AI Bookkeeping Agents: What They Do Alone and What You Must Review

Today's agents work alone mostly on rule-matched and high-confidence tasks: coding recurring bank lines, matching payments to invoices, reconciling clear matches and flagging oddities. A bookkeeper still reviews low-confidence items, new payees, sales-tax coding, journals, transfers, anything touching payroll or equity, and a sample of the confident postings.

Vendors differ on how much runs without you and how much you can switch off. Xero lets an admin turn automatic reconciliation on or off for each bank account, Intuit's agent auto-posts a few kinds of exact match such as bills paid through its own Bill Pay, and Double and Booke AI send what they are unsure about to a person. The design question is the same everywhere: which task may post, which must queue, and who signs.

Follow me on Instagram@sagnikteaches

Three levels of agent action, and where the gate belongs

Sort what an agent does by how far its action travels. At the first level it suggests: a category, a match or a draft sits in a queue and nothing in the books changes until a person acts. At the second it posts under conditions, which means it acts alone when a rule or a confidence test is met. At the third it acts outward, by sending a message, chasing a payment, paying a bill or filing a return, and the effect leaves your books and reaches other people.

Connect on LinkedInSagnik Bhattacharya

The gate sits between the first and second levels for anything new, and between the second and third levels always. An agent can earn level two for a narrow task after a clean run of sampled months. Nothing earns level three in bookkeeping without a person pressing the button, because a mistake there cannot be corrected with a journal.

Subscribe on YouTube@sagnikteaches

What each product's own pages say it does alone

This table comes from each vendor's pages as I read them in November 2026. The wording is theirs, and the last column is what I would look for.

Two points matter more than the feature lists. First, "highly confident" and "high confidence" are the vendors' words, and none of the pages gave a number for the threshold, so a line the agent trusts is not proof that the line is right. Second, controls differ: only Xero's page describes a switch for each bank account, which makes it the easiest agent to roll out one account at a time. Best AI accounting agents compared sets the three ledgers side by side on features, and this tutorial stays on what you review.

A review gate for each agent task

For each task below: what an agent can reasonably do alone, what a person checks, and the evidence worth keeping. The gates are my recommendations, not vendor settings.

Bank feed categorisation

• Alone: lines that match an existing rule, or a repeat payee with a stable category.
• Review: every new payee; every line marked medium or low confidence (QuickBooks shows these as blue and orange signals); anything above a value you set; payees with several plausible categories, such as a marketplace, a card processor or a big-box supplier.
• Evidence: the rule or label the agent cited, and the month's exception list.

A rule that went wrong shows why. A rule saying "contains AMAZON, post to office supplies" quietly posted a $1,240 stock purchase as office supplies for three months, because it had no amount condition. The monthly sample caught it only because the policy asked for every line over $500. The fix was an amount limit on the rule and a reclassification of three months of lines.

Matching payments to invoices and bills

• Alone: exact matches on amount and reference, and bills paid through the ledger's own bill-pay service.
• Review: one payment covering several invoices, short payments, payments with fees netted off, matches that straddle a period end, and customers with similar names.
• Evidence: the match list with any difference shown.

Consider a card-processor payout of $4,812.30 that an agent matched to a single $4,950.00 invoice. The $137.70 gap is the processor's fee. Posted as a match, it leaves the invoice partly open or hides the fee, and either way the expense is missing. The check is to read every match where the amounts differ and record the fee as its own line.

Bank reconciliation

• Alone: clear matches, when the agent is confident. Xero's page says JAX reconciles "when it's highly confident" and suggests the rest.
• Review: the closing balance against the statement, anything cleared with a date outside the period, old or large unreconciled items, duplicates, and transfers.
• Evidence: the reconciliation report stored beside the statement PDF.

The one non-negotiable check is that the reconciled balance equals the statement's closing balance, and a person looks at both numbers. An agent that reports "reconciled" has told you its own opinion. For setup and the matching rules in Xero, Xero JAX bank reconciliation: set it up and check the matches goes through them screen by screen.

Receipt and bill capture

• Alone: reading supplier, date and totals from a clear document.
• Review: tax amounts against the expected rate, currency, supplier name against your list, duplicates (the same invoice number captured from email and from a portal), personal items, and line-level coding on mixed invoices such as stock plus delivery.
• Evidence: the source document attached to the bill.

Sales-tax and GST coding

• Alone: nothing. Treat any tax-code change as a suggestion, because the right code depends on what was supplied and to whom, and an agent may default to the code last used for that supplier.
• Review: zero-rated against exempt against out-of-scope, reverse-charge supplies, mixed baskets, the tax amount on each bill against the expected rate, and the return's totals against the ledger.
• Evidence: a tax-code exception report each period.

Two sibling tutorials go deep on the rules: GST/HST coding and remittance checks with AI and AI bookkeeping tools checked against Making Tax Digital rules.

Journals, accruals and adjustments

• Alone: nothing. Double's page says its agents draft complex journal entries, and that is the right level: drafting.
• Review: every journal, with the supporting schedule, the period, the reversal date, a check that no automatic entry already covers it, and a named approver.
• Evidence: the journal with its source attached and the approver's initials.

Transfers, loans, owner draws and payroll

• Alone: a matched transfer between your own accounts when both sides are present. QuickBooks posts the other side of a transfer or card payment when you post one.
• Review: any line where the payee is a person, a lender, a tax authority or a payroll provider; owner draws and contributions; loan payments split between principal and interest; anything that moves equity or a liability.
• Evidence: the loan statement or payroll register behind the line.

A typical miss: a $2,000 loan payment coded entirely to interest expense. The loan schedule says $1,640 principal and $360 interest, so $1,640 sits in the wrong place, overstating expenses and leaving the liability too high. An agent has no way to know the split unless the schedule is in front of it.

Anomaly flags and variance checks

• Alone: flagging only.
• Review: the flags it raised and, more importantly, the ones it did not. Run three checks yourself: each account against last month, any negative balance that should not be negative, and the uncategorised or suspense balance, which should be zero at close.
• Evidence: the variance sheet with a comment on every movement you chose to explain.

Messages to customers, clients and suppliers

• Alone: drafting only. A reminder to a customer who has already paid, or a chase to a supplier whose invoice you are disputing, is a relationship problem or a cash problem, not a bookkeeping one.
• Review: balance, recipient, tone, attachments and whether anything changed since the draft.
• Evidence: a log of what was sent and who approved it.

Chat answers and report commentary

• Alone: answering questions from your data. Xero's JAX page says it does not run in the background unless you ask a question, and Intuit's page says Intuit Intelligence explains each recommendation before you act.
• Review: totals against the report, the period used, whether unposted items are included, and any answer that sounds like tax or legal advice.
• Evidence: the report you tied the answer back to.

How much of the confident work to check each month

Checking everything defeats the point of an agent, and checking nothing makes you the person who signed off unseen work. A workable middle is a fixed sampling rule that you write down, so the amount of checking does not depend on how busy the month was. These are my suggested rules, not vendor settings.

1. Check every line the agent marked low or medium confidence, and every new payee.
2. Check every line above a value you set, for example $500.
3. Check every transfer, loan, payroll and owner line.
4. Check a random one in ten of the high-confidence postings, with a minimum of twenty lines.
5. Check twenty random lines from rule-based or exact-match postings, because a wrong rule repeats itself every month.

Here is an illustrative month for an online clothing store with 640 bank lines. The groups add up to 640, and the sample covers about a third of them.

On the time side, say a bookkeeper codes bank lines by hand at about 1.2 minutes each, which is roughly 12 hours 48 minutes for 640 lines. Reviewing 218 lines at about 1.5 minutes each is 327 minutes, or about 5 hours 27 minutes, a saving of around 7 hours 20 minutes. Treat those minutes as placeholders and time your own month, because review speed depends on how clean the exception list is.

The sample also needs a stop rule. Suppose two of the twenty high-confidence lines turn out wrong, a card-processor payout posted as sales and a customer refund posted as an expense. That is one in ten. It does not prove that 18 of the 180 lines are wrong, since a sample of twenty is small and the true rate could be well below or above that. It does mean the category is not ready to run alone. The rule: more than one error in twenty sampled lines sends a second sample of twenty, and any further error sends that category back to review-all until the cause is fixed.

Misses a sample catches, and how each one shows up

A second AI as an extra net, not a replacement for the sample

A general assistant can read an export of the month's postings and point at lines that look wrong. Strip account numbers and personal names first, and use a business plan or switch off training on a consumer account. Consumer ChatGPT and Claude plans let you turn off the use of chats for model training in privacy settings, and business plans do not train on business content by default.

Below are 25 transactions an accounting agent posted this month: date, payee, amount, category, confidence.
List any line that looks miscoded or needs a supporting document, and give a one-line reason for each. Do not recode anything and do not guess at information that is not in the table.

This is an illustrative answer for a clothing store's export.

One hit, one false alarm and one miss is a normal result. The second assistant widens the net and costs a few minutes, but it has no more insight into your client than the first agent did, so the human sample stays.

A written gate policy you can adopt

Write the rules down once, per client or per practice, and the review stops depending on memory. The thresholds below are examples; set your own.

AGENT POSTING POLICY (example thresholds)
May post alone: bank lines that match an existing rule; payments matched exactly to an invoice or bill by amount and reference; both sides of a matched transfer between our own accounts.
Must queue for review: any line marked medium or low confidence; new payees; any line over $500; refunds; tax-code changes; lines in loans, equity, payroll liabilities or suspense.
Never unattended: journals and accruals; messages to customers or suppliers; payments; tax returns; changes to the chart of accounts or tax codes.
Sampling each month: 1 in 10 of high-confidence postings (minimum 20) and 20 rule-based postings.
Stop rule: more than 1 error in 20 sampled lines turns that category back to review-all until the cause is fixed.
Sign-off: the bookkeeper initials the exception report, and the reconciled balance equals the statement balance.

When to turn autonomy back down

• A new client file. Little history means weak suggestions. Intuit's help page says a clean match in a new company file may not show green until you have accepted that kind of match a few times, so start suggest-only.
• A change to the chart of accounts or tax settings. Rules written against the old structure keep posting to it. Review rules the same week.
• A vendor feature that changes. The QuickBooks Intuit Intelligence page carries a BETA label today. Read release notes, and re-run a sample after any update.
• Year-end and tax periods. Close more slowly, sample more, and keep journals human.
• A breached stop rule. One category goes back to review-all, not the whole file, until the cause is understood.

Native agent or add-on: what changes for the reviewer

An agent built into the ledger comes with the plan or a small add-on (Xero lists auto-reconcile at $5 a month on its cheapest plan), and its reasoning appears where you already work, such as the label Xero attaches to each automatic match or the badges QuickBooks shows. An add-on such as Booke AI adds a separate queue, a rationale on each flagged item and an audit trail, but also a monthly fee (listed at $129 per business) and another company that holds your client's financial data. Check its security and data terms before connecting, and ask what happens to your data if you stop paying.

Either way, test before you trust. Testing a bookkeeping AI tool on your own client files describes a trial month. If you pay for a specific firm-side agent, Double's features and cost per client is the place to check the economics; its own pricing page currently lists firm-wide add-ons rather than a per-client rate, so ask for a quote based on your client count.

Questions about letting agents post to the books

Can an AI agent post journal entries on its own?

Treat journals as draft-only. Double's page says its agents draft journal entries and that reviewing is your first task, and no vendor page I read promises safe unattended journals. Accruals, depreciation and adjustments rest on judgement and period-end facts an agent cannot see, so a bookkeeper approves each one with the source document attached.

How big a sample should I check?

Check everything the agent marked low or medium confidence, every new payee, every line above a value you set, and a random slice of the high-confidence work, such as one line in ten with a floor of twenty. If more than one in twenty sampled lines is wrong, widen the sample until you know why.

Should auto-posting be off for a new client file?

Yes, for the first month or two. A new file has little history to learn from, and Intuit's help page says even a clean match may not show green until QuickBooks has seen you accept that kind of match a few times. Run suggest-only, correct what is wrong, then allow rule-based posting for categories with a clean record.

Do these agents replace a bookkeeper?

Not on the vendors' own wording. Intuit says exceptions go to you or your accountant for review, Double says reviewing is your task, Booke AI says your team remains in control of review decisions, and Xero says its tools let you review the output. The work shifts from typing entries to checking and explaining them.

Further reads

• AI Agents for CPA Firms: Starting Jobs and a Review Gate for Each — Starting jobs and a review gate for each, from a firm's side.
• Does QuickBooks Have AI Agents? Plans, Prompt Limits and Add-Ons — Which QuickBooks plans include which AI features, and the limits.
• AI Bookkeeping Software Challenges: Where It Goes Wrong and the Fix — Where AI bookkeeping tends to go wrong, with fixes.
• Claude for Bookkeepers: Month-End Checks and Client Queries — Using Claude for month-end checks and client queries.
• How to Use QuickBooks AI Agents Without Posting Wrong Categories — The QuickBooks banking-page steps in more detail.
• Can AI Do a Restaurant's Bookkeeping? What a Person Still Checks — One restaurant month worked through: what AI bookkeeping handles for receipts, deposits and sales, and what a person still checks on tips, tax and stock.
• AI Accounting in Singapore: GST, ACRA Filings and Software Choices — How to use an AI assistant for GST F5 coding checks, a filing calendar and software questions, while a named person files with IRAS and ACRA.
• Catch-Up Bookkeeping With AI: Rescue Months of Untouched Books — A bookkeeper's month-by-month recovery method, with exception lists, client queries, reconciliation checks and a scoped clean-up quote.
• Can AI Help Accountants? What It Handles and What You Review — A small-practice split of accounting jobs into AI-reliable, AI-then-review and accountant-only, with a named reviewer and sign-off for each.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Intuit Intelligence page; Intuit help articles on AI suggestions for bank transactions, Accounting Agent features and Ready to post (updated October and November 2026); Xero pricing, bank reconciliation and JAX pages; Double homepage and pricing page; Booke AI homepage.