On this page
Can an AI agent do bookkeeping?
Partly, and only with review. An AI agent for bookkeeping, as accounting software ships it today, can suggest categories for your bank and card transactions, match them to invoices and bills, and reconcile the ones it is confident about. A person still reviews what it proposes, decides unusual items and signs off each reconciliation.
This runs in production at two independent accounting software vendors, by their own account. Intuit told investors in February 2026 that its accounting agent categorized over 237 million transactions in January 2026, which it said was over half of all transactions categorized that month. Xero reports that its Auto Bank Reconciliation, still in beta, has reconciled more than 100 million transactions. Both are vendor figures, and neither vendor publishes an error rate. Microsoft offers a similar reconciliation assist in Business Central, labeled a preview.
So treat this as a supervised hire. The AI moves your bookkeeping from typing to reviewing. Someone who understands your books still owns them.
This page covers the weekly loop of categorizing and reconciling. For supplier invoices see accounts payable, for closing the month see month-end close, and for the role as a whole see bookkeeper.
What the agent does and what stays with you
Agent
1of 8
- Post routine payroll and bill-payment matches
Agent with review
5of 8
- Suggest categories, customers and vendors for bank and card transactions
- Match bank statement lines to invoices, bills and ledger entries
- Post unmatched bank amounts to suggested GL accounts
- Ask for missing receipts and context on transactions
- Find reconciliation and categorization issues
Human
2of 8
- Decide the treatment of unusual items such as prepayments, splits and owner transactions
- Sign off each reconciliation and close the period
8 tasks: 1 agent, 5 agent with review, 2 human
| Task | Who does it | Why |
|---|---|---|
| Suggest categories, customers and vendors for bank and card transactions | Agent with review | QuickBooks groups transactions its Accounting AI is confident about into a ready-to-post list that a person reviews, can change and posts. |
| Post routine payroll and bill-payment matches | Agent | QuickBooks documents auto-posting matches for paychecks and payroll taxes run through its payroll and for bills paid with QuickBooks Bill Pay, labeled with an AUTO badge so a person can check them afterward. |
| Match bank statement lines to invoices, bills and ledger entries | Agent with review | Xero auto-reconciles only transactions it is highly confident about and records the method used, and Microsoft's Copilot saves no proposed match until a person selects Keep it. |
| Post unmatched bank amounts to suggested GL accounts | Agent with review | Microsoft documents Copilot proposing a G/L account for unmatched lines, which a person reviews line by line, can change, and keeps or discards. |
| Ask for missing receipts and context on transactions | Agent with review | QuickBooks documents sending one link to request details and files, with its AI helping verify that the right receipts are uploaded. |
| Find reconciliation and categorization issues | Agent with review | QuickBooks documents AI that explains issues it finds during reconciliation and flags duplicates and miscategorizations in reports, which Intuit calls a supplemental review covering most, not all, accounts. |
| Decide the treatment of unusual items such as prepayments, splits and owner transactions | Human | No cited source documents an agent deciding accounting treatment; QuickBooks only flags some issues, such as income that should have been deferred. |
| Sign off each reconciliation and close the period | Human | No cited source documents an agent signing off a period, and Microsoft posts a fully matched reconciliation automatically only if a person turns that option on. |
These products leave review controls in different places: QuickBooks groups confident suggestions for posting and auto-posts a limited set of matches; Xero auto-reconciles only when confidence is high; Microsoft presents proposed matches for a person to review and keep.
- QuickBooks. Intuit's documentation says its Accounting AI groups transactions it is confident about into a "ready to post" list that you review and post, with changes if needed. It auto-posts a narrow set of matches (payroll run through Intuit's payroll, bills paid through QuickBooks Bill Pay) and marks them with an AUTO badge. On higher plans it looks for issues during reconciliation and analyzes balance sheet and profit and loss reports for problems such as duplicate transactions and miscategorizations. Intuit calls that analysis a supplemental review covering a majority of accounts, not all. Intuit also describes the AI asking questions about transactions that need more context and updating them with the answers.
- Xero. Xero says Auto Bank Reconciliation, powered by its AI assistant JAX, reconciles a transaction only when it has a high level of confidence and leaves the rest for you. Its Reconciled page shows whether each item was reconciled by a match, a rule, memory, a prediction or by you, and lets you filter the automatic ones. Xero reports that nearly three in four customers say it saves them time, based on its own survey of 203 small businesses and 120 accountants and bookkeepers using Xero globally, in April 2026.
- Microsoft Business Central. Microsoft documents Copilot running after the built-in automatic matching to propose more matches by date, amount and description, including one bank line paying several invoices. No proposal is saved until a person selects Keep it. For lines it cannot match, Copilot suggests a G/L account that you can change before posting.
What none of the cited sources show is an agent deciding how to treat an unusual item or signing off a period. Those stay with you.
A worked example: one week of bank lines
This is an illustration, not a customer story, and the accounting treatment is an example, not advice for your books. Assume, for this example, that your software's AI behaves as the documentation above describes.
On Monday, 140 bank and card lines from the previous week sit in the feed. The AI matches a $3,200 customer payment to the open invoice for the same amount and reconciles it. It matches the payroll run to the payroll records. It puts 96 card charges into the ready-to-post list with suggested categories, and leaves 12 lines it is not confident about.
Your bookkeeper filters the auto-reconciled items by method and opens the five matched by prediction. One $1,240 charge from a software supplier was coded to "Software subscriptions", but the bill attached covers two years, so the bookkeeper records half as prepaid. They post the ready-to-post list after changing two categories, then work the 12 open lines: three transfers between your own accounts, one owner withdrawal, and eight card charges with no receipt, which go to the request-for-information link. At month end, a person checks that each reconciled balance agrees with the statement before closing the period.
What you need to provide
The agent touches financial data (every bank and card transaction, balances and payees) and personal data, because payees include employees, customers and sole traders. Give it the least access that does the job:
- Bank and card feeds in your accounting system, or statement files for accounts that are not connected. Microsoft's assist, for example, needs an imported statement file when the bank account is not linked online.
- A user role that can categorize and match but cannot pay bills, change bank details or close periods.
- Your chart of accounts, existing bank rules and enough posted history for the AI to learn how you code each payee.
- The inbox or folder where receipts and bills arrive, so matches have documents behind them.
- Written rules for the items that trip up any bookkeeper: transfers, loan payments, owner transactions and prepayments.
Check the feature's plan and limits before you rely on it. Intuit's reconciliation and report analysis are on QuickBooks Online Plus, Advanced and Intuit Enterprise Suite only. Microsoft's assist handles around 3,000 open bank ledger entries in a statement's date range on current versions, and requires an administrator to activate the capability.
Records stay your responsibility. The IRS says you must keep records as long as needed to prove the income or deductions on a tax return, and that the burden of proof for entries on your return is yours. Software doing the matching does not change who has to back up the numbers.
Job brief template
Use this as the brief for an operator, a bookkeeping firm or your own team setting up the AI. Replace the bracketed values with your own.
Copy-ready brief
- Goal
- Every bank and card transaction from the previous week is categorized, matched to its document and reconciled, with exceptions listed, by the second business day of each week.
- Deliverable
- Bank and card accounts in the accounting system reconciled to the statement balance, with an exceptions list (uncategorized items, missing receipts, unusual amounts) sent to the bookkeeping owner.
- Inputs
- Bank and credit card feeds connected to the accounting system, or statement files for unconnected accounts
- A user role in the accounting system that can categorize and match but cannot pay bills, change bank details or close periods
- Chart of accounts, existing bank rules and 6–12 months of posted history
- Access to the inbox or folder where receipts and bills arrive
- Written rules for recurring items: owner transactions, transfers between accounts, loan payments and prepayments
- Acceptance criteria
- Each reconciled account's balance agrees with the bank or card statement balance for the period.
- On a 50-transaction sample, the category matches what the bookkeeping owner would choose with at most 2 errors.
- On the same sample, every transaction matched to an invoice or bill points to the right document, with 0 errors.
- No transaction older than 10 business days is left uncategorized without appearing on the exceptions list.
- Every transaction above an agreed amount (for example $500) has a receipt or bill attached or is on the exceptions list.
- Systems
- accounting system (for example QuickBooks Online, Xero or Business Central)
- bank and credit card feeds
- receipt capture or email inbox
- Constraints
- Bank data shows supplier, customer and employee names and amounts; access is limited to the accounting system and the receipts inbox.
- The agent's account cannot pay, transfer money, change bank details or lock a period.
- Your record retention rules apply to copies of receipts the provider stores.
- Volume
- Example: 300–3,000 bank and card transactions a month across 2–6 accounts
- Cadence
- Example: daily categorization, weekly reconciliation, monthly sign-off
- Human checkpoints
- Daily or weekly: the bookkeeping owner reviews the ready-to-post or proposed matches and the auto-reconciled list filtered by method.
- Unusual items (prepayments, splits, owner transactions, transfers): the bookkeeping owner decides the treatment.
- Month end: a person compares each reconciled balance with the statement and signs off before the period is closed.
Keep the acceptance criteria. A 50-transaction sample checked against your own coding tells you more about your books than any vendor's usage number.
Risks and human checkpoints
A transaction is matched to the wrong invoice or bill, for example two invoices with the same amount from one customer.
How you notice it
Filter auto-reconciled items by method each week and open every match made by prediction or memory above your amount threshold.
A plausible but wrong category is posted, such as a two-year subscription coded entirely to this year's expense.
How you notice it
Sample 50 posted transactions a month against the bookkeeping owner's coding, and read the AI's report insights on the balance sheet and profit and loss.
One payment covering several invoices is left unmatched or matched to only one of them.
How you notice it
Xero names this as a case it is still working on; review open receivables after each reconciliation for partly paid invoices.
Transactions fall outside the tool's limits and are silently left for manual work.
How you notice it
Check the uncategorized and unreconciled counts weekly; Microsoft documents a limit of around 3,000 open bank ledger entries in the statement's date range for its assist.
Entries are posted without the receipts you would need to support them on a tax return.
How you notice it
Run a monthly report of transactions above your threshold with no attachment and clear it before close.
Human checkpoints
- Daily or weekly: the bookkeeping owner reviews the ready-to-post or proposed matches and the auto-reconciled list filtered by method.
- Unusual items (prepayments, splits, owner transactions, transfers): the bookkeeping owner decides the treatment.
- Month end: a person compares each reconciled balance with the statement and signs off before the period is closed.
The main risk is not that the AI leaves work undone; that shows up in the uncategorized count. It is that it does work confidently and wrongly, and the error sits in reconciled books until a tax return or a lender asks about it. Xero's own design answer is to show the method behind each reconciliation, and Intuit labels what it posts automatically. Use those labels: they tell you where to spend review time.
On cost, the closest official benchmark for the person doing this work is the 2025 national median annual wage of $50,670 for bookkeeping, accounting and auditing clerks (SOC 43-3031), as published on O*NET OnLine from BLS data, wage only. The AI features cited here are included in accounting subscriptions rather than priced per transaction, so this page does not put a number on the agent side. The real cost to measure is your reviewer's time.
How to run a trial
Run the AI on one account for a month alongside your current process.
- Pick one busy account, such as the main operating account or a company card.
- Turn on the AI features for that account only and limit the user role as described above.
- Each week, sample 50 transactions the AI categorized or matched and compare them with what your bookkeeper would have done. Count errors against the brief's criteria.
- Open every auto-reconciled item matched by prediction or memory above your amount threshold.
- Time your reviewer. Minutes per week spent reviewing and fixing is the number that decides whether it pays.
- At month end, confirm the reconciled balance agrees with the statement and list what the AI left undone.
Pick one busy account
Such as the main operating account or a company card
Turn on the AI features for that account only
Limit the user role as described above
Each week, sample 50 transactions
Compare with what your bookkeeper would have done and count errors against the brief's criteria
Open every auto-reconciled item
Those matched by prediction or memory above your amount threshold
Time your reviewer
Minutes per week spent reviewing and fixing is the number that decides whether it pays
At month end, confirm the reconciled balance
It should agree with the statement; list what the AI left undone
If two months in a row meet the criteria, extend it to the next account. For more on structuring the test, see how to run a paid trial with an AI agent, and for writing the brief, how to write a job description for an AI agent. Related jobs: expense review and audit and the finance overview.
This page is not financial, tax or accounting advice.
Sources
- Learn about Accounting AI features
Intuit QuickBooksPrimaryAccessed Nov 14, 2026
- Overview of Intuit AI in QuickBooks Online
Intuit QuickBooksPrimaryAccessed Nov 14, 2026
- Intuit second-quarter fiscal 2026 conference call remarks (February 26, 2026)
Intuit Inc.VendorAccessed Nov 14, 2026
- What's new in Auto Bank Reconciliation
XeroVendorAccessed Nov 14, 2026
- Reconcile bank accounts with Copilot (preview)
Microsoft LearnPrimaryAccessed Nov 14, 2026
- 43-3031.00 - Bookkeeping, Accounting, and Auditing Clerks
O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026
- Recordkeeping
Internal Revenue ServicePrimaryAccessed Nov 14, 2026
What we could not verify
- Independent accuracy rates for AI transaction categorization or reconciliation. The only usage figures found are vendor-reported (Intuit's 237 million transactions categorized in January 2026, Xero's 100 million transactions auto-reconciled and its survey of Xero users).
- Whether Intuit's 237 million 'categorized' transactions were posted by a person from the ready-to-post list or posted automatically; Intuit's earnings remarks do not say.
- A per-transaction or per-month price for the AI features alone. The cited features are bundled into accounting subscriptions, and plan prices were not checked, so no cost comparison is given.
- How much reviewer time the features save for a given business. Xero quotes customers, but no source measures it independently.
- Production evidence for Microsoft's bank account reconciliation assist; Microsoft's documentation labels it a preview and gives no usage figures.
Related
- AI agents for finance: jobs they do and how to hire one
- Can AI replace bookkeepers? A task-by-task answer
- AI agent for month-end close: what it does and needs
- AI agent for accounts payable: what it does and costs
- AI agent for expense audit: what it checks, who approves
- AI agent job description: how to write one, with template
- How to test an AI agent before buying: run a paid trial