Claude for Bookkeepers: Month-End Checks and Client Queries

A bookkeeper can use Claude to run month-end checks and draft client query lists by connecting it read-only to the ledger, giving it a written checklist, and having it return findings, proposed journals and questions as text. Nothing posts until the bookkeeper reads each item and enters or approves it. Add write access only later, for one narrow task.

Most setup guides stop at "connect your accounting software". The missing piece is the gate: a point in the process where a person approves what Claude found, with a record of who approved what. This tutorial builds the routine around that gate, using the Claude for Small Business plugin where it fits, and a plain checklist where it does not.

Follow me on Instagram@sagnikteaches

What Claude for Small Business gives you, and what it does not

Anthropic's small business page describes a plugin installed inside Claude, covering a Monday-morning business brief, bookkeeping checks against payment tools with plain-English monthly summaries, proposals, lead replies and marketing calendars. It names Intuit QuickBooks, HubSpot, Shopify, Stripe, Gusto and Canva "and 25 more" as connectable tools, and Anthropic's Academy guide to the plugin lists a month-end skill that reconciles the books against payment processors, flags what is off and writes a close packet for the accountant. On approvals, the page says that by default nothing sends or pays without your approval, bills wait for approval before payment, Claude shows its plan before acting, and the permissions you set in each tool still apply.

Connect on LinkedInSagnik Bhattacharya

Read that carefully. It is a default behaviour, described in marketing copy, not a contractual control. Your own practice should add a gate that does not depend on how a feature is worded this month, which the sections below set out. Also note the plan: Anthropic's FAQ says Pro suits one person and Team a business with more than one person, both month to month. On the pricing page, Team lists Standard seats at $20 a seat a month billed annually or $25 monthly, from two seats, and says Team does not train on your content by default. Confirm figures on the page before you quote them to a client.

Subscribe on YouTube@sagnikteaches

The two main ledgers behave differently, and the difference shapes the gate. Claude's connector directory lists a Xero connector built by Xero and marked Anthropic verified; its listing says Claude connects with read-only access, and its tools read bank account transactions, aged receivables and payables, bills, invoices, the chart of accounts and the profit and loss. That suits month-end review, and it means journals stay something you enter in Xero yourself. Intuit's QuickBooks Labs connector, still labelled beta, goes further: its listing says it can categorise bank-feed lines, reconcile accounts and record bills, payments and journal entries, and that it previews payment links and reminders for your confirmation before they go out. With QuickBooks, read-only is something you set, not something you get.

The running example: a solo bookkeeper with fourteen clients

Take an illustrative bookkeeper who looks after fourteen small clients: a mix of trades, a café, a consultancy and a retailer. Month-end is the crunch. For each client she needs to check that bank accounts reconcile, that nothing sits in suspense, that sales and payment-processor totals agree, that bills are coded sensibly and that unusual items are queried. Then she sends a query list so the client can explain what she cannot.

On a bad month this takes about ninety minutes a client, of which perhaps half is hunting for oddities and typing the same questions. Those are the parts to hand to Claude, with her review in between.

Set access in layers: read-only first

Begin with the smallest permission that lets the task work.

1. Layer one, no connector. Export a trial balance and a bank transaction list as spreadsheets, remove account numbers and upload them to a Claude Project. Claude reads the files and returns findings. Nothing connects to the live ledger. This is the safest start and surprisingly capable.
2. Layer two, read-only connector. Connect the ledger so Claude can read transactions and reports but not change them. With Xero the connector is read-only already. With QuickBooks Labs, open the connector's tool permissions and set every tool that creates, updates, categorises or sends to Blocked; on a Team plan an owner can restrict a connector's actions for the whole organisation, and members cannot override it. Run the month-end checks live. Claude still cannot post.
3. Layer three, one narrow write. After a full month with no unreviewed errors, allow a single action, such as creating journal entries, and set that one tool to Needs approval so Claude asks before each one. Do not enable bill payment or sending client emails from Claude in the early months.

Plugins are added from the Customize menu's Plugins tab, and Anthropic's help centre says they are available on all paid plans; the Academy guide walks through this plugin's setup skill, which asks about the business and connects tools one at a time. Connectors sit under Customize, then Connectors. On a Team workspace, an owner must first add a connector for the organisation, and each member then signs in to it individually. Remove any connector you do not need for this task.

The month-end checklist, written so Claude can follow it

Claude is only as good as the instruction. Put the checklist in the Project instructions so every run starts the same way.

You are assisting a bookkeeper with month-end review for ONE client
at a time. You have read-only access. You must not post, edit, send
or pay anything.

For the period [dates], check and report on:
1. Bank reconciliation: unreconciled items, and items older than 30 days.
2. Suspense or uncategorised accounts: any balance, with transactions.
3. Sales vs payment processor: total sales against payouts plus fees;
   explain any difference.
4. Accounts receivable: invoices overdue more than 60 days.
5. Accounts payable: bills due in the next 14 days; duplicates by
   supplier, amount and date.
6. Unusual items: any transaction more than 3 times the supplier's
   average, round-sum transfers, and new suppliers.
7. Payroll liabilities: do balances look like one pay run's worth?
Output: a table of findings with transaction references, a list of
proposed journals as TEXT ONLY (debit, credit, amount, reason), and
a list of questions for the client. Mark anything you are unsure of
as UNSURE. Never invent a reference or a figure.

That "TEXT ONLY" line is the heart of the early approach. Claude proposes; the bookkeeper enters. Over time, you learn which proposals you never change and can consider automating those, but only with evidence.

What the output looks like, and what to fix in it

An illustrative result for the café client, trimmed: three unreconciled items older than 30 days, a $412 suspense balance, a payment-processor difference of $38.20, one possible duplicate supplier bill and one transaction flagged as unusual. For the suspense balance, Claude proposes "Dr Repairs and maintenance 412.00, Cr Suspense 412.00, reason: matches a supplier invoice for oven repair".

The bookkeeper's review catches that the invoice is for a new oven, which is a fixed asset, not a repair. Claude saw a match by amount and supplier; it lacked the judgement. She changes the journal, and notes in her log "capital vs expense confusion". That kind of correction is precisely what the approval gate exists for, and the log tells you which categories need a rule in the instructions, such as "any single purchase over $1,000 is queried as possible fixed asset".

Client query lists that clients actually answer

The query list is where time is wasted: vague questions, long spreadsheets and clients who reply "not sure". Ask Claude to write queries in a form a busy owner can answer in a line.

From the findings above, write the client query list. Rules:
- Group by what the client must do: confirm, find a document, or decide.
- Each query: date, amount, payee, and one plain question.
- Offer a likely answer where evidence supports one ("Looks like a
  supplier refund, correct?") so they can reply yes or no.
- Maximum 10 items; combine small items into one question.
- No accounting jargon. Friendly, brief. Sign as [your name].

Illustrative output: "1. 12 March, $412.00, paid to [supplier]: is this the new oven or a repair? A photo of the invoice helps. 2. 19 March, $1,500 transfer in from [name]: loan, owner's money or a customer payment?" Offering a likely answer raises response rates because the client confirms instead of researching. Always check each item against the ledger before sending; a wrong amount in a query makes you look careless. The prompt wording for other query types is in ChatGPT prompts for bookkeepers, which transfers directly to Claude.

Three more clients, three different checks

The checklist flexes by client type, and each type shows a different weak spot.

An online retailer. Sales come from a shop platform and payouts arrive from a payment processor net of fees and refunds. The key check is item 3, sales against payouts. Illustrative month: platform sales $31,200, refunds $1,150, fees $905, payouts $29,145. Claude should reconcile $31,200 minus $1,150 minus $905 to $29,145 and flag any gap. A common cause of a gap is timing, where a payout for the last days of the month lands in the next period, so ask Claude to list payouts by date alongside the sales cut-off.

A consultancy with a few foreign-currency invoices. The risk is exchange differences posted to the wrong account. Add a line to the instructions: "list every transaction in a currency other than the base currency and show the rate used." Claude cannot decide the correct treatment, but it can gather the items so you can decide in one pass.

A trades business with job costing. Here bills are coded to jobs, and the usual error is a material bill with no job attached. Ask for "bills over $200 with no job or class" and a count by supplier. The query list then asks the client which job each belongs to, with the supplier and date, so the answer is a job name.

Before and after: one query, rewritten

A first-draft query from a tired bookkeeper reads: "Please explain the transactions in the attached list." The client opens a forty-row spreadsheet and closes it. After the prompt above, the same month becomes four questions, each with a date, amount, payee and a suggested answer. Replies come back the same day, because each can be answered with a yes, a no or a photo. The gain is not clever AI; it is a rule that every query must stand alone and be answerable in one line.

One further rule helps: cap the list. Ten questions get answered, thirty do not. If there are more than ten, combine the small items ("six card payments under $20 at the same supplier: all office supplies?") and push anything immaterial to next month with a note.

How to check that the routine is working

Judge the routine by outcomes you can count, not by how impressive the summary reads. Track four numbers per month: minutes per client, issues found by Claude that were real, issues you found that Claude missed, and false alarms. A healthy pattern after two or three months is falling minutes, real issues found rising to match your own, misses close to zero and false alarms steady or falling.

Here is an illustrative log across fourteen clients for the first three months:

The numbers are invented for illustration, but the shape is what to look for. Misses are the figure to watch closely: a rising count means the checklist has a gap, or a connector has stopped showing something. Never reduce your own review time below what the misses justify.

Prompts for the awkward bits

Two extra prompts earn their place. The first finds duplicates and near-duplicates without relying on exact matches.

List possible duplicate payments or bills in this period: same
supplier, amounts within 1 percent, dates within 10 days. For each
pair show both references. Do not conclude they are duplicates;
say what evidence would settle it.

The second drafts the short internal summary you keep with the month's file.

Summarise this client's month in 120 words for my file: headline
balances, what changed from last month, items still open and who
owns each. Use only the findings above. Mark any figure you could
not see with NOT SEEN.

"NOT SEEN" is the counterpart of "NOT IN FILE" in other workflows: a visible marker for a gap, which is far better than a confident guess. Treat any figure marked that way as an action for you, not a detail to skip.

The approval gate in writing

Make the gate a document, not a habit. One page, signed off by whoever runs the practice:

• Nothing posts without a person. Claude may only propose. A named bookkeeper enters or approves every journal.
• Approval has a record. For each client-month, a line: date, reviewer, number of proposals, number accepted unchanged, number changed, number rejected.
• Thresholds escalate. Any proposed adjustment over a set amount, say $1,000, or touching tax accounts, payroll or loans, needs a second person or the client's confirmation.
• No payments and no client emails from Claude. Messages go out from your own mailbox after reading.
• Scope is documented. Which connectors, which clients, read-only or not, and who can change that.

A realistic weekly tally from the log might read: 62 proposals, 51 accepted unchanged, 9 changed, 2 rejected. An 82 percent unchanged rate is a number you can use to decide whether to widen the scope. If the changed share creeps up after a connector update or a new client type, tighten instructions and go back a layer.

Where it goes wrong, and how to spot it

Four failure patterns recur. First, plausible but wrong matches: Claude pairs a payment to the wrong invoice because amounts are similar. Spot it by checking that every proposed match shows the date and reference of both sides. Second, silent gaps: an account the connector cannot see, so the check passes without covering it. Ask Claude to list which accounts and date ranges it actually read. Third, stale data: a connector showing figures from before a late import. Compare the headline balance to the ledger screen. Fourth, over-confident narrative: a tidy summary that hides a flagged UNSURE. Make "UNSURE" items impossible to miss by putting them first.

A quick sum is a good sanity check on any run. If sales total $18,400, processor payouts $17,790 and fees $610, the difference is zero. If Claude reports a clean match but your own arithmetic shows $17,790 plus $610 is $18,400, fine; if it does not add up, the summary is wrong. Do the sum yourself for any figure you plan to send.

A month-end day, step by step

For one client, the routine takes roughly thirty minutes instead of ninety. Open the client's Project, confirm the period, run the checklist prompt and read the findings. Spot-check three flagged items directly in the ledger. Enter the journals you agree with. Generate the query list, edit it, and send it from your own email. Log the counts. For fourteen clients, schedule them across three days, finishing before the clients' own deadlines.

The closing packet is also worth drafting, so see the Claude month-end packet for QuickBooks and PayPal, and the QuickBooks close checklist for where AI helps and where it does not. For rolling this out across a team, rolling out AI in a bookkeeping practice covers the people side. The wider context is the AI guide.

One last worked case shows the gate earning its keep. In a busy week, Claude proposes reclassifying a $2,300 payment from "Loan repayments" to "Expenses" because the payee name resembles a supplier. The reviewer sees the amount exceeds the $1,000 escalation threshold and touches a loan account, so it goes to the client as a query instead of being entered. The client confirms it is a loan instalment. Without the written threshold, a rushed Friday afternoon might have let the change through, overstating expenses and understating the loan balance. The threshold cost two minutes and prevented a reporting error that would have surfaced, awkwardly, at year end.

Review the thresholds themselves every quarter. If almost everything escalates, the limit is too low and people will start approving without reading. If nothing ever escalates, it is too high. Adjust until a handful of items a month reach the second person, and note the change with the date.

Client consent and data handling

Before connecting a client's books, check your engagement letter. It should tell clients that AI tools may assist with review, that they are used on a business workspace that does not train on content by default, and that a person reviews all output. Keep each client in a separate Project so one client's data never appears in another's chat, and name Projects by client code, not by business name, where that helps.

Avoid pasting bank account numbers, tax identifiers or payroll details unless the task needs them. When a client leaves, remove their Project and revoke the connector, and note the date. These steps are small, and they are what a client will ask about first.

Trying it on one client before all fourteen

Choose a client with a clean ledger and a past month already closed. Run the checklist on that old month and compare Claude's findings with what you found manually at the time. Count the issues you found that Claude missed and the issues Claude flagged that were false alarms. If it misses more than a couple of real issues, add rules to the instructions. If it floods you with false alarms, set thresholds. Repeat on a second past month with a messier client.

Only then run a live month, still read-only, and keep the log. After two clean live months you can decide whether any write permission is worth the risk. For many practices the honest answer is no: the time saved is in finding problems and drafting queries, and typing a journal takes seconds. The gate costs little and the practice keeps control.

Claude for bookkeepers: common questions

Can Claude post entries to the ledger for me?

It can if you connect it with write access, but start without it. Read-only access lets Claude find problems and draft journals as text for you to enter yourself. Add write access later, for one narrow task, once a month of read-only review shows the drafts are reliable.

Is there a switch called approval mode?

Not by that name. Anthropic's small business page says that by default nothing sends or pays without your approval and that Claude shows its plan before acting. The nearest real setting is each connector's tool permissions, where every action can be set to Always allow, Needs approval or Blocked. Use those, and add your own written gate on top.

Which plan do I need?

Anthropic's pricing page lists Pro for one person and Team, from two seats, for a business. Team lists Standard seats at $20 a seat a month billed annually, or $25 monthly. Check the page for current figures and for which plan your connectors need.

Should client bank details go into Claude?

Only through connectors your engagement terms allow, on a Team workspace that does not train on your content by default. Strip account numbers from anything you paste, tell clients how you use AI tools and keep the connected scope as narrow as the task allows.

Further reads

• AI Receipt Capture and Bank Categorisation: A Bookkeeper's Setup — Receipt capture and bank-line categorisation setup.
• Xero vs QuickBooks AI: Which Saves More Bookkeeping Time? — Which ledger gives you more AI help.
• How to Speed Up Month-End Close in a Small Business With AI — Speeding up the close in a small business.
• Is AI Worth It for a One-Person Bookkeeping Business? — Whether AI pays off for a one-person practice.
• Can AI Do My Bookkeeping? What Still Needs an Accountant — What still needs an accountant.
• AI Accounting Workflow: Client Upload to Reviewed Books, Step by Step — Five stages from client upload to partner review, each with an owner, tool and check, shown through one worked month for a small contractor client.
• AI Bookkeeping Agents: What They Do Alone and What You Must Review — A task-by-task review checklist for AI bookkeeping agents: what each may post alone, what must queue for a person, and how big a sample to check.
• AI Bookkeeping in Canada: GST/HST Coding and Remittance Checks — A bookkeeper's review routine for GST/HST coding and payroll remittances: formulas for the maths, AI for the judgement calls, CRA pages for the dates.
• AI for Australian Bookkeepers: BAS Prep, GST Checks and Xero — A quarter-end BAS review that pairs Xero's automation with invoice checks, reconciliations and a documented sign-off.
• AI for Fractional CFOs: Monthly KPI Packs and Client Briefings — How a fractional CFO can run monthly KPI packs for several clients with AI: client templates, export-based data, narrative drafts and a review checklist.
• Claude for Excel for Accountants: Reconciliations a Manager Can Sign Off — A reviewable Excel reconciliation workflow: copies, source totals, exception notes, formula tests and a manager approval gate.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic's Claude pricing page and small business solutions page (approval behaviour, connectors, plans); Anthropic Academy's Small Business plugin guide; Claude connector directory listings for Xero and Intuit QuickBooks Labs; Claude help centre articles on plugins and connectors.