

Read the entries page by page, then replay them in order against the daily balances printed by the bank. Three months on this job existed only as photographs. There was no text to copy. The AI read 171 transactions from those pages, and I checked the sequence against 55 daily balance checkpoints.

All 55 matched exactly. That's roughly 18 checks a month instead of one. A monthly total can hide two mistakes that cancel each other out. It can also match when the right amount appears on the wrong day. Checking the daily sequence can catch both.

The statements also contained photographs of checks and deposit slips. Those showed payee names and itemized deposits that weren't in the plain transaction list. Reading them answered dozens of questions I would otherwise have emailed to the client.

IRS Publication 583 lists "Bank deposit slips" among supporting documents for gross receipts. It also says, "If you do not have a canceled check, you may be able to prove payment with certain financial account statements prepared by financial institutions." That doesn't mean statements answer every question. Your accountant decides what supporting records they need.

Anything still unclear went to the client. I used dates, amounts and shop names rather than asking them to choose accounting categories, and I kept each batch of questions small. For the final round, I grouped 35 open items into eight yes or no questions, which came back answered that morning.

What breaks when a business switches banks partway through the year?

A sorting rule can miss a recurring payment because the new bank describes it differently. The client changed banks partway through the year. The same payments appeared under two spellings, one from each bank. My rules didn't consistently handle both.

Six of the eight silent sorting bugs came from that cause. After the fifth, I had the AI build a checker that hunts for that exact pattern. It found the sixth on its first run. So I added a check for the pattern, rather than continuing to fix one entry at a time.

Payee names caused another problem near delivery. One shop's name sat inside another shop's name. Matching on part of the name nearly put four correct entries in the wrong place. I matched the client's final answers by exact date and amount instead.

Partial name matching caused trouble three times on this job. What I've found is that a name is useful evidence, but a piece of a name isn't enough to identify an entry.

What do you do when the balance sheet won't balance?

Trace the difference back to the entries and require the explanation to account for the whole gap. Near the end, my balance sheet was out by a six figure amount. Two diagnostic tools explained it confidently. Both were guessing.

Earlier, another difference had received a plausible explanation: "year-boundary timing." That explanation was wrong. My system had counted refunds as payments. The timing explanation would have left the actual mistake in the books.

For the larger gap, I required a third tool to work out what the books did with every one of the 2,246 transactions. Its explanation had to add up to the difference to the penny. Its first run missed by a wide margin. A single sign error was the cause, and the requirement to match the gap caught it.

Once the explanation tied, it identified 44 transfers back into the account holding customers' money. Those 44 transfers returned customers' money to the account holding it, so my sorting rules shouldn't have counted them as business income. Correcting those entries was the biggest correction of the job.

Ten automated checks ran after every change. Before delivery, a model from another company, GPT-5.6 through Codex, reviewed the work three times. It found that one check couldn't fail. That check calculated a number but compared it to nothing.

The model that wrote it had read it many times without noticing the problem. Basically, a check needs something to check against. There's more on that in my post on how to verify AI output.

What should the accountant actually get from you?

The accountant should get the finished reports and a short summary of the decisions they need to make. I delivered four files: the three reports and a two-page summary. The summary listed seven decisions, with the facts behind each one. The accountant decides the tax treatment.

At delivery, every transaction had a category, every account matched its statement and the balance sheet tied to the penny. Getting there took repeated corrections to my own system. Three spreadsheets alone would have moved unfinished work onto the accountant's desk.

The SBA's finance guide says, "Maintaining proper bookkeeping can help keep your business running smoothly." It lists bank reconciliation, which means checking the books against the bank's records, among the jobs an accounting professional handles.

I also made a communication mistake. During the work, I sent the client a profit figure three times. Each was honest when I wrote it, and each changed within a day. That cost more trust than any sorting error.

My rule now is that I don't send a number while any work behind it remains open. The year-end books case study has more detail about the checking and the delivery. I offer bookkeeping for small businesses, with tax decisions left to the accountant.

Frequently asked questions

How long does rebuilding a year of books take?

This job took nine days from the intake call to delivery, finishing 15 days before the deadline. That was one job, so treat it as an example. Most of my time on it went into checking.

Can AI rebuild the books for me?

AI can help read statements, sort entries and build reports. On this job, it also produced confident wrong answers and wrote a check that couldn't fail. I directed the work and checked the results. I wouldn't treat an AI-produced spreadsheet as finished without checking it against the records.

Are bank statements enough if I don't have receipts?

Bank statements may help prove payment, but your accountant decides whether they need receipts or other supporting records. IRS Publication 583 says, "If you do not have a canceled check, you may be able to prove payment with certain financial account statements prepared by financial institutions."

What should I hand my accountant?

Give them the requested reports, supporting records and a short list of unresolved decisions with the facts behind each one. There's a free checklist for your books with seven checks to run before the handover.

Is checking the monthly total enough?

A monthly total is useful, but it can hide offsetting errors or an entry on the wrong day. Check the daily balances when the statements provide them. For the three photographed months here, that gave me 55 checkpoints instead of three.