Accounting Agent Skill

Version: 2.0 Date: 2026-02-17

Agent Definition

Field	Value
Role	Financial Intelligence Specialist
Goal	Receive financial documents from the Assignment Agent, perform detailed financial analysis (price verification, date validation, currency checks), standardise the data into a structured format, and route the results — either to the Admin Agent for expense registration or to the Data Analysis Agent for further reporting.
Backstory	You are a senior financial analyst with deep expertise in travel-industry accounting. You receive invoices, receipts, and confirmations in multiple formats (PDF, DOCX, XLSX, images) and multiple languages (Thai, Chinese, English). You extract every financial detail with precision, cross-check totals against line items, flag discrepancies, translate descriptions into English, and produce a clean, standardised JSON record. You never submit data downstream without verifying its integrity first.

Core Workflow (4 Steps)

code

┌──────────────────────────────────────────────────────────────────┐
│  Step 1: RECEIVE                                                 │
│  ← Assignment Agent sends file (PDF/DOCX/XLSX/image) or data    │
└──────────────────┬───────────────────────────────────────────────┘
                   ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 2: ANALYSE                                                 │
│  Extract raw text → validate prices, dates, currencies, totals   │
│  Detect anomalies (mismatched totals, missing fields, bad dates) │
└──────────────────┬───────────────────────────────────────────────┘
                   ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 3: DEFINE                                                  │
│  Map extracted data to Standardised Invoice JSON Schema           │
│  Translate descriptions → English                                │
│  Classify expense type and assign routing tag                    │
└──────────────────┬───────────────────────────────────────────────┘
                   ▼
┌──────────────────────────────────────────────────────────────────┐
│  Step 4: ROUTE                                                   │
│  ┌─ expense_register  → Admin Agent (charge record creation)     │
│  └─ analysis_only     → Data Analysis Agent (reporting/BI)       │
└──────────────────────────────────────────────────────────────────┘

Tools

Tool Class	Description
Step 1 — Receive
`ReceiveDocumentTool`	Accept a file path or binary payload from the Assignment Agent. Detect file type (PDF, DOCX, XLSX, PNG/JPG).
Step 2 — Analyse
`ExtractTextTool`	Extract raw text from the document. Uses `pdfplumber` for PDF, `python-docx` for DOCX, `openpyxl` for XLSX, and `pytesseract` OCR for images.
`FinancialAnalysisTool`	Use the LLM to parse extracted text and perform financial checks: verify line-item totals match the grand total, validate date ranges, confirm currency codes, detect duplicate charges, and flag anomalies.
`CurrencyValidationTool`	Validate currency codes against ISO 4217, check exchange rates against reference rates, and flag suspicious conversions.
Step 3 — Define
`StandardiseInvoiceTool`	Map extracted data to the Standardised Invoice JSON Schema (see schema below). Fill all required fields, set missing fields to defaults.
`TranslateFieldsTool`	Use the LLM to translate all descriptive text fields (descriptions, notes, activity details) into English while preserving the JSON structure.
`ClassifyExpenseTool`	Classify the document type and determine the routing tag: `expense_register` (needs to be recorded on the website) or `analysis_only` (informational, for reporting).
Step 4 — Route
`RouteToAdminTool`	Package the standardised JSON and send it to the Admin Agent for expense registration on `/charges_group/create`.
`RouteToAnalysisTool`	Package the standardised JSON and send it to the Data Analysis Agent for inclusion in booking/financial reports.
`SaveFinancialRecordTool`	Persist the analysis results and standardised data to `data/financial_records.json`.

Step-by-Step Detail

Step 1: Receive

The Accounting Agent is activated when the Assignment Agent delegates a financial task. Input arrives as:

Input Type	Description	Example
File upload	A document attached via LINE message, forwarded by the Assignment Agent	Invoice PDF, confirmation DOCX, receivables XLSX, receipt image
Structured data	A JSON object from another agent or from a CSV import	`{ "tour_code": "...", "amount": ..., "pax": ... }`
Text message	A natural-language financial instruction from the user	"Register expense 5000 THB for tour JAPAN7N-001"

Actions:

•Detect the input type (file vs. structured data vs. text)
•If file → determine format (PDF/DOCX/XLSX/image) and queue for extraction
•If structured data → skip extraction, proceed directly to analysis
•If text → use LLM to parse intent and extract parameters

Step 2: Analyse (Financial Analysis)

This is the intelligence core. The agent examines every financial detail in the document.

2.1 Text Extraction

File Format	Extraction Method	Library
PDF (text-based)	Direct text extraction	`pdfplumber`
PDF (image-based)	OCR → text	`pytesseract`
DOCX	Paragraph + table extraction	`python-docx`
XLSX	Cell-by-cell extraction with formula resolution	`openpyxl`
PNG / JPG	OCR → text	`pytesseract`

2.2 Financial Validation Checks

Check	Description	Flag Level
Line-item total	Sum of `quantity × unit_price` for each item must equal `total_price`	ERROR if mismatch > 1%
Grand total	Sum of all line-item totals must equal `subtotal`; `subtotal + tax = grand_total`	ERROR if mismatch
Date validity	All dates must be parseable and logically ordered (issue_date ≤ start_date ≤ end_date)	WARNING if illogical
Date format	Convert all dates to `YYYY-MM-DD` regardless of source format	AUTO-FIX
Currency code	Must be a valid ISO 4217 code (THB, RMB, USD, EUR, CNY, etc.)	ERROR if unrecognised
Exchange rate	If multi-currency, exchange rate must be > 0 and within reasonable bounds	WARNING if suspicious
PAX count	`adults + children + escorts = total_pax`	WARNING if mismatch
Duplicate detection	Check if same `tour_code + amount + date` exists in `data/financial_records.json`	WARNING if duplicate found
Required fields	`tour_code` or `tour_name`, `grand_total`, `currency` must be present	ERROR if missing
Payment deadline	Flag if payment deadline is in the past or within 24 hours	WARNING

2.3 Analysis Output

json

{
  "analysis_status": "passed | warnings | errors",
  "total_checks": "number",
  "passed_checks": "number",
  "warnings": [
    {
      "check": "string",
      "field": "string",
      "message": "string",
      "expected": "string",
      "actual": "string"
    }
  ],
  "errors": [
    {
      "check": "string",
      "field": "string",
      "message": "string",
      "expected": "string",
      "actual": "string"
    }
  ]
}

Step 3: Define (Standardise & Classify)

3.1 Standardised Invoice JSON Schema

The agent maps all extracted data into this unified schema, designed for travel-industry invoices:

json

{
  "document_info": {
    "document_title": "string",
    "document_type": "Confirmation | Invoice | Quote | Summary Statement | Receipt",
    "issue_date": "YYYY-MM-DD",
    "internal_ref_id": "string",
    "external_ref_id": "string"
  },
  "supplier": {
    "company_name": "string",
    "contact_person": "string",
    "phone": "string",
    "email": "string",
    "address": "string"
  },
  "customer": {
    "company_name": "string",
    "contact_person": "string",
    "phone": "string",
    "email": "string",
    "address": "string",
    "source_market": "string"
  },
  "tour_details": {
    "tour_name": "string",
    "tour_code": "string",
    "start_date": "YYYY-MM-DD",
    "end_date": "YYYY-MM-DD",
    "duration_days": "number",
    "participants": {
      "adults": "number",
      "children": "number",
      "tour_escorts": "number",
      "total_pax": "number"
    },
    "service_standards": {
      "accommodation_standard": "string",
      "meal_plan": "string",
      "transportation_mode": "string",
      "guide_notes": "string",
      "insurance_notes": "string"
    },
    "itinerary": [
      {
        "day_number": "number",
        "date": "YYYY-MM-DD",
        "transportation": "string",
        "activities": "string (translated to English)",
        "meals": {
          "breakfast": "boolean",
          "lunch": "boolean",
          "dinner": "boolean"
        },
        "accommodation": "string"
      }
    ]
  },
  "financials": {
    "currency": "string (ISO 4217)",
    "line_items": [
      {
        "item_id": "string",
        "description": "string (translated to English)",
        "item_type": "Flight | Accommodation | Tour Fee | Visa | Tip | Allowance | Taxi | Adjustment",
        "quantity": "number",
        "unit_price": "number",
        "total_price": "number",
        "currency": "string",
        "notes": "string"
      }
    ],
    "subtotal": "number",
    "tax_amount": "number",
    "grand_total": "number",
    "total_in_words": "string",
    "payment_summary": {
      "amount_receivable": "number",
      "amount_collected": "number",
      "amount_paid": "number",
      "amount_outstanding": "number"
    }
  },
  "payment_details": {
    "payment_deadline": "YYYY-MM-DD HH:MM:SS",
    "payment_terms": "string",
    "bank_accounts": [
      {
        "account_name": "string",
        "account_number": "string",
        "bank_name": "string",
        "bank_branch": "string",
        "currency": "string",
        "country": "string",
        "notes": "string"
      }
    ]
  },
  "confirmation_details": {
    "requires_signature": "boolean",
    "confirmation_status": "Pending | Confirmed",
    "confirmation_request_text": "string",
    "signatories": [
      {
        "party": "Supplier | Customer",
        "company_name": "string",
        "signer_name": "string",
        "date": "YYYY-MM-DD"
      }
    ]
  },
  "raw_text": "string (full extracted text for audit)"
}

3.2 Translation Rules

Source Language	Fields Translated	Target Language
Chinese (中文)	descriptions, activities, notes, payment terms	English
Thai (ไทย)	descriptions, activities, notes, payment terms	English
Mixed	All non-English descriptive fields	English

The JSON structure and field names remain unchanged — only the values of descriptive fields are translated.

3.3 Expense Classification & Routing Tag

Document Type	Contains Actionable Expense?	Routing Tag	Destination Agent
Invoice with specific tour charge	Yes — needs to be registered as expense on website	`expense_register`	Admin Agent
Receipt / proof of payment	Yes — may need recording	`expense_register`	Admin Agent
Confirmation (no payment yet)	No — informational only	`analysis_only`	Data Analysis Agent
Summary Statement / Receivables	No — for reporting	`analysis_only`	Data Analysis Agent
Quote / Estimate	No — for budgeting	`analysis_only`	Data Analysis Agent

The routing tag is embedded in the output:

json

{
  "routing": {
    "tag": "expense_register | analysis_only",
    "destination_agent": "admin | data_analysis",
    "reason": "string (why this routing was chosen)",
    "priority": "high | medium | low"
  }
}

Step 4: Route

Based on the routing tag, the agent delivers the standardised JSON to the appropriate downstream agent.

Route A: To Admin Agent (`expense_register`)

The Accounting Agent extracts the fields the Admin Agent needs for the charges form:

json

{
  "routing": { "tag": "expense_register", "destination_agent": "admin" },
  "expense_data": {
    "tour_code": "string (from tour_details.tour_code)",
    "program_code": "string (if identifiable)",
    "pax": "number (from tour_details.participants.total_pax)",
    "amount": "number (from financials.grand_total)",
    "description": "string (summary of line items)",
    "charge_type": "string (mapped from line_items.item_type)",
    "payment_date": "YYYY-MM-DD",
    "currency": "string",
    "exchange_rate": "number",
    "remark": "string (source document ref + analysis notes)"
  },
  "standardised_invoice": { "...full schema..." },
  "analysis_result": { "...validation checks..." }
}

Route B: To Data Analysis Agent (`analysis_only`)

json

{
  "routing": { "tag": "analysis_only", "destination_agent": "data_analysis" },
  "standardised_invoice": { "...full schema..." },
  "analysis_result": { "...validation checks..." },
  "financial_highlights": {
    "grand_total": "number",
    "currency": "string",
    "tour_code": "string",
    "supplier": "string",
    "customer": "string",
    "date_range": "YYYY-MM-DD to YYYY-MM-DD",
    "pax_count": "number"
  }
}

Both Routes: Save to Local Record

Regardless of routing, every processed document is persisted:

Output file: data/financial_records.json

Complete Output Schema

json

{
  "task_id": "string (UUID)",
  "source_file": "string (original filename)",
  "source_type": "pdf | docx | xlsx | image | structured_data | text",
  "processing_steps": {
    "received_at": "ISO-8601",
    "extraction_completed_at": "ISO-8601",
    "analysis_completed_at": "ISO-8601",
    "standardisation_completed_at": "ISO-8601",
    "routed_at": "ISO-8601"
  },
  "analysis_result": {
    "analysis_status": "passed | warnings | errors",
    "total_checks": "number",
    "passed_checks": "number",
    "warnings": [],
    "errors": []
  },
  "standardised_invoice": { "...full Standardised Invoice JSON Schema..." },
  "routing": {
    "tag": "expense_register | analysis_only",
    "destination_agent": "admin | data_analysis",
    "reason": "string",
    "priority": "high | medium | low"
  },
  "expense_data": { "...if routed to admin..." },
  "financial_highlights": { "...if routed to data_analysis..." },
  "timestamp": "ISO-8601"
}

Input

The Accounting Agent accepts tasks exclusively from the Assignment Agent. Input types:

Input Type	Format	Required Fields	Example
Invoice file	PDF, DOCX, XLSX, PNG, JPG	File path or binary	`/uploads/invoice_xinjiang_260118.pdf`
Structured data	JSON	`tour_code`, `amount` (minimum)	`{ "tour_code": "2UKWL3NN", "amount": 123120, "currency": "RMB" }`
Text instruction	Natural language	Must mention financial context	"Register 5000 THB expense for JAPAN7N-001"
CSV batch	CSV file	`tour_code`, `pax`, `amount` columns	`data/tour_charges.csv`

Execution Order

Layer 2 — Financial Intelligence (runs after Data Analysis Agent completes Layer 1)

Dependencies:

•Receives tasks from the Assignment Agent
•May receive supplementary booking data from the Data Analysis Agent for cross-referencing

Downstream consumers:

•Admin Agent — receives expense_register tagged records for website data entry
•Data Analysis Agent — receives analysis_only tagged records for reporting
•Executive Agent — consumes data/financial_records.json for financial summaries

Relationship to Other Agents

Agent	Relationship
Assignment Agent	Receives all financial tasks from Assignment Agent. Reports processing status back.
Admin Agent	Sends `expense_register` records with pre-filled expense data for web form entry. The Admin Agent handles the actual browser automation on `/charges_group/create`.
Data Analysis Agent	Sends `analysis_only` records for inclusion in booking and financial reports. May receive booking data for cross-referencing tour codes.
Market Analysis Agent	No direct interaction. Financial data flows to Market Analysis via the Executive Agent.
Executive Agent	Provides financial summaries, analysis results, and processed invoice counts for the executive report.

Configuration

Environment Variables

Variable	Description	Default
`OPENAI_API_KEY`	OpenAI API key for LLM extraction, analysis, and translation	Required
`OPENAI_MODEL`	LLM model for financial analysis	`gpt-4o-mini`
`TESSERACT_PATH`	Path to Tesseract OCR binary (for image processing)	System default
`FINANCIAL_RECORDS_PATH`	Path to persist processed records	`data/financial_records.json`
`UPLOAD_DIR`	Directory for received document files	`uploads/`
`TRANSLATION_TARGET`	Target language for translated fields	`en`

Agent-Level Configuration

python

ACCOUNTING_AGENT_CONFIG = {
    "enabled": True,
    "timeout": 300,
    "max_iterations": 20,
    "verbose": True,
    "llm_model": "gpt-4o-mini",
    "translation_enabled": True,
    "translation_target_language": "en",
    "validation_strict_mode": False,
    "duplicate_check_enabled": True,
    "max_file_size_mb": 25,
    "supported_formats": ["pdf", "docx", "xlsx", "png", "jpg", "jpeg", "csv"],
    "anomaly_threshold_percent": 1.0
}

Implementation Status

PARTIALLY IMPLEMENTED — Core extraction and standardisation logic exists in the auto-receipt project:

•backend/extractors.py — File text extraction (PDF, DOCX, XLSX)
•backend/llm_processor.py — OpenAI integration for standardisation
•backend/schemas.py — Pydantic models for the invoice schema
•backend/main.py — FastAPI endpoints for upload and processing

Still needed:

• FinancialAnalysisTool — price/date/currency validation checks
• CurrencyValidationTool — ISO 4217 validation and exchange-rate checking
• ClassifyExpenseTool — routing logic (expense_register vs. analysis_only)
• RouteToAdminTool — integration with Admin Agent
• RouteToAnalysisTool — integration with Data Analysis Agent
• SaveFinancialRecordTool — persistent record storage
• CrewAI agent definition in src/accounting_agent/agents.py
• CrewAI task definitions in src/accounting_agent/tasks.py
• Integration with Assignment Agent task delegation

Error Handling

•File format not supported: Return UNSUPPORTED_FORMAT error with the detected MIME type. Do not attempt processing.
•Text extraction failure: If pdfplumber fails, fall back to pytesseract OCR. If OCR also fails, return EXTRACTION_FAILED with the error detail.
•LLM parsing failure: If the LLM returns malformed JSON or fails to extract required fields, retry once with a more explicit prompt. If the second attempt fails, return STANDARDISATION_FAILED.
•Financial validation errors: Do not block routing on warnings. Only block on critical errors (missing grand_total, unrecognisable currency). Attach all warnings/errors to the output so downstream agents are aware.
•Duplicate detection: If a potential duplicate is found, attach a DUPLICATE_WARNING but still process and route. Let the destination agent decide whether to proceed.
•Translation failure: If translation fails for a field, keep the original text and add a TRANSLATION_FAILED flag for that field.
•Routing failure: If the destination agent is unavailable, save the record to data/financial_records.json with status PENDING_ROUTING and retry on the next cycle.
•File too large: If the file exceeds max_file_size_mb, return FILE_TOO_LARGE without processing.

MCP Integration

When exposed via the MCP server, the Accounting Agent provides:

MCP Tool Name	Description	Input Parameters	Output
`accounting_process_invoice`	Process a single invoice document through the full 4-step pipeline	`file_path`, `translation` (optional, default `true`)	Full output JSON with analysis, standardised data, and routing
`accounting_analyse_file`	Extract and analyse only (steps 1-2), without standardising or routing	`file_path`	Analysis result with extracted text and validation checks
`accounting_standardise`	Standardise pre-extracted text into the JSON schema (step 3 only)	`raw_text`, `source_type`	Standardised Invoice JSON
`accounting_check_status`	Check processing status of a previously submitted task	`task_id`	Status and result if completed

Testing Strategy

Unit Tests

python

def test_extract_text_pdf():
    """Verify ExtractTextTool handles text-based PDFs."""

def test_extract_text_docx():
    """Verify ExtractTextTool handles DOCX with tables."""

def test_extract_text_xlsx():
    """Verify ExtractTextTool handles multi-sheet XLSX."""

def test_extract_text_image_ocr():
    """Verify ExtractTextTool handles image-based documents via OCR."""

def test_financial_analysis_valid():
    """Verify FinancialAnalysisTool passes a clean invoice."""

def test_financial_analysis_mismatched_totals():
    """Verify FinancialAnalysisTool flags line-item vs. grand-total mismatch."""

def test_financial_analysis_invalid_dates():
    """Verify FinancialAnalysisTool flags illogical date ranges."""

def test_standardise_invoice():
    """Verify StandardiseInvoiceTool produces valid JSON matching the schema."""

def test_translate_fields():
    """Verify TranslateFieldsTool translates Chinese/Thai to English."""

def test_classify_expense_register():
    """Verify ClassifyExpenseTool tags invoices as expense_register."""

def test_classify_analysis_only():
    """Verify ClassifyExpenseTool tags confirmations as analysis_only."""

def test_route_to_admin():
    """Verify RouteToAdminTool packages expense_data correctly."""

def test_route_to_analysis():
    """Verify RouteToAnalysisTool packages financial_highlights correctly."""

Integration Tests

python

def test_full_pipeline_invoice_pdf():
    """End-to-end: PDF invoice → extract → analyse → standardise → route to admin."""

def test_full_pipeline_confirmation_docx():
    """End-to-end: DOCX confirmation → extract → analyse → standardise → route to data_analysis."""

def test_full_pipeline_receivables_xlsx():
    """End-to-end: XLSX receivables → extract → analyse → standardise → route to data_analysis."""

def test_batch_csv_processing():
    """Process CSV of tour charges → standardise each → route all to admin."""

Accounting Agent Skill

Agent Definition

Core Workflow (4 Steps)

Tools

Step-by-Step Detail

Step 1: Receive

Step 2: Analyse (Financial Analysis)

2.1 Text Extraction

2.2 Financial Validation Checks

2.3 Analysis Output

Step 3: Define (Standardise & Classify)

3.1 Standardised Invoice JSON Schema

3.2 Translation Rules

3.3 Expense Classification & Routing Tag

Step 4: Route

Route A: To Admin Agent (expense_register)

Route B: To Data Analysis Agent (analysis_only)

Both Routes: Save to Local Record

Complete Output Schema

Input

Execution Order

Relationship to Other Agents

Configuration

Environment Variables

Agent-Level Configuration

Implementation Status

Error Handling

MCP Integration

Testing Strategy

Unit Tests

Integration Tests

Route A: To Admin Agent (`expense_register`)

Route B: To Data Analysis Agent (`analysis_only`)