Implementation Guide: Produce financial narratives and management commentary for monthly close packages

Hardware Procurement

Software Procurement

Prerequisites

• Active QuickBooks Online (Plus or Advanced), Xero (Standard or Premium), or Sage Intacct subscription with admin-level API access credentials
• Minimum 12 months of historical financial data loaded in the accounting platform for meaningful comparative and variance analysis
• Budget and/or forecast data loaded into the accounting platform or available in Excel format for variance commentary generation
• Clean, well-structured Chart of Accounts with consistent naming conventions and proper account categorization (revenue, COGS, operating expenses, etc.)
• Microsoft 365 Business Standard or higher licenses assigned to all users who will use Copilot for Finance (minimum 2–3 seats for typical small firm)
• Microsoft Entra ID (Azure AD) tenant configured with user accounts for SSO across all SaaS platforms
• Stable business internet connection with minimum 25 Mbps download speed
• Modern web browser (Chrome 120+, Edge 120+, or Safari 17+) on all workstations
• Designated 'AI Narrative Review Lead' — a licensed CPA or senior accountant who will own the human review and approval workflow
• Written engagement letter language approved by the firm's managing partner addressing AI tool usage in client deliverables
• Credit card or payment method on file for OpenAI API account (only if implementing custom n8n pipeline)
• SharePoint Online or OneDrive for Business document library created for storing generated narrative reports and maintaining version history

Installation Steps

Step 1: Establish Fathom Account and Connect Accounting Platform

Create the Fathom organization account, configure SSO, and connect the firm's accounting platform (QuickBooks Online or Xero). This establishes the data pipeline that feeds the AI narrative engine. Select the Fathom Pro plan which supports up to 20 company connections and includes the Commentary Writer feature at no additional cost.

Step 2: Configure Chart of Accounts Mapping and KPI Definitions in Fathom

Map the client's chart of accounts to Fathom's standardized financial categories. Define non-financial KPIs that should be referenced in management commentary (headcount, units sold, etc.). This step is critical because the AI Commentary Writer uses these mappings to generate accurate, contextual narratives.

Step 3: Enable and Configure Commentary Writer AI Feature

Activate Fathom's Commentary Writer within the report builder. Configure narrative preferences including tone, detail level, and specific metrics to highlight. Generate initial test narratives to establish a quality baseline.

Step 4: Deploy Microsoft 365 Copilot for Finance Licenses

Assign Microsoft 365 Copilot licenses to the designated finance users via the Microsoft 365 Admin Center. This enables in-Excel AI-powered variance analysis and narrative drafting that complements Fathom's report-level commentary with granular spreadsheet-level analysis.

Step 5: Configure SharePoint Document Library for Narrative Storage and Version Control

Create a structured SharePoint Online document library to store AI-generated narratives, maintain version history, and support the review/approval workflow. This provides the audit trail required for compliance.

Step 6: Create OpenAI API Account and Configure API Keys

Set up the OpenAI platform account that will power the custom n8n narrative pipeline. Configure API keys with appropriate usage limits and security controls. This step is only required if implementing the advanced custom pipeline (Steps 10-12).

Step 7: Deploy n8n Workflow Orchestration Platform

Deploy n8n as the workflow automation backbone that connects the accounting platform API, OpenAI API, and output delivery channels. Choose between n8n Cloud (recommended for simplicity) or self-hosted Community Edition (for maximum data control).

# OPTION A: n8n Cloud (Recommended)
# Navigate to https://app.n8n.cloud/register
# Select Pro plan ($50/month, 10,000 executions)
# Complete registration and access the workflow editor

# OPTION B: Self-Hosted via Docker (For data sovereignty requirements)
# On a Linux VM (Ubuntu 22.04 LTS, 4 vCPU, 8GB RAM):
sudo apt update && sudo apt install -y docker.io docker-compose
mkdir -p /opt/n8n && cd /opt/n8n

cat > docker-compose.yml << 'EOF'
version: '3.8'
services:
  n8n:
    image: n8nio/n8n:latest
    restart: always
    ports:
      - '5678:5678'
    environment:
      - N8N_BASIC_AUTH_ACTIVE=true
      - N8N_BASIC_AUTH_USER=admin
      - N8N_BASIC_AUTH_PASSWORD=CHANGE_THIS_STRONG_PASSWORD
      - N8N_ENCRYPTION_KEY=GENERATE_A_RANDOM_32_CHAR_KEY
      - WEBHOOK_URL=https://n8n.yourmspdomain.com/
      - DB_TYPE=postgresdb
      - DB_POSTGRESDB_HOST=postgres
      - DB_POSTGRESDB_PORT=5432
      - DB_POSTGRESDB_DATABASE=n8n
      - DB_POSTGRESDB_USER=n8n
      - DB_POSTGRESDB_PASSWORD=CHANGE_THIS_DB_PASSWORD
    volumes:
      - n8n_data:/home/node/.n8n
    depends_on:
      - postgres
  postgres:
    image: postgres:16
    restart: always
    environment:
      - POSTGRES_USER=n8n
      - POSTGRES_PASSWORD=CHANGE_THIS_DB_PASSWORD
      - POSTGRES_DB=n8n
    volumes:
      - postgres_data:/var/lib/postgresql/data
volumes:
  n8n_data:
  postgres_data:
EOF

docker-compose up -d

# Verify n8n is running:
curl -s http://localhost:5678/healthz

Step 8: Configure Identity, SSO, and Access Controls

Set up Single Sign-On using Microsoft Entra ID across all platforms (Fathom, n8n, ChatGPT Team). Configure role-based access controls to enforce the principle of least privilege. This is essential for compliance with GLBA Safeguards Rule and firm data security policies.

Step 9: Build the Review and Approval Workflow

Establish the human-in-the-loop review and approval process that ensures every AI-generated narrative is verified by a qualified professional before inclusion in client deliverables. This is the most critical compliance step — AI-generated content in financial close packages MUST be reviewed by a CPA or senior accountant.

Step 10: Build Custom n8n Narrative Generation Workflow

Create the advanced n8n workflow that pulls financial data from the accounting platform API, applies structured prompt templates, sends data to GPT-5.4 Mini for narrative generation, and routes the output for review. This provides maximum customization beyond what Fathom's built-in Commentary Writer offers.

Step 11: Configure Prompt Templates and Narrative Styles

Deploy the carefully engineered prompt templates that guide the LLM to produce accurate, professional financial narratives. These templates are the core intellectual property of the implementation and should be version-controlled.

Step 12: Pilot Testing with 2-3 Client Companies

Run the AI narrative generation system in parallel with the firm's existing manual narrative process for 2-3 selected pilot clients. Compare AI output quality, accuracy, and completeness against manually written commentary. This dual-run approach validates the system before full rollout.

Step 13: Production Rollout and Monthly Close Process Integration

Expand the AI narrative system to the full client base. Integrate narrative generation into the firm's existing monthly close workflow and checklist. Establish the production cadence and standard operating procedures.

Step 14: Compliance Documentation and Audit Trail Setup

Finalize all compliance documentation including the AI governance policy, updated engagement letters, data processing agreements, and audit trail procedures. This step ensures the firm meets AICPA, SOX, GLBA, and applicable privacy law requirements.

Custom AI Components

Financial Narrative System Prompt

Type: prompt

The master system prompt that instructs GPT-5.4 Mini to generate professional financial narratives from structured accounting data. This prompt enforces accuracy, appropriate tone, materiality thresholds, and accounting-specific language conventions. It is the core intellectual property of the custom narrative pipeline.

Implementation:

You are a senior financial analyst and CPA preparing management commentary for monthly close packages. You generate professional, accurate financial narratives based exclusively on the structured financial data provided to you.

CRITICAL RULES

OUTPUT FORMAT

Return the narrative in clean markdown with the following sections:

• ### Executive Summary
• ### Revenue & Gross Margin Analysis
• ### Operating Expense Analysis
• ### Balance Sheet Highlights
• ### Cash Flow & Liquidity
• ### Items Requiring Management Attention

Do not include any preamble, disclaimers about being an AI, or meta-commentary. Output only the narrative text.

Financial Data Transformation Function

Type: integration n8n Code node that transforms raw QuickBooks Online API JSON responses into a structured, LLM-friendly text format. This function extracts key financial metrics, calculates variances against prior period and budget, and assembles a data summary that serves as the user prompt for the narrative LLM.

Implementation:

// n8n Code Node: Transform QBO Financial Data for LLM Narrative Generation
// Input: Items from QBO P&L and Balance Sheet HTTP Request nodes
// Output: Structured text summary for LLM prompt

const pnlData = $('Fetch QBO PnL').item.json;
const bsData = $('Fetch QBO Balance Sheet').item.json;

// Helper function to extract row value from QBO report JSON
function getRowValue(report, rowName) {
  if (!report || !report.Rows || !report.Rows.Row) return null;
  for (const row of report.Rows.Row) {
    if (row.Summary && row.Summary.ColData) {
      if (row.Summary.ColData[0].value === rowName) {
        return row.Summary.ColData.map((col, idx) => idx === 0 ? col.value : parseFloat(col.value) || 0);
      }
    }
    if (row.Rows && row.Rows.Row) {
      for (const subRow of row.Rows.Row) {
        if (subRow.ColData && subRow.ColData[0].value === rowName) {
          return subRow.ColData.map((col, idx) => idx === 0 ? col.value : parseFloat(col.value) || 0);
        }
      }
    }
  }
  return null;
}

// Extract P&L columns (Current Month, Prior Month, Budget if available)
const columns = pnlData.Columns.Column.map(c => c.ColTitle);

// Build structured summary sections
let sections = [];
sections.push('# FINANCIAL DATA FOR NARRATIVE GENERATION');
sections.push(`## Report Period: ${pnlData.Header.StartPeriod} to ${pnlData.Header.EndPeriod}`);
sections.push(`## Company: ${pnlData.Header.ReportName}`);
sections.push(`## Columns: ${columns.join(' | ')}`);
sections.push('');

// Extract all P&L rows recursively
function extractRows(rows, indent = 0) {
  let results = [];
  if (!rows || !rows.Row) return results;
  for (const row of rows.Row) {
    if (row.ColData) {
      const label = row.ColData[0].value;
      const values = row.ColData.slice(1).map(c => c.value || '0');
      results.push(`${'  '.repeat(indent)}${label}: ${values.join(' | ')}`);
    }
    if (row.Header && row.Header.ColData) {
      results.push(`${'  '.repeat(indent)}**${row.Header.ColData[0].value}**`);
    }
    if (row.Rows) {
      results = results.concat(extractRows(row.Rows, indent + 1));
    }
    if (row.Summary && row.Summary.ColData) {
      const label = row.Summary.ColData[0].value;
      const values = row.Summary.ColData.slice(1).map(c => c.value || '0');
      results.push(`${'  '.repeat(indent)}**${label}: ${values.join(' | ')}**`);
    }
  }
  return results;
}

sections.push('## PROFIT & LOSS STATEMENT');
sections = sections.concat(extractRows(pnlData.Rows));
sections.push('');
sections.push('## BALANCE SHEET');
sections = sections.concat(extractRows(bsData.Rows));
sections.push('');
sections.push('## INSTRUCTIONS');
sections.push('Generate a management commentary narrative for this financial data following your system prompt guidelines.');
sections.push('Compare current period to prior period. Flag all material variances (>5% AND >$1,000).');
sections.push('If budget column is present, also compare actuals to budget.');

return [{ json: { prompt_text: sections.join('\n'), generated_at: new Date().toISOString() } }];

Monthly Narrative Generation Workflow

Type: workflow Complete n8n workflow definition that orchestrates the end-to-end monthly narrative generation process: scheduled trigger → data fetch from QBO → data transformation → LLM narrative generation → quality checks → SharePoint storage → Teams notification. This is exported as n8n JSON for direct import.

Implementation:

{
  "name": "Monthly Financial Narrative Generator",
  "nodes": [
    {
      "parameters": { "rule": { "interval": [{ "field": "cronExpression", "expression": "0 9 5 * *" }] } },
      "name": "Monthly Schedule Trigger",
      "type": "n8n-nodes-base.scheduleTrigger",
      "position": [250, 300]
    },
    {
      "parameters": { "url": "https://quickbooks.api.intuit.com/v3/company/{{$json.companyId}}/reports/ProfitAndLoss", "authentication": "oAuth2", "method": "GET", "queryParameters": { "parameters": [{ "name": "start_date", "value": "={{$now.minus({months:1}).startOf('month').toFormat('yyyy-MM-dd')}}" }, { "name": "end_date", "value": "={{$now.minus({months:1}).endOf('month').toFormat('yyyy-MM-dd')}}" }, { "name": "accounting_method", "value": "Accrual" }] } },
      "name": "Fetch QBO PnL",
      "type": "n8n-nodes-base.httpRequest",
      "position": [470, 300]
    },
    {
      "parameters": { "url": "https://quickbooks.api.intuit.com/v3/company/{{$json.companyId}}/reports/BalanceSheet", "authentication": "oAuth2", "method": "GET", "queryParameters": { "parameters": [{ "name": "as_of", "value": "={{$now.minus({months:1}).endOf('month').toFormat('yyyy-MM-dd')}}" }] } },
      "name": "Fetch QBO Balance Sheet",
      "type": "n8n-nodes-base.httpRequest",
      "position": [470, 500]
    },
    {
      "parameters": { "jsCode": "// See Financial Data Transformation Function component for full code" },
      "name": "Transform Financial Data",
      "type": "n8n-nodes-base.code",
      "position": [690, 400]
    },
    {
      "parameters": { "model": "gpt-5.4-mini", "messages": { "values": [{ "role": "system", "content": "={{ $('Load System Prompt').item.json.prompt }}" }, { "role": "user", "content": "={{ $json.prompt_text }}" }] }, "options": { "temperature": 0.3, "maxTokens": 2500 } },
      "name": "Generate Narrative via GPT-5.4 Mini",
      "type": "@n8n/n8n-nodes-langchain.openAi",
      "position": [910, 400]
    },
    {
      "parameters": { "jsCode": "// Post-processing: validate no hallucinated numbers\nconst narrative = $input.first().json.message.content;\nconst sourceData = $('Transform Financial Data').first().json.prompt_text;\n\n// Extract all dollar amounts from narrative\nconst narrativeNumbers = narrative.match(/\\$[\\d,]+\\.?\\d*/g) || [];\n\n// Flag any numbers not found in source data\nlet warnings = [];\nfor (const num of narrativeNumbers) {\n  const cleanNum = num.replace(/[$,]/g, '');\n  if (!sourceData.includes(cleanNum) && parseFloat(cleanNum) > 100) {\n    warnings.push(`WARNING: ${num} cited in narrative but not found in source data`);\n  }\n}\n\nreturn [{ json: { narrative, warnings, has_warnings: warnings.length > 0, generated_at: new Date().toISOString() } }];" },
      "name": "Validate Narrative Accuracy",
      "type": "n8n-nodes-base.code",
      "position": [1130, 400]
    },
    {
      "parameters": { "resource": "file", "operation": "upload", "siteId": "YOUR_SHAREPOINT_SITE_ID", "folderId": "YOUR_DRAFT_FOLDER_ID", "fileName": "={{$now.minus({months:1}).toFormat('yyyy-MM')}}-narrative-draft.md", "fileContent": "={{ $json.narrative }}" },
      "name": "Save to SharePoint",
      "type": "n8n-nodes-base.microsoftSharePoint",
      "position": [1350, 400]
    },
    {
      "parameters": { "resource": "chatMessage", "operation": "create", "teamId": "YOUR_TEAMS_TEAM_ID", "channelId": "YOUR_CHANNEL_ID", "message": "=📊 **AI Narrative Draft Ready for Review**\n\nPeriod: {{$now.minus({months:1}).toFormat('MMMM yyyy')}}\nWarnings: {{$json.has_warnings ? $json.warnings.join('; ') : 'None — all figures verified'}}\n\nPlease review and approve in SharePoint." },
      "name": "Notify Reviewer via Teams",
      "type": "n8n-nodes-base.microsoftTeams",
      "position": [1570, 400]
    }
  ],
  "connections": {
    "Monthly Schedule Trigger": { "main": [[{ "node": "Fetch QBO PnL" }, { "node": "Fetch QBO Balance Sheet" }]] },
    "Fetch QBO PnL": { "main": [[{ "node": "Transform Financial Data" }]] },
    "Fetch QBO Balance Sheet": { "main": [[{ "node": "Transform Financial Data" }]] },
    "Transform Financial Data": { "main": [[{ "node": "Generate Narrative via GPT-5.4 Mini" }]] },
    "Generate Narrative via GPT-5.4 Mini": { "main": [[{ "node": "Validate Narrative Accuracy" }]] },
    "Validate Narrative Accuracy": { "main": [[{ "node": "Save to SharePoint" }]] },
    "Save to SharePoint": { "main": [[{ "node": "Notify Reviewer via Teams" }]] }
  }
}

AI Governance Policy Template

Type: prompt A comprehensive AI governance policy document template tailored for accounting firms deploying AI-generated financial narratives. This document satisfies AICPA, GLBA, and SOX compliance requirements and should be customized and signed by the firm's managing partner.

Implementation:

# AI GOVERNANCE POLICY — [FIRM NAME]

Effective Date: [DATE] | Version: 1.0 | Approved by: [MANAGING PARTNER NAME]

1. PURPOSE

This policy governs the use of artificial intelligence tools in the preparation of client financial deliverables at [Firm Name]. It ensures all AI-assisted work products meet professional standards, protect client confidentiality, and comply with applicable regulations.

2. SCOPE

This policy applies to all firm personnel, contractors, and managed service providers who use AI tools in connection with client engagements, specifically:

• AI-generated financial narratives and management commentary
• AI-assisted variance analysis and trend identification
• AI-powered report generation and formatting

3. APPROVED AI TOOLS

The following AI tools are approved for use with client financial data:

4. HUMAN REVIEW MANDATE

All AI-generated content intended for client deliverables MUST be reviewed and approved by a qualified professional (CPA, CMA, or senior accountant with 3+ years of experience) before delivery. AI output is a DRAFT until human-reviewed. Specific requirements:

• Reviewer must verify all cited figures against source GL data
• Reviewer must assess narrative accuracy, completeness, and appropriateness of conclusions
• Reviewer must document their review via the SharePoint approval workflow
• Reviewer's name and date must be recorded in the engagement file

5. DATA HANDLING

• Client financial data may only be processed through approved tools with executed DPAs
• No client data may be stored on personal devices or non-approved cloud services
• All AI-processed data must comply with GLBA Safeguards Rule requirements
• For EU-resident client data, GDPR data residency requirements must be verified

6. AUDIT TRAIL

The following records must be maintained for every AI-generated narrative:

• Input data snapshot (what financial data was provided to the AI)
• AI tool and model version used
• Complete AI output (unedited)
• Reviewer identity and approval/rejection decision
• Final edited version delivered to client
• Retention period: 7 years minimum

7. CLIENT DISCLOSURE

[Firm Name] will include the following language in all engagement letters: 'Our firm utilizes AI-assisted tools to support the preparation of financial analyses and management commentary. All AI-generated content is reviewed and approved by qualified professionals before delivery. Your data is processed only through enterprise-grade platforms with appropriate security certifications and data processing agreements.'

8. INCIDENT RESPONSE

If an AI tool produces materially incorrect content that is delivered to a client:

9. REVIEW CYCLE

This policy will be reviewed quarterly and updated as needed. Next review date: [DATE + 90 DAYS].

Signature: _________________________ Date: _____________ [MANAGING PARTNER NAME], [FIRM NAME]

Narrative Quality Scoring Agent

Type: agent

A quality assurance agent that evaluates AI-generated financial narratives against a scoring rubric before they reach the human reviewer. It catches common LLM errors like hallucinated numbers, incorrect percentage calculations, missing material variances, and inappropriate forward-looking statements. Runs as a second LLM call in the n8n pipeline.

Implementation:

// n8n Code Node: Quality Scoring Agent
// This node takes the generated narrative and source data, sends them to a second LLM call
// for quality assurance scoring, and flags issues for the human reviewer.

const narrative = $('Generate Narrative via GPT-5.4 Mini').first().json.message.content;
const sourceData = $('Transform Financial Data').first().json.prompt_text;

// QA System Prompt
const qaSystemPrompt = `You are a financial narrative quality assurance auditor. You receive two inputs:
1. SOURCE DATA: The raw financial data that was provided to the narrative generator
2. GENERATED NARRATIVE: The AI-generated management commentary

Your job is to score the narrative on the following rubric and flag any issues:

SCORING RUBRIC (1-5 scale for each)

OUTPUT FORMAT

Return ONLY a JSON object (no markdown fences):

{
  "scores": {
    "numerical_accuracy": <1-5>,
    "completeness": <1-5>,
    "professional_tone": <1-5>,
    "objectivity": <1-5>,
    "structure": <1-5>
  },
  "overall_score": <average of above, rounded to 1 decimal>,
  "pass": <true if overall >= 3.5, false otherwise>,
  "issues": [
    {"severity": "critical|warning|info", "description": "<specific issue description>", "location": "<which section>"}
  ],
  "suggested_edits": ["<specific suggested edit 1>", "<specific suggested edit 2>"]
}

const qaUserPrompt = `## SOURCE DATA\n${sourceData}\n\n## GENERATED NARRATIVE\n${narrative}`;

// Return the QA prompt for the next OpenAI node
return [{
  json: {
    qa_system_prompt: qaSystemPrompt,
    qa_user_prompt: qaUserPrompt,
    original_narrative: narrative
  }
}];

// After this Code node, add another OpenAI Chat node:
// Model: gpt-5.4-mini
// System: {{ $json.qa_system_prompt }}
// User: {{ $json.qa_user_prompt }}
// Temperature: 0.1 (maximum consistency for QA)
// Max Tokens: 1000
//
// Then add a Code node to parse the QA JSON response:
// const qaResult = JSON.parse($input.first().json.message.content);
// const narrative = $('Quality Scoring Agent').first().json.original_narrative;
// if (!qaResult.pass) {
//   // Route to 'Needs Revision' path — regenerate with issue context
//   return [{ json: { ...qaResult, narrative, action: 'revise' } }];
// }
// return [{ json: { ...qaResult, narrative, action: 'approve' } }];

Testing & Validation

• NUMERICAL ACCURACY TEST: Generate a narrative for a pilot client. Independently calculate the top 5 variance percentages from the raw GL data using a calculator or Excel. Compare each percentage cited in the AI narrative — all must match within 0.1 percentage points. If any figure is wrong, check the data transformation function and prompt instructions.
• COMPLETENESS TEST: Review the source P&L for the pilot month. Identify all line items with variances exceeding 5% AND $1,000 (the defined materiality threshold). Verify that every material item appears in the AI narrative. Count: AI should discuss at least 90% of material variances.
• HALLUCINATION TEST: Search the AI-generated narrative for every dollar amount cited. Cross-reference each amount against the raw QBO/Xero data export. Flag any amount that cannot be traced to the source data. Zero hallucinated numbers is the passing threshold.
• TONE AND COMPLIANCE TEST: Have a CPA read the narrative and flag any instances of: (a) forward-looking statements or predictions, (b) investment advice, (c) casual or unprofessional language, (d) AI self-references ('As an AI...'), (e) unsupported opinions about business strategy. The narrative must contain zero instances of any of these.
• INTEGRATION TEST — FATHOM: In Fathom, disconnect and reconnect a test company. Generate Commentary Writer output. Verify that the narrative accurately reflects the most recent month's data and that prior period comparisons use the correct time frame. Click through every cited figure to verify it traces back to the GL.
• INTEGRATION TEST — QBO API: In n8n, manually trigger the workflow for a test company. Verify the HTTP Request nodes return 200 status codes. Verify the P&L date range parameters correctly target the prior month. Verify the Balance Sheet as-of date is the last day of the prior month.
• INTEGRATION TEST — SHAREPOINT: Verify that generated narrative documents appear in the correct SharePoint folder path (/Clients/{ClientName}/{Year}/{Month}/Draft-Narratives/). Verify version history is enabled and captures each upload. Verify the Power Automate approval flow triggers automatically when a new file is created.
• INTEGRATION TEST — TEAMS NOTIFICATION: After a narrative is saved to SharePoint, verify that a notification appears in the #narrative-review Teams channel within 5 minutes. Verify the notification includes the correct client name, period, and a clickable link to the SharePoint document.
• QUALITY SCORING AGENT TEST: Feed the QA agent a narrative with an intentionally wrong percentage (e.g., change a 12% increase to 22%). Verify the agent catches it and returns a 'critical' severity issue with numerical_accuracy score below 3. Then feed it a correct narrative and verify it passes with overall score >= 3.5.
• END-TO-END TIMING TEST: Measure the complete cycle time from workflow trigger to Teams notification for 3 different clients. The entire automated pipeline should complete in under 5 minutes per client. Fathom Commentary Writer should generate in under 30 seconds.
• PARALLEL COMPARISON TEST: For each pilot client, generate the AI narrative AND have the firm produce their traditional manual narrative. Present both (anonymized) to the lead CPA and ask them to rate quality on a 1-10 scale. The AI narrative should score within 2 points of the manual narrative to pass.
• SECURITY TEST: Verify that the OpenAI API key is not visible in any n8n workflow export, log file, or SharePoint document. Verify it is stored only in n8n's encrypted credentials store or Azure Key Vault. Attempt to access the n8n instance without SSO credentials — access should be denied.

Client Handoff

The client handoff meeting should be a 90-minute session with the firm's managing partner, senior accountants (narrative reviewers), and any staff who will interact with the AI tools. Cover the following topics in order:

Maintenance

Monthly Maintenance Tasks (MSP Responsibility):

• Review n8n workflow execution logs for errors or anomalies. Check that all scheduled runs completed successfully. Investigate and resolve any failed executions within 4 business hours during close periods (1st-10th of each month).
• Monitor OpenAI API usage and costs via the OpenAI dashboard. Verify costs are within the expected $20-70/month range for a 10-client practice. Alert the client if costs exceed 150% of baseline.
• Verify Fathom data sync status for all connected companies. Re-authenticate OAuth connections if any have expired (QBO tokens expire every 100 days).
• Export and archive n8n execution logs to SharePoint for long-term compliance retention (n8n Cloud only retains 30 days by default).

Quarterly Maintenance Tasks (MSP Responsibility):

• Prompt Optimization Review: Sample 5-10 generated narratives from the quarter. Assess quality trends. Refine prompt templates based on recurring reviewer edits. Document all prompt changes with version numbers and rationale.
• Compliance Audit: Sample 3-5 random client narratives and verify: (a) documented human review approval exists, (b) AI engine identification metadata is present, (c) SharePoint version history is intact, (d) all vendor DPAs are current. Produce a 1-page quarterly compliance summary.
• Software Update Review: Check for Fathom feature updates, n8n version updates, and OpenAI model changes. Test any updates in a non-production environment before deploying. GPT-5.4 Mini model updates may change output characteristics — always re-test prompts after model updates.
• Client Satisfaction Check: Brief survey or conversation with the lead reviewer — are narratives meeting quality expectations? Are there new report types or narrative sections they'd like to add?

Semi-Annual Tasks:

• AI Governance Policy Review: Update the policy document to reflect any new tools, regulatory changes, or lessons learned. Obtain managing partner re-approval if material changes are made.
• Engagement Letter Audit: Verify all active client engagement letters include the AI usage disclosure addendum.
• Security Review: Rotate API keys (OpenAI, QBO OAuth). Review access permissions across all platforms. Remove access for any departed staff. Verify MFA enforcement.

Model Retraining / Update Triggers:

• OpenAI announces a new GPT-5.4 Mini version → test all prompt templates against new model, compare output quality, update n8n model parameter if quality is maintained or improved.
• Client adds a new service line or significant chart of accounts changes → update KPI definitions and prompt context.
• Fathom releases Commentary Writer updates → test new features, update SOPs and training materials.
• Regulatory changes (new AICPA guidance on AI, PCAOB requirements) → update governance policy and compliance procedures.

SLA Considerations:

• During monthly close period (business days 1-10): 4-hour response time for critical issues (workflow failures, data sync errors), 8-hour response for non-critical.
• Outside close period: Next business day response.
• Severity 1 (narrative pipeline completely down): 2-hour response, 8-hour resolution target. Fallback: manual narrative process.
• Severity 2 (quality degradation, partial functionality): 4-hour response, 24-hour resolution.
• Severity 3 (cosmetic issues, feature requests): Next business day acknowledgment, scheduled for next quarterly optimization cycle.

Escalation Path: