How AI bookkeeping works: from bank feed to closed month

AI bookkeeping connects to your bank and cards, cleans up each transaction, applies your rules first, and lets an AI model categorize the rest with a confidence score. Transactions it is unsure of go to you for review. Then it matches transfers and payouts, reconciles each account to the bank statement, and closes the month.

Updated September 30, 2026 · 5 min read · By the Accountable team

The short version

The steps are the same in most tools: connect feeds, clean vendor names, apply rules, categorize with a model, route low confidence to a person, reconcile, close.
Rules handle repeat vendors the same way every time, and the AI model handles new or ambiguous ones.
A confidence score decides who sees a transaction: high confidence posts, low confidence becomes a question for you.
Matching turns raw bank lines into correct books, for example by recognizing a transfer between your own accounts as not income or spending.
Reconciliation to the bank statement is the step that turns AI suggestions into books you can trust.

AI bookkeeping does the same seven steps a human bookkeeper does, faster

Step 1: Connect the sources. Bank, card, payroll and payment accounts send transactions through read-only connections, usually within minutes of posting.
Step 2: Clean each line. A raw description like “SQ *BLUE BOTTLE 0042” becomes a vendor name, and duplicates are dropped.
Step 3: Match what is not new activity. A move from checking to savings is a transfer, and a Stripe payout is a deposit of money already earned.
Step 4: Apply your rules. “Figma is Software” runs before any AI is asked.
Step 5: Categorize the rest with a model that reads the vendor, amount, memo and history, and returns a category with a confidence score.
Step 6: Send doubts to a person. Low-confidence transactions wait in a review queue with a plain question.
Step 7: Reconcile and close. Each account's ledger balance is compared with the bank statement, and the month is locked when they match.

Rules handle the repeats and the model handles the new

A rule is an exact instruction: when the vendor contains “Figma”, categorize as Software. It is fast, free and predictable, which is why it should run first. A model is for the long tail, the odd vendor you have never paid before or a charge that could go two ways.

Intuit's own researchers describe the same split for QuickBooks. Their published model, Rel-Cat, reads the transaction description together with the relationships between a business, its vendors and its accounts, and they report that it beats QuickBooks' production model while handling new customers with little data. Mercury says its AI pre-fills accounting codes from your categorization history, and lets you overwrite each suggestion. The pattern is general: history first, a model for what history cannot answer, and a person for the rest.

In Accountable, every rule runs before the model, a rule can be created from any correction with “Always categorize this vendor this way”, and each transaction's panel shows why it was categorized: a rule's reason, the AI with how sure it was, the bank or payroll feed, or a person.

A confidence score decides who looks at a transaction

A model does not just name a category. It also reports how sure it is. The tool then uses that number as a gate: confident answers post, unsure answers wait. The gate is the difference between automation that saves time and automation that hides mistakes.

Accountable shows the gate on each transaction: Sure (95%), Fairly sure or Unsure. Anything it is not sure about goes to a Needs review tab, grouped by vendor, where one answer such as “Yes, Software” can categorize all 14 payments to Webflow and make a rule. In Accountable's own test of 223 labeled startup transactions, every transaction posted without review was right, and about 1% went to the review queue. That is our test, not an independent audit. Digits says its AI auto-books more than 95% of transactions; that figure is Digits' claim.

What the confidence score does to a transaction
Confidence	What happens	Who decides
High	The category posts and the reason is recorded	The rule or the model, reviewable at any time
Low	The transaction waits in a review queue with a question	You or your teammate, in one tap
Answered	The answer applies and can become a rule	You, once per vendor

Matching is what makes bank lines into correct books

A bank feed shows money moving. Books need to know why. Three matches do most of the work.

Transfers between your own accounts are matched on both sides and never counted as income or spending. Miss this and revenue and expenses both look bigger than they are.
Payment processor payouts are matched to the bank deposit. A $337.53 Stripe payout is booked as cash in, against the $399.00 of sales, the fees and the refund behind it. The Stripe accounting guide shows the entries.
Receipts and invoices are matched to the transaction they prove, so the category has a document behind it.

Reconciliation is the step that makes AI bookkeeping trustworthy

Categorizing tells you what each transaction was. Reconciling proves none were missed, duplicated or changed. For each account, the ledger balance must equal the bank statement's ending balance, with a difference of exactly $0.00 once pending items are explained. Nothing else in the process can check the AI independently.

That is why a benchmark that let AI models close a real company's books month after month found drift of more than 15% after several months even though the models started within 1% of a CPA: they were free to reach a matching number in the wrong way. See the guide to AI bookkeeping accuracy for how to check it yourself.

People still decide the judgment calls

AI is strong on volume and weak on intent. Some decisions depend on facts that are not in the bank feed, so a person should make them, or confirm the AI's suggestion.

Whether money from a founder is a loan or equity.
Whether a $3,000 laptop is an expense this year or an asset that is depreciated.
Whether a payment to a person is wages or a contractor fee.
Whether a refund belongs against this month's revenue or last month's.
Anything a tax position depends on, which belongs with your CPA.

Questions founders ask

How does AI categorize transactions?

It reads the vendor name, amount, memo and your history, compares them with known patterns, and returns a category with a confidence score. Rules you create run first, and low-confidence transactions go to you for review.

What is a confidence score in bookkeeping?

It is the model's estimate of how likely its category is right. Tools use it as a gate: high scores post automatically and low scores wait for a person.

Does AI bookkeeping replace a bookkeeper?

It replaces the repetitive work: categorizing, matching and reconciling. A person or your CPA still decides judgment calls such as loans versus equity, capitalizing purchases and tax positions.

Does AI bookkeeping learn from my corrections?

In most tools a correction can become a rule, so the same vendor is handled the same way next time. In Accountable, ticking Always categorize this vendor this way creates that rule.

Is AI bookkeeping accurate?

It is accurate on routine transactions and needs checking on unusual ones. The proof is reconciling every account to its bank statement to $0.00.

Every transaction categorized, with the reason

Accountable categorizes each transaction by a rule or by its AI, shows how sure it was, asks you about the rest, and matches every account to your bank before the month closes.

Start free

Sources, checked September 30, 2026: