AI bookkeeper for small business: what it should and shouldn't do
An AI bookkeeper for small business should learn from your past books, ask instead of guessing, and never make tax calls. A 10-point checklist before you trust one.
By Oliver Carlin, founder of Allmata
You close the week and the books are still open. A stack of uncategorized charges sits there from the card, the store, and vendors you pay every month. You know where most of them go. You also know that a rushed pass will show up later as a profit number you cannot defend.
That pile is what an AI bookkeeper for small business is for. Not the tax return. Not a ruling on whether something is deductible. The repeated work of filing what already happened, in the way you already file it.
Most tools skip the part that keeps you safe. They guess a category from a vendor name and move on. One quiet miss is easy to shrug off. A year of them bends the picture you use to run the company. Accuracy is the job. Speed is a side effect, and only after the record is right.
Why one wrong expense matters
Bookkeeping is the memory of the business. Pricing, hiring a person, a talk with your accountant, a slow month you need to understand: all of it reads that memory. If an expense lands in the wrong account, the story of the month shifts a little. If a charge never lands at all, your costs look lighter than they were. You will not feel either mistake on the day it happens.
Profit is where the miss becomes a decision. Understated costs make a weak stretch look acceptable. A personal or one-time charge dumped into operating expenses can make a healthy stretch look like trouble. You might delay a hire you can afford, or spend against a margin you do not have. The software will not feel that. You will.
So the standard is dull on purpose. Every ordinary charge filed the way you file it. Every uncertain charge asked, not guessed. Every change visible before it hits the books. A tool that cannot meet that standard will make the file look finished while making it less true.
What it should learn before it asks
A careful bookkeeper does not arrive with a generic chart and a confident hand. They look at how this company already works. An AI bookkeeper should do the same, and it should do it before it asks you to teach it one transaction at a time.
The history worth reading is yours. Which account you use for a vendor. Which bills you split. Which deposits are income and which are transfers. What you recode after the bank feed lands, and what you leave alone. Two quiet years of those choices beat any default category list.
Then it can ask. The ask should be small enough to answer between other work. One transaction. One recommended category. One reason, tied to your history or the receipt. A plain signal of how sure it is. What it cost, in dollars, to prepare the card. If you want the evidence, it should be there: the past charges, the lesson it used, how to undo the result.
When the card is almost right, you should not have to fix the books by hand. You say what to change in your own words. The role rewrites the suggestion. You approve the rewrite. That is a correction. Retyping the entry yourself is you still doing the job.
After the approval, the write happens in your books system, not in a side spreadsheet. Then the role reads the line back. The books either match the card or they do not. “It should be updated” is not a confirmation.
What it must refuse
Taxes are a different profession. An AI bookkeeper must not decide what you can deduct, how a trust is treated, what you owe, or how to answer a notice. Your accountant or tax preparer stays the authority. A cleaner file is a gift to them. A tax opinion from a model is a risk to you. Any product that mixes “we categorize” with “we will tell you what to file” has blurred a line you should keep sharp.
It must not invent numbers in sentences. A plain-language tour of the P&L is useful when you are tired of report layouts. It is safe only when every figure was computed in code from your books. A model that writes a dollar amount from memory will eventually write a smooth wrong one. Let the wording be flexible. Keep the math fixed.
It must not see tools you did not open. The bookkeeper needs the books. It does not need the help desk, and the help desk does not need the ledger. Private channels stay private unless you say otherwise.
It must not change the file in the dark. The connection starts read-only. The first write is an approved card that states exactly what will happen. Money movement, and anything you cannot undo, needs a second confirm. That pause is the same one a careful human bookkeeper would want.
It must remember. If you answered this vendor last month, the same question should not return as a stranger. Answers become dated lessons. When your rule changes, the newer lesson wins and the old one stays archived, so you can see what used to be true.
Your accountant or tax preparer stays the authority.
P&L figures are computed in code from your books.
The bookkeeper sees the books, not the help desk.
It starts read-only. The first write is a card you approve.
A checklist for any tool
Take this to any vendor before you connect a feed. Vague answers are a no.
- Does it connect read-only until you approve a specific card?
- Does it learn from your own categorization history first? How many months?
- Does each card show the category, the reason, how sure it is, and the cost of preparing it?
- Can you correct it in your own words and get a rewrite, instead of editing the ledger yourself?
- Does it write into your books, then read the entry back?
- Does it refuse tax advice, trust questions, and filing judgments?
- Are P&L figures computed from the books, never composed by the model?
- Do your answers stick as dated lessons, one private memory for your company?
- Can you see the dollars of AI work on each task, without a credit pack to decode?
- If one narrow kind of task is ever allowed to run on its own, do you switch that on yourself? Does your correction pause it?
That last item is the one to treat as later, not as a default. A long run of work you did not have to fix is the evidence. A switch you control is the permission. A pause when you correct it is the brake. Full access to the whole ledger on day one is not a feature. It is a wager.
When a person should make the call
Keep a human bookkeeper or your tax preparer for judgment. A loan, a partner buy-in, a messy accrual, a letter from a tax agency, a question that could change what you owe: those are conversations. Someone has to sit with the facts and stand behind an answer. Software can assemble the file. It should not sit in that chair.
Keep a person when the history is the problem. If the last two years were filed three different ways, a system that learns from history will learn the conflict. That is a useful mirror. It is a bad moment to hope a tool “cleans it up” alone. Have your preparer set the rule. Then let the tool follow the rule without drifting.
Keep a person when you cannot explain a figure. Ask the software for the computed number and the accounts under it. Ask your preparer what it means for the business. A paragraph of fluent explanation is not a substitute for understanding.
There is also the odd transaction that is odd on purpose: a one-time legal bill, a personal charge that hit the business card, a refund that is not revenue. Those are asks, not guesses. If a tool cannot pause on them, it is not ready.
| Bookkeeper | A person | |
|---|---|---|
| Known vendors | Drafts a card, you approve | Sets the rule if history conflicts |
| P&L figures | Computed in code from your books | Explains what they mean |
| Tax, trusts, filings | Does not decide | Your accountant or preparer |
| Loans, notices | Can assemble the file | Sits with the facts |
How Allmata does it
Allmata is an AI team for owners. The Bookkeeper is live today. It works inside your books system: live with Finaloop, and with QuickBooks or Xero on request. The Bookkeeper page is the short version of that scope.
Connections start read-only. The Bookkeeper learns from up to 24 months of your categorization history before it asks. Each uncategorized transaction becomes a card: recommended category, why, and how sure. You approve, it categorizes in your books, then reads the entry back to confirm.
It never guesses on taxes or trusts. Your accountant or tax preparer stays the tax authority. Allmata gives no tax advice. A P&L is explained in plain words, and every number is computed in code from your books, never written by a model.
Your answers become dated lessons in one private company Brain, shared only by that company’s agents. It also learns how you’ve filed every transfer account before, so owner draws and money moved between your own accounts aren’t counted as income or expenses. You can read, change, and download that Brain. Money, and anything that cannot be undone, needs a second confirm. Letting one kind of task run on its own, after a long record where you never had to fix the work, is coming soon. You switch it on. A correction pauses it.
The same order is spelled out in Read-only first, and the full trust model is on security. Systems that connect are listed on integrations.
Where to start
Pull the last month. Look at what is still uncategorized, and at what you recoded by hand. That set is the test. An AI bookkeeper for small business should treat those charges the way you already do, or ask in one card you can answer quickly. If it cannot, keep it away from the file.
If you are unsure what an AI team can take this week, get the free AI Workforce Blueprint. A few questions, then a clear look at the work that can move. If you already want the Bookkeeper on your own books, the 14-day trial includes $10 of AI work on us, and it does not ask for a card. Plans start at $79 a month. You see the dollars each task cost.
The inbox has the same shape of risk. AI customer service for a Shopify or WooCommerce store is the matching standard: look up the fact, draft the action, wait for your OK.
Questions people ask
Does an AI bookkeeper replace my accountant?
No. It can sort transactions from your own history and explain a P&L with figures computed from your books. Your accountant or tax preparer remains the authority on taxes, trusts, and filings. Allmata gives no tax advice.
How far back should it study my categories?
It should learn how you already file vendors before it asks you anything. Allmata's Bookkeeper learns from up to 24 months of your categorization history.
Can it change my books without approval?
Not under this model. The connection starts read-only. A category is written only after you approve a card that states the change. Anything that moves money needs a second confirm.
What happens when a suggestion is wrong?
You say what to change in your own words and the role rewrites the card. That answer becomes a dated lesson, so the same vendor is not treated as a brand-new question next month.
Which books systems does the Bookkeeper use?
It is live with Finaloop. QuickBooks and Xero are available on request. After you approve a card, it categorizes in your books and reads the entry back to confirm.