Journal entries with Copilot, Claude or ChatGPT: draft and check before posting

Copilot in Excel, Claude and ChatGPT can all turn a month-end support schedule into a draft set of journal entries, but none of the three vendors documents its tool as a substitute for review before posting. Microsoft's Copilot in Excel FAQ states that Copilot "can sometimes make mistakes, misinterpret information, or produce inaccurate results" and tells users to verify output before relying on it, and it specifically cautions against relying on it for decisions in finance. A draft from any of these assistants is therefore a starting point for the preparer, and the posting decision stays with the person who signs the entry.

The practical workflow is short. The accountant gives the assistant the source facts for each entry (an invoice estimate, a policy schedule, a fixed asset register line, a contract term, a loan agreement), asks for balanced entries in a fixed column layout, and then checks the result against an independent calculation before anything reaches the general ledger. The six October 2026 adjusting entries below, which total $14,142 in debits and $14,142 in credits, give a complete answer key that a reader can compare with whatever the AI produces.

The documentation for all three products was checked on November 3, 2026. Plan names, rollout stages and file limits change often, so the requirements table later on reflects what the vendors' help pages said on that date.

··········

HOW AI-DRAFTED JOURNAL ENTRIES FIT INTO THE MONTH-END CLOSE.

Most adjusting entries at month end follow a small number of patterns: accruals for expenses incurred but not yet invoiced, amortization of prepaid costs, depreciation, recognition of deferred revenue, interest accruals and reclassifications of amounts coded to the wrong account. Each pattern has a fixed debit and credit direction and an amount that comes from a schedule. The repetitive part of the work is turning those schedules into entry lines with the right accounts, amounts, dates and descriptions, and that part suits an AI assistant well because the logic is rule-based and the inputs are already in a spreadsheet.

The judgemental part stays with the accountant. Whether an estimate is reasonable, whether a cost belongs in this period, which day-count convention a loan agreement uses and whether an accrual should auto-reverse are questions the assistant can only answer from what it is told. When those facts are missing from the prompt, a language model will often fill the gap with a plausible assumption and present it with the same confidence as the facts it was given. The debit and credit rules themselves are covered in our article on journal entry debit and credit rules with practical examples, and the review steps below assume that knowledge.

Microsoft has also announced accounting capabilities for Copilot in Excel that draft accruals, reclassifications, allocations, reversals and recurring entries with checks for required inputs and balanced entries. According to Microsoft's October 29, 2026 blog post, those accounting capabilities are available in Frontier, Microsoft's early-access program, while a separate set of finance skills in Copilot in Excel is generally available. Our coverage of Microsoft's Copilot for finance launch with accounting skills in Excel describes that announcement in detail. The method in the rest of the piece works with the generally available assistants as well.

··········

COPILOT IN EXCEL, CLAUDE AND CHATGPT: DOCUMENTED REQUIREMENTS AND LIMITS.

The three tools work in different places. Copilot runs inside the Excel workbook and can edit it directly. Claude and ChatGPT work on uploaded files in a sandboxed Python environment and return a new or edited file. The table summarizes what each vendor's help center documents.

........

Item	Copilot in Excel	Claude	ChatGPT
Where it runs	Inside Excel (Windows, Mac, web)	Claude apps, with file creation enabled	ChatGPT, data analysis with Python
Plans documented	Personal or Family with AI credits, Microsoft 365 Premium, commercial Copilot, or Copilot Chat-eligible business plans	Free, Pro, Max, Ultra, Team, Enterprise	Not stated on the data analysis page
Spreadsheet formats	Modern formats such as .xlsx; Strict Open XML not supported	Creates and edits .xlsx; reads CSV and TSV	.xls, .xlsx, CSV
File size limit	Not documented on the FAQ page	30MB per file	About 50MB for spreadsheets
Setting to check	Calculation Options set to Automatic for editing	Settings, Capabilities, file creation toggle	Upload limit of 10 files per conversation

........

Two documented details affect journal entry work directly. Copilot editing is supported only when Excel's Calculation Options are set to Automatic, so a workbook kept on manual calculation for performance reasons has to be switched before Copilot can write entries into it. On the Claude side, the help center notes that file creation draws on the same usage limits as the plan and consumes more of them than an ordinary chat, which matters for a close with dozens of entries.

ChatGPT's help center explains that the data analysis environment writes and runs pandas code on the uploaded file and that the environment cannot make outbound network requests. It also recommends descriptive column headers in the first row and one row per record. A support schedule laid out that way gives every assistant a better chance of reading the amounts correctly.

··········

WORKED EXAMPLE: SIX OCTOBER 2026 ADJUSTING JOURNAL ENTRIES.

The dataset is illustrative. It describes a small US company closing October 2026 on a monthly basis, with six adjustments that cover the most common entry patterns. Each row holds everything an assistant needs to draft the entry, and nothing it would have to guess.

........

Ref	Item	Source facts (illustrative)
JE-01	Utilities accrual	October bill not received; estimate = average of Jul $3,960, Aug $4,180, Sep $4,420, rounded to the dollar
JE-02	Prepaid insurance	12-month policy paid July 1, 2026 for $18,600, straight-line
JE-03	Depreciation	Equipment cost $84,000, salvage $6,000, 5-year straight-line life
JE-04	Deferred revenue	$27,000 annual subscription billed September 1, 2026, earned evenly over 12 months
JE-05	Interest accrual	$250,000 note at 6.6% a year, 30/360 day count, interest paid quarterly
JE-06	Reclassification	$3,480 software subscription coded to office supplies expense

........

The prompt should fix the output format, ask for the calculation behind each amount and forbid the assistant from inventing missing facts. A version that works with any of the three tools, after uploading or opening the schedule, reads as follows.

You are preparing October 2026 month-end adjusting journal entries for review.
Use only the facts in the table "Adjustments". Do not assume any fact that is not in the table.
For each row, return lines with these columns: Ref, Date (2026-10-31), Account, Debit, Credit, Description, Calculation.
Show the arithmetic for every amount in the Calculation column. Round to whole dollars only where the row says so.
Each Ref must balance. After the lines, show total debits and total credits.
List any row where you needed information that was not provided, instead of guessing.
Flag which entries should be set to auto-reverse on 2026-11-01.

The last two instructions do most of the control work. Asking for a list of missing information gives the model a legitimate place to put its uncertainty, and asking which entries should reverse forces a decision that is easy to forget when entries are drafted in bulk. In this dataset the utilities accrual is the clear reversing candidate, because the actual invoice will be booked in November; whether the interest accrual reverses depends on how the company records the quarterly payment.

··········

EXPECTED JOURNAL ENTRIES FOR THE OCTOBER CLOSE EXAMPLE.

The amounts below were computed independently in Python from the source facts, so they serve as the answer key for the AI draft. The utilities estimate is $12,560 divided by 3, which is $4,186.67 and rounds to $4,187. The insurance charge is $18,600 over 12 months, the depreciation is $78,000 of depreciable cost over 60 months, the subscription revenue is $27,000 over 12 months, and the interest is $250,000 multiplied by 6.6% and divided by 12.

........

Ref	Account	Debit	Credit
JE-01	Utilities expense	4,187
JE-01	Accrued liabilities		4,187
JE-02	Insurance expense	1,550
JE-02	Prepaid insurance		1,550
JE-03	Depreciation expense	1,300
JE-03	Accumulated depreciation		1,300
JE-04	Deferred revenue	2,250
JE-04	Subscription revenue		2,250
JE-05	Interest expense	1,375
JE-05	Accrued interest payable		1,375
JE-06	Software subscriptions expense	3,480
JE-06	Office supplies expense		3,480
Total		14,142	14,142

........

Two balance checks confirm the schedules behind the entries. After the October charge, four of the twelve insurance months have been used, so prepaid insurance should stand at $12,400 on October 31, 2026. Two of the twelve subscription months have been earned since September 1, so deferred revenue for that contract should stand at $22,500. If the ledger balances after posting do not match these figures, either an earlier month was missed or the draft used the wrong number of periods.

The interest entry shows why the day-count fact belongs in the prompt. Under the 30/360 convention stated in the illustrative note, October interest is $1,375. If the same note used actual days over a 365-day year, the 31 days of October would give $1,401.37. An assistant given only the principal and the rate has to pick one, and a $26 difference on one entry is easy to miss in review unless the convention is stated and checked.

The accrual pattern behind JE-01 is explained in more depth in our article on accrued expenses, their accounting treatment and journal entries.

··········

SIX CHECKS BEFORE POSTING AN AI-DRAFTED JOURNAL ENTRY.

The first check is balance by entry as well as in total. A draft can show equal total debits and credits while two entries are each out of balance by offsetting amounts. In a sheet where column A holds the Ref, column D the debit and column E the credit, a helper column with a standard formula flags any reference that does not net to zero.

=ROUND(SUMIFS($D:$D,$A:$A,A2)-SUMIFS($E:$E,$A:$A,A2),2)

Any result other than zero points to the entry that needs attention, and a final cell with =SUM(D:D)-SUM(E:E) confirms the batch as a whole.

The second check is direction. Deferred revenue and reclassifications are the entries where models most often reverse the sides, because both involve debiting an account that usually carries a credit balance or crediting an expense. In JE-04 the liability is debited and revenue is credited; in JE-06 the correct expense is debited and the wrong one is credited.

The third check is recomputation. Every amount in the draft should be recalculated from the source schedule outside the assistant, with a formula in the workbook or a separate calculation, and compared with the answer key. Asking the assistant to show its arithmetic in a Calculation column makes this quick, because an error in the method is visible next to the amount.

The fourth check is the period count. Amortization and revenue entries go wrong when a model divides the remaining balance by the remaining months instead of the original amount by the original term. For straight-line schedules the two methods happen to give the same answer, which hides the error until a schedule with a mid-term change produces a different figure. The test is to compare the post-entry balance with the expected balance, as done above for prepaid insurance and deferred revenue.

The fifth check is unsupported assumptions. Any fact the assistant introduced that was not in the source (a day-count convention, a useful life, a reversal date, an account that does not exist in the chart of accounts) should be removed or confirmed with documentation before posting. Entries should also be mapped to the company's actual account codes, since an assistant will happily invent a reasonable-sounding account name.

The sixth check is approval and evidence. The draft, the support schedule, the reviewer's sign-off and any changes made after review should be retained together, under the same segregation of duties that applies to manually prepared entries. Microsoft's FAQ notes that saved Copilot changes are visible to anyone with access to the file and can be reverted by undoing them or restoring prior versions, so version history can serve as part of the trail for workbooks edited in Excel. The same checks fit into the wider sequence described in our month-end close checklist with Copilot in Excel, Claude and ChatGPT.

··········

DATA PRIVACY SETTINGS FOR LEDGER DATA IN COPILOT, CLAUDE AND CHATGPT.

Ledger detail, vendor names and payroll-related accruals are confidential, so the account type matters as much as the tool. Microsoft states that Copilot processes prompts and responses within the Microsoft 365 service boundary with enterprise data protection, honors existing permissions and sensitivity labels, and does not use prompts, responses or documents to train the underlying foundation models. Copilot in Excel cannot be turned off separately from Copilot Chat, according to the FAQ; it falls under the administrator settings for Microsoft Copilot Chat.

Anthropic's help center describes file creation as running in an isolated sandbox that is never shared between users, with network egress on by default for Free, Pro, Max and Ultra and restricted by default for Team and Enterprise organizations. It also warns that hidden instructions in external files or websites could lead Claude to run untrusted code or send information from its context to third parties, and recommends monitoring the work and stopping it if Claude accesses data unexpectedly. Finance teams that upload ledger extracts should keep network access restricted for that work.

OpenAI's help center states that its business offerings, including ChatGPT Enterprise, are not used to train its models, and that users of personal plans can turn off model training in their settings. For any of the three tools, an export that removes employee names and bank details before upload reduces the exposure without affecting the entry logic.

··········

AI-DRAFTED JOURNAL ENTRIES: WHAT THE TOOLS DOCUMENT AND WHAT STAYS WITH THE CONTROLLER.

The vendors' documentation supports a clear division of work. Copilot in Excel, Claude and ChatGPT can read a support schedule, apply standard debit and credit patterns and return balanced, described entry lines in a fixed layout, and Microsoft's newer accounting capabilities add input and balance checks inside Excel for organizations enrolled in Frontier. Microsoft's documentation tells users to review and verify Copilot output and names finance decisions as an area where the tool should not be relied on without verification, and Anthropic's help center recommends monitoring Claude while it works on files and stopping it if it accesses data unexpectedly.

The October example shows where review effort pays off: entry-level balance, debit and credit direction on deferred revenue and reclassifications, recomputation of every amount, period counts on amortization, and facts such as day-count conventions that change the number without changing the pattern. A reviewer with the $14,142 answer key, the expected closing balances of $12,400 and $22,500, and the SUMIFS balance column can confirm or reject an AI draft in minutes. Estimates, cut-off decisions, reversal settings and the final approval remain the accountant's responsibility, and the evidence retained for them should be the same as for any manual entry.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org