This blueprint is one of the workflows that most obviously benefits from Claude Managed Agents — the new hosted agent runtime Anthropic launched on April 8, 2026. Long-running, stateful, multi-tool, credential-sensitive: it’s a textbook match.
A no-code, step-by-step workflow for corporate finance and accounting teams who want to use AI to compress their month-end close — covering ERP data export, account reconciliation, variance analysis, narrative generation, and audit-ready documentation. Built for NetSuite, SAP, Oracle, and QuickBooks environments. No developers required.
Summary Card
What it does: Automates the most time-consuming parts of the month-end close — account reconciliation, variance flagging, management narrative drafting, and audit documentation — using AI tools you can operate without any technical background
Who it’s for: Controllers, staff accountants, FP&A analysts, and CFOs at companies using NetSuite, SAP, Oracle, or QuickBooks who want to close faster and with more consistent documentation
Time to implement: 45–90 minutes for the manual workflow; 3–5 hours for a semi-automated setup with Zapier or Make
Tools required: Your existing ERP (NetSuite, SAP, Oracle, or QuickBooks) + ChatGPT Enterprise or Claude.ai Teams/Pro + optional Zapier or Make for automation
Cost estimate: $20–30/month (Claude Pro or ChatGPT Plus) for the manual workflow; $0 additional if your company already has ChatGPT Enterprise or Microsoft Copilot
Difficulty: Beginner — no code, no API keys, no IT department required for the core workflow
GAAP context: Designed for US GAAP reporting environments; includes SEC-relevant variance documentation guidance
Last tested: March 2026 with Claude Sonnet 4, ChatGPT-4o, NetSuite, and QuickBooks Online
What the research says
A joint MIT/Stanford study found AI assistance can reduce monthly financial close time by up to 7.5 days. Gartner research cited by CFO Dive projects that AI-equipped ERP environments could cut close times by 30% by 2028. The bottleneck is not the technology — it is knowing exactly how to structure the workflow. That is what this blueprint covers.
The month-end close is one of the most predictable, repeatable workflows in corporate finance — and yet most teams still run it the same way they did ten years ago. Trial balance pulled on day one. Reconciliations worked manually through day four. Variance comments written from scratch on day six. Audit workpapers assembled on day seven. By day ten, the controller signs off and everyone exhales.
The repetition is the opportunity. Because the steps are the same every month, AI is exceptionally good at handling the structured parts: spotting what changed, explaining why it matters, and drafting the documentation that takes so long to write from a blank page.
This blueprint shows you exactly how to do it — using tools you can access today, without writing a single line of code, and with specific guidance on how to handle financial data confidentiality at every step.
Before You Start: How to Handle Financial Data Confidentiality
This section is not optional reading. Financial data is among the most sensitive information your organization handles, and the decision of which AI tool you use should be made before you export a single row from your ERP.
The core question is simple: does the AI provider you are using have access to the data you paste into it, and can they use it for model training? Expand the section that matches your company type below.
Select Your Company Type — Expand for Recommendations
▶ Public Company (SEC Filer)Highest Caution
Recommended tools: ChatGPT Enterprise, Claude for Enterprise (API with enterprise agreement), or Microsoft Copilot for Finance (via M365 E5 + Copilot add-on). Azure OpenAI Service if your IT team has it deployed.
Why: Pre-announcement financial data is material non-public information (MNPI). Your legal and compliance team should review AI tool data processing agreements before any financial data is used. Enterprise tiers explicitly state data is not used for training and is not stored beyond the session.
Safe shortcut: Anonymize account names in your export before uploading — replace “Revenue – SaaS Subscriptions” with “Account A.” The AI does not need to know the account names to do reconciliation math. See the data preparation step below.
▶ Private CompanyStandard Caution
Recommended tools: Claude.ai Pro or Teams plan, or ChatGPT Team or Enterprise. Both explicitly exclude uploaded files from training data on paid plans.
Why: Private company financials are confidential but not subject to SEC insider trading rules. The main risk is business confidentiality. Enterprise/Teams tiers address this.
Check this: In Claude.ai, go to Settings → Privacy → ensure “Improve Claude for everyone” is OFF. In ChatGPT, go to Settings → Data Controls → turn OFF “Improve the model for everyone.”
▶ Small Business / StartupStandard Paid Plan
Recommended tools: Claude Pro ($20/month) or ChatGPT Plus ($20/month). On paid plans, your conversations are not used to train the model by default.
Good practice: Still avoid pasting your exact customer names or SSNs. Use account category names rather than full client names in any data you upload.
Free tier warning: Do not use the free tier of any consumer AI tool for financial data. Free tiers typically use your conversations for model training.
What This AI Workflow Actually Does — And What It Does Not Do
Let me be precise about where AI helps and where humans must stay in the loop. This is not about replacing your accounting team. It is about removing the parts of close that are time-consuming but not judgment-intensive — the math-checking, the first-draft writing, the documentation formatting — so that your team can spend their time on the decisions that actually require accounting expertise.
AI handles well ✓
Comparing this month’s trial balance to last month’s and flagging what changed by more than a threshold you set
Identifying accounts where the balance moved in an unexpected direction
Matching intercompany entries to confirm they net to zero
Drafting the management narrative explaining variances in plain English
Producing a structured reconciliation summary with account-by-account commentary
Generating a checklist of items requiring human sign-off, organized by priority
Formatting output in US GAAP-consistent language for audit workpapers
Human must own ✗
Final review and sign-off on all reconciliations
Judgment calls on accruals, estimates, and reserves
Any disclosure decisions (especially for public companies)
Audit trail documentation — AI can draft it, but a human must certify it
Interpretation of unusual transactions (M&A activity, one-time items, restatements)
Journal entry approval and posting
The Full Workflow Architecture
Here is how the six-step workflow fits together from ERP export to signed-off documentation.
Workflow Overview
1
ERP Export
Trial balance + transaction detail as CSV/Excel
→
2
Data Prep
Clean, format, anonymize if needed
→
3
Reconciliation
AI flags variances and mismatches
→
4
Variance Analysis
Compare to prior period and budget
→
5
Narrative Draft
AI writes management commentary
→
6
Audit Docs
Human reviews + signs off
You can run this entirely manually — exporting from your ERP and uploading to an AI chat interface — or you can automate the data movement using Zapier or Make once you have the manual version working. I strongly recommend starting manually. Once you understand exactly what the AI is doing with your data and what outputs you want, automation becomes a simple add-on, not a prerequisite.
Step 1: Export Your Trial Balance and Transaction Detail from Your ERP
You need two exports for this workflow: your trial balance (account-level summary) and your transaction detail for any accounts you want to reconcile at the line level. Expand your ERP below for the exact export path.
Expand Your ERP System for Export Instructions
▶ NetSuite
Trial Balance
Navigate to Reports → Financial → Trial Balance
Set your date range to the period you are closing (e.g., March 1–31)
Optional: set the subsidiary filter if you run consolidated reporting
Click Customize and ensure Account Number, Account Name, Beginning Balance, Debit, Credit, and Ending Balance columns are all visible
Click Export → CSV (top right of the report)
Transaction Detail (General Ledger)
Navigate to Reports → Financial → General Ledger
Set the same date range and filters
Export to CSV — this gives you every journal entry line for the period
▶ SAP
Trial Balance
Transaction code S_ALR_87012301 (Balance Sheet / P&L Statement) — or use the Fiori app “Display Financial Statement”
Set fiscal year, posting period, and company code
Execute the report, then use List → Export → Spreadsheet to download as XLSX
Transaction Detail
Transaction code FAGLL03 (G/L Account Line Item Display) — or Fiori “Display Line Items in G/L Accounts”
Enter your account range, company code, and posting date range
Execute and export to Excel via Ctrl+Shift+F7 or the spreadsheet export button
▶ Oracle Fusion
Trial Balance
Navigate to General Ledger → Reports and Analytics → Financial Reporting Center
Select Trial Balance Report from the standard report library
Set your ledger, period, and currency parameters
Click Submit, then open the completed report and use Actions → Export → Excel
Journal Details
Navigate to General Ledger → Journals → Manage Journals
Filter by period and status (Posted), then use the Export option to download to XLSX
▶ QuickBooks
QuickBooks Online
Navigate to Reports → All Reports → Trial Balance
Set your report period using the date range dropdowns at the top
Click Export → Export to Excel (or CSV) in the top-right corner
QuickBooks Desktop
Go to Reports → Accountant & Taxes → Trial Balance
Set the date range and click Excel → Create New Worksheet
For transaction detail, use Reports → Accountant & Taxes → General Ledger
What if my ERP exports are very large?
AI tools like Claude and ChatGPT can handle files up to 25–50MB. If your trial balance is larger, split it by account category — run Balance Sheet accounts first, then Income Statement. You can combine the summary outputs at the narrative stage. Always upload the file using the attachment icon — do not paste raw data directly into the chat.
Step 2: Prepare Your Data Before Uploading to the AI
The quality of your AI output is directly proportional to how cleanly your input is structured. Five minutes of preparation here saves you from confusing outputs later.
Data Preparation Checklist
Include column headers on row 1. Your spreadsheet must have clear column names — Account Number, Account Name, Beginning Balance, Debit, Credit, Ending Balance. The AI uses these to understand what each column means.
Include prior period data for comparison. Add a column for “Prior Period Ending Balance” (last month or same month last year). Without this, the AI cannot do meaningful variance analysis.
Include account type classification. Add a column indicating whether each account is an Asset, Liability, Equity, Revenue, or Expense. If your ERP export does not include this automatically, you can add it with a VLOOKUP or manually for your key accounts.
Remove rows with zero activity if your file is large. Filter out accounts with $0.00 beginning balance, $0 debits, $0 credits, and $0 ending balance. They add noise without adding information.
For public companies or sensitive environments: anonymize. Replace specific account names with generic labels if needed (“Revenue – Product A” → “Revenue Account 1”). Maintain a separate mapping key for your own reference. The AI only needs numbers and categories.
Note your materiality threshold. Decide what dollar amount or percentage change qualifies as “significant” for your organization — $5,000? 10%? You will include this in your prompt so the AI knows what to flag.
Save as CSV or XLSX. Both work. CSV is preferred for large files. Name the file clearly: “TrialBalance_March2026.csv”
Step 3: Run the Account Reconciliation Prompt
Now you are ready to upload your prepared trial balance to Claude or ChatGPT and run the reconciliation prompt. Open a new conversation, attach your CSV or XLSX file using the paperclip/attachment icon, and paste the following prompt. Customize the two bracketed fields before sending.
Prompt — Account Reconciliation Analysis
You are a senior accountant reviewing a month-end trial balance for a US GAAP-compliant company. I have uploaded a trial balance CSV with the following columns: Account Number, Account Name, Account Type, Prior Period Ending Balance, Current Period Debit Activity, Current Period Credit Activity, Current Period Ending Balance.
Please perform the following analysis:
1. BALANCE SHEET RECONCILIATION
- For each asset, liability, and equity account, calculate the change from prior period to current period (in dollars and as a percentage)
- Flag any account where the change exceeds [INSERT YOUR MATERIALITY THRESHOLD — e.g., $10,000 or 10%]
- For flagged accounts, note whether the movement direction is expected (e.g., a receivable account increasing is expected if revenue is growing; a reversal would be flagged as unusual)
2. INCOME STATEMENT CHECK
- Identify the top 5 largest revenue variances from prior period
- Identify the top 5 largest expense variances from prior period
- Flag any expense account that exceeded its balance in the same period last year by more than [INSERT THRESHOLD]
3. INTERCOMPANY / ELIMINATIONS CHECK (if applicable)
- Identify any account number containing "interco" or "elimination" in the name
- Confirm whether corresponding debit and credit entries net to zero
- List any intercompany accounts that do not balance
4. UNUSUAL ITEMS
- Flag any account balance that is directionally inconsistent with US GAAP presentation (e.g., a negative cash balance, a credit balance in an asset account, a debit balance in a liability account) as these require human review
Output Format:
- Section 1: Table of flagged Balance Sheet accounts with: Account Name, Prior Balance, Current Balance, $ Change, % Change, Flag Reason
- Section 2: Top revenue and expense variances table
- Section 3: Intercompany reconciliation summary
- Section 4: Unusual items list with brief explanation of why each is flagged
- Section 5: Summary — total accounts reviewed, number flagged, estimated time to resolve flagged items
Write in the tone of a reconciliation workpaper that a CPA would review during audit fieldwork. For every number you cite, include the Account Number in brackets so I can verify it against my source data.
Customize before running
Replace [INSERT YOUR MATERIALITY THRESHOLD] with your threshold. A common starting point for mid-market companies: flag any account where the change is greater than $25,000 or greater than 15%, whichever is lower.
If you have no intercompany activity, delete Section 3 from the prompt entirely.
Step 4: Run the Variance Analysis Against Budget or Prior Year
The reconciliation step tells you what changed. The variance analysis step tells you whether it was expected — and drafts the explanation for each significant item. For this step, you will need a second file: either your budget or your prior year actuals for the same period. If you have both, run the prompt twice.
Prompt — Budget vs. Actual Variance Analysis
I am attaching two files: File 1 is the Actual Trial Balance for [MONTH] [YEAR] and File 2 is the Budget for [MONTH] [YEAR]. Both files share the same column structure: Account Number, Account Name, Account Type, and Balance.
Please perform a Budget vs. Actual variance analysis with the following requirements:
1. REVENUE VARIANCE ANALYSIS
- For each revenue account, calculate: Actual, Budget, $ Variance (Actual minus Budget), % Variance
- Classify each variance as Favorable (F) or Unfavorable (U) from a P&L perspective
- Flag variances greater than [INSERT THRESHOLD, e.g., $50,000 or 10%]
- For each flagged revenue account, draft a one-sentence possible explanation in plain English
2. EXPENSE VARIANCE ANALYSIS
- Repeat the same analysis for all expense accounts
- Flag any expense account where actuals exceeded budget by more than [THRESHOLD]
- For each flagged expense, draft a one-sentence possible explanation
3. GROSS MARGIN CHECK
- Calculate total actual revenue, total budgeted revenue, and the variance
- Calculate actual gross margin % and budgeted gross margin %
- Flag if actual gross margin % differs from budget by more than 2 percentage points
4. OPERATING EXPENSE CHECK
- Calculate total actual OpEx vs. budget
- Identify which three expense categories have the largest unfavorable variances
5. NARRATIVE DRAFT
Based on all of the above, draft a 3–4 sentence management commentary paragraph that could appear in a monthly reporting package. Write it in third-person, past tense, US GAAP-appropriate language.
Important notes:
- I will review and edit all variance explanations — treat these as first drafts, not final commentary
- If you cannot determine a plausible cause from the data alone, write "Requires management review — no data-supported explanation available" rather than speculating
- Present all figures in thousands ($000s) with one decimal place
Human-in-the-loop checkpoint — do not skip this
The AI will draft variance explanations based on patterns in the data, but it does not know about the business events behind the numbers. Before moving to Step 5, review each flagged variance and add a note about the actual business driver where you know it. This takes 10–15 minutes and produces dramatically better narratives. The AI writes the draft; you add the business context.
Step 5: Generate the Management Narrative
Once you have reviewed and annotated the variance analysis output, you are ready to generate the full management narrative. Continue in the same conversation — the AI already has the context from the previous steps. Paste your business context notes into the bracketed section below.
Prompt — Management Narrative Generation
Using the variance analysis we just completed, please draft the full management narrative for [COMPANY NAME]'s [MONTH] [YEAR] financial close package. Here is the additional business context you should incorporate:
[PASTE YOUR NOTES HERE — e.g.: "The revenue increase was driven by the XYZ contract signed March 15. The engineering overage was due to three new hires starting mid-month. The accounts receivable increase reflects the XYZ invoice which is not yet due."]
Please produce the narrative in the following sections:
SECTION 1: EXECUTIVE SUMMARY (2–3 sentences)
High-level summary of the month — was it above or below expectations, and what were the one or two biggest drivers?
SECTION 2: REVENUE (1 paragraph)
Explain total revenue vs. budget and vs. prior year. Quantify the variance. Identify the primary business drivers. Note any revenue recognition items that affected the period.
SECTION 3: COST OF GOODS SOLD AND GROSS MARGIN (1 paragraph)
Explain gross margin performance. Quantify any changes in COGS. Note significant cost drivers.
SECTION 4: OPERATING EXPENSES (1 paragraph)
Walk through major expense categories with variances above threshold. Note any one-time or non-recurring expenses that should be excluded from run-rate analysis.
SECTION 5: BALANCE SHEET HIGHLIGHTS (1 paragraph)
Comment on cash position and movement, accounts receivable aging if relevant, and any significant balance sheet changes.
SECTION 6: ITEMS REQUIRING MANAGEMENT DECISION
Bulleted list of items from the close that require a management decision or follow-up action.
Formatting requirements:
- Third-person, past tense throughout
- US GAAP-appropriate language
- Dollar amounts in thousands ($000s) with one decimal
- Percentages to one decimal place
- Avoid the word "significant" — use specific dollar and percentage figures instead
- Mark with [REVIEW NEEDED] anywhere you have inserted a placeholder or are unsure of the business driver
- Plain text only — no markdown, no asterisks, no pound signs. Use ALL CAPS for section headers.
Step 6: Generate Audit-Ready Workpapers and Sign-Off Documentation
The final step turns your AI-assisted analysis into documentation that can withstand audit scrutiny. For public companies this means PCAOB-compatible workpapers; for private companies it means clean, organized documentation that supports a smooth year-end audit. Every reconciliation needs a human name and date on it.
Prompt — Audit Workpaper Generation
Using the reconciliation analysis and variance commentary from this conversation, please produce audit-ready close documentation for [COMPANY NAME] for the period ending [DATE], prepared under US GAAP.
DOCUMENT 1: CLOSE PACKAGE COVER PAGE
- Company name, reporting period, date prepared
- Preparer name: [INSERT NAME] | Title: [INSERT TITLE]
- Reviewer name: [INSERT NAME — leave blank if not yet reviewed]
- List of all documents included in the close package
- Status of each document: Complete, In Review, or Open Item
DOCUMENT 2: TRIAL BALANCE CERTIFICATION
- Confirmation that the trial balance ties to the general ledger
- Total debits and total credits (confirm they are equal)
- Any out-of-balance items and their resolution status
- Prepared by / Date / Reviewed by / Date fields (for human completion)
DOCUMENT 3: RECONCILIATION SUMMARY
For each flagged account from the reconciliation analysis:
- Account name and number
- Prior period balance
- Current period balance
- Explanation of the change
- Status: Reconciled, Unreconciled, or Requires Follow-Up
- Reconciled by / Date field (blank for human completion)
DOCUMENT 4: OPEN ITEMS LOG
- List all items flagged as [REVIEW NEEDED] or unresolved
- For each item: Description, Dollar Amount, Responsible Party (blank), Due Date (blank), Status
DOCUMENT 5: MANAGEMENT SIGN-OFF CHECKLIST
Produce a checklist of all standard US GAAP close procedures:
- Trial balance reviewed and approved
- All intercompany eliminations confirmed
- Revenue recognition reviewed per ASC 606
- Accruals and prepaid expenses reviewed
- Fixed assets roll-forward reconciled
- Bank reconciliations completed and reviewed
- Payroll accrual confirmed
- Debt covenants reviewed (if applicable)
- Tax provision reviewed (if applicable)
- All open items from prior period cleared or escalated
- Management narrative reviewed and approved
- Financial statements tied to trial balance
Format all of the above as clean plain text that can be pasted into a Word document. Use ALL CAPS for document titles and section headers. Include [SIGNATURE LINE] placeholders wherever a human sign-off is required. Plain text only — no markdown, no asterisks.
Public company note (SEC filers)
AI-generated workpapers must be reviewed, edited, and signed off by a CPA. The AI output is a draft framework, not a completed workpaper. A cover note stating “Initial draft prepared with AI assistance; reviewed and approved by [Name, CPA] on [Date]” is good practice and consistent with how firms are handling AI-assisted work in 2026.
Automating the Data Movement: Zapier and Make Options
Once you have run the manual workflow for two or three months and know exactly what inputs and outputs work for your organization, you can automate the data transfer steps using Zapier or Make. Expand each step below to see what is realistically automatable without any coding — and what must stay manual.
Automation Decision Matrix — Click Each Step to Expand
▶ ERP data exportPartial
NetSuite: NetSuite has a native Saved Search export that Zapier can trigger via the NetSuite connector — but setup requires your NetSuite admin to create the saved search and grant the Zapier integration read-access. If your admin can do this, you can schedule a monthly data pull automatically.
SAP / Oracle: These ERPs typically require IT involvement to set up scheduled exports. The simplest no-code path is to manually export your trial balance CSV each period and drop it into a designated Google Drive or SharePoint folder — then Zapier or Make can detect the new file and trigger the rest of the workflow.
QuickBooks: QuickBooks has a direct Zapier connector. You can trigger a trial balance export on a schedule using the “New Report” Zapier action (available on QuickBooks Online Advanced plan).
▶ Data preparation and cleaningKeep Manual
Automating data preparation requires knowing in advance exactly what cleaning steps are needed every month. In practice there are almost always month-specific quirks — a new account that needs classification, a format change in the ERP export, an unusual transaction that needs separate handling.
Keep this step manual for at least the first six months. Once your format is completely stable month over month, you can use a Make or Zapier “Formatter” step to apply standard cleaning rules automatically.
▶ Sending data to AI and running promptsFully Automatable
Both Zapier and Make have Claude and ChatGPT connectors. Once your CSV is in Google Drive or Dropbox, you can build a workflow that: detects the new file in your designated folder → reads the file content → sends it to Claude or ChatGPT with your saved prompt → captures the AI response → saves the response to a Google Doc → sends an email notification.
Zapier path: New File in Google Drive → Read File → Claude: Send Message (with file content + saved prompt) → Create Google Doc with output → Email notification
Make path: Watch Google Drive Files → Read File → Anthropic or OpenAI module → Create Google Doc → Gmail notification. Setup time: approximately 2–3 hours for someone comfortable with Zapier or Make. No coding required.
▶ Variance review and narrative editingMust Stay Human
This is the most important human step in the entire workflow. The AI identifies what changed and offers draft explanations based on patterns — but it does not know about the sales contract you signed, the vendor dispute that caused a delayed payment, or the strategic decision to accelerate hiring.
There is no automation shortcut here. This step is where your expertise creates value — and where the AI’s output goes from “plausible draft” to “accurate and defensible.”
▶ Distributing the close packageFully Automatable
Once the final narrative and workpapers are saved to a shared Google Drive or SharePoint folder, Zapier or Make can automatically send an email notification to your distribution list with a link to the close package.
Zapier path: File Updated in Google Drive → Filter (only when file name contains “FINAL”) → Send email via Gmail or Outlook with sharing link. This removes the manual “please find attached” email that typically ends every close cycle.
What Breaks and How to Fix It
I tested this workflow with multiple trial balance formats across two AI tools. Here are the failure modes you will encounter and how to resolve them.
▶ The AI cites a balance that doesn’t match my file
What happens: In rare cases the AI will cite a balance that differs slightly from what is in your file — usually when accounts have very similar names.
Fix: Add this line to the end of your reconciliation prompt: “For every number you cite in your output, include the Account Number in brackets so I can verify it against my source data.” This forces specificity and makes spot-checking fast.
▶ My file is too large to upload
What happens: Claude and ChatGPT both have file size limits (typically around 25MB). Very large trial balances hit this.
Fix: Split by financial statement section — Balance Sheet first (assets, liabilities, equity), then Income Statement (revenue, COGS, opex). Run the prompts separately and combine both outputs at the narrative stage.
▶ The variance explanations are too generic to be useful
What happens: The AI produces explanations like “the increase may be due to increased business activity” — which is useless.
Fix: This always happens when you skip the human annotation step in Step 4. Go back, add your business context notes (even just brief bullet points), and re-run the narrative prompt with that context included. The specificity of the output scales directly with the specificity of your context notes.
▶ The output looks messy when pasted into Word
What happens: The AI output uses markdown formatting (asterisks, hashes) that appears as symbols when pasted into Word.
Fix: The prompts above already include the plain-text instruction. If you write a custom prompt, always add: “Plain text only — no markdown, no asterisks, no pound signs. Use ALL CAPS for section headers and standard hyphens for bullet points.”
▶ The AI misclassifies an account type
What happens: If your trial balance does not have an Account Type column, the AI infers types from account names — and occasionally gets it wrong (e.g., classifying “Deferred Revenue” as a revenue account instead of a liability).
Fix: Always include an Account Type column in your export. It takes two minutes to add and prevents this entire class of error.
US GAAP and SEC Context: What to Keep in Mind
If your company files with the SEC — or is preparing for an IPO or debt offering — there are additional considerations for using AI in the close process.
Disclosure controls (SOX Section 302/906): Using AI to assist with close documentation does not impair your CEO/CFO certifications — but the human review chain must be clearly documented. Every AI-assisted workpaper should show a human reviewer’s name and date.
Management’s Discussion & Analysis (MD&A): The variance narrative from Step 5 is a strong starting point for MD&A drafting. However, MD&A language in your 10-Q/10-K is subject to SEC review and materiality standards. Have your legal/IR team review any AI-drafted language before it goes into a public filing.
Revenue recognition (ASC 606): The prompts in this blueprint do not perform revenue recognition analysis — they flag variances for human review. ASC 606 judgments (performance obligations, variable consideration, contract modifications) must be made by your accounting team.
Non-GAAP measures: If your reporting includes adjusted EBITDA or other non-GAAP measures, add a prompt step specifically for the non-GAAP reconciliation. Include: “This is a non-GAAP measure. All adjustments must be labeled as such and individually described per SEC Regulation G requirements.”
My Notes After Testing This Workflow
I tested this workflow against a real 120-account trial balance export from QuickBooks Online and a 340-account export in NetSuite format. Here is what I observed.
The reconciliation step worked better than expected. Claude Sonnet 4 was able to correctly identify directional inconsistencies — a credit balance in an asset account, a debit balance in a liability account — without any explicit accounting rules in the prompt. When I added the account type column to my export, the accuracy of variance flagging was essentially perfect on the accounts I spot-checked.
The variance narrative was the highest-leverage step. Without business context notes, the AI produced technically accurate but completely generic commentary. With even three or four bullet points of business context, it produced commentary that was 80% final-ready on the first pass. This step compressed what typically takes 90 minutes of writing into about 20 minutes of editing.
The workpaper formatting required the most cleanup. The AI-produced workpaper structure was logically sound but needed formatting work when pasted into Word. The plain text instruction in the prompts above helps significantly. I recommend pasting the output into Google Docs first and then converting to your preferred format.
Where I hit a wall: Complex intercompany reconciliations with more than four entities required manual review even after the AI flagged the relevant accounts. The AI correctly identified which intercompany pairs did not net to zero, but the investigation of why required tracing back to specific journal entries inside the ERP. The AI gets you to the right question; you still have to find the answer.
Tested: March 2026 using Claude Sonnet 4 (claude.ai Teams) and ChatGPT-4o (ChatGPT Team plan). QuickBooks Online Advanced and NetSuite OneWorld formats. Materiality threshold set at $10,000 or 5%, whichever was lower.
Related Blueprints
Blueprint
AI Agent for DPIA Workflow
A similar audit-documentation approach applied to GDPR/CCPA privacy risk assessments. Useful if your finance team also handles data privacy compliance.
Blueprint
No-Code Compliance Agent with Microsoft Copilot
If your company runs on Microsoft 365, this companion blueprint covers the same document-review workflow using tools already in your enterprise account.
Stack Guide
n8n vs. Make vs. Zapier for AI Agents
A full comparison of the three automation platforms if you decide to build the automated version of this workflow — with pricing, pros and cons, and the specific use cases each handles best.
Ahmad works in AI agent operations at G42. Before that, he spent 17 years working in communications and software development. He builds and maintains AI workflows in production daily and writes the blueprints he wishes someone had given him when he started. Every guide on this site is AI-assisted and human-tested. All articles reflect his personal opinions and thoughts and do not necessarily represent those of his employer (G42).