mcs-finance-statement-agent

Finance Statement Agent

Summary

Conversational AI for credit financial statement extraction. An analyst uploads a PDF financial report in Microsoft Teams; a Copilot Studio agent orchestrates an Azure-hosted pipeline that extracts the Income Statement, Balance Sheet, Cash Flow, and computed Ratios into a downloadable Excel workbook ready for credit spreading.

Multi-language: English, Chinese, Japanese, French (auto-detected, label-reconciled to a canonical schema).

Data flow: Analyst → Teams → Copilot Studio → Custom Connector → Azure Functions (HTTP 202) → Poll until complete → Excel generated → SAS download link in chat.

Frameworks

Prerequisites

• Microsoft 365 tenant with Copilot Studio licensed
• Power Platform environment with Dataverse and custom connector permissions
• Azure subscription with rights to create Resource Group, Function App, Content Understanding (or Document Intelligence), Azure OpenAI (with gpt-4.1 deployment), and Storage Account
• Azure CLI 2.50+
• Azure Functions Core Tools 4.x
• Power Platform CLI (pac)
• Python 3.11+, Node.js 18+

Contributors

mcs-finance-statement-agent | Shaji Sivaraman (@sgshaji), Microsoft

Version history

Version	Date	Author	Comments
1.0	April 30, 2026	Shaji Sivaraman	Initial release

Disclaimer

THIS CODE IS PROVIDED AS IS WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING ANY IMPLIED WARRANTIES OF FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR NON-INFRINGEMENT.

---

Minimal Path to Awesome

The sample is split across three deployables: an Azure Functions backend, a Copilot Studio agent, and an optional Power Apps Code App for human-in-the-loop review.

1. Clone and authenticate

git clone https://github.com/pnp/copilot-pro-dev-samples.git
cd copilot-pro-dev-samples/samples/mcs-finance-statement-agent/src
az login
az account set --subscription "<subscription-id>"

2. Provision Azure resources

Create the resources listed in Prerequisites. Grant the Function App's system-assigned managed identity these roles at the resource group scope:

• Cognitive Services User — Content Understanding / Document Intelligence
• Cognitive Services OpenAI User — Azure OpenAI
• Storage Blob Data Contributor — Storage Account

3. Configure and run the backend

cd azure-functions
cp .env.example .env
# edit .env with your endpoints (no API keys — Managed Identity handles auth)
pip install -r requirements.txt
func start                      # http://localhost:7071/api/health

4. Deploy the Function

func azure functionapp publish <your-function-app> --python

5. Set up the custom connector

1. Power Platform → Custom connectors → New → Import OpenAPI file → docs/custom-connector-swagger.yml
2. Update the Host to your Function App
3. Authentication: API key (header x-functions-key) — retrieve via:

   az functionapp keys list --name <your-function-app> --resource-group <your-rg>

4. Create a connection using the new connector

6. Push the Copilot Studio agent

cd ../copilot-studio-agent
pac copilot push --environment <env-id>

7. (Optional) Deploy the HITL review Code App

cd ../code-app
npm install && npm run build
pac code push

Features

This sample demonstrates an end-to-end agentic pattern for processing long-running document workloads from Copilot Studio:

• Async extraction with polling — Power Platform custom connectors have a default 30-second synchronous request timeout. The pipeline returns HTTP 202 {jobId} in ~100 ms; the Copilot Studio topic polls /extract/status/{jobId} every 30 s until completed (ConditionGroup + GotoAction loop bounded by max-attempts)
• Pluggable extraction backend — Content Understanding (default), Document Intelligence, Textract, or local pdfplumber, selectable via EXTRACTION_BACKEND
• 5-stage pipeline — analyze → select → extract → enrich → validate. Backends emit a common markdown + HTML-table format so Stages 2–5 are reusable
• Multi-language label reconciliation — Azure OpenAI maps source-language labels to a canonical English schema for English, Chinese, Japanese, and French statements
• Managed Identity end-to-end — no API keys for Azure service-to-service auth. The only secret is the Function key consumed by the Power Platform custom connector
• Job state in Blob with 30-min TTL — bounded storage; SAS URLs returned in chat for the generated Excel
• Human-in-the-loop review — optional Power Apps Code App provides an analyst grid backed by Dataverse for correcting extracted values before downstream credit spreading
• Multi-row column-header parsing — handles statements where Q4 and FY columns share a parent header (e.g., Meta Income Statement)
• >5 MB upload handling — uses Copilot Studio Question node bound to FilePrebuiltEntity (direct Activity.Attachments access fails for files > 5 MB)

Repository layout (under src/)

src/
├── azure-functions/         # Python backend (HTTP 202 async pipeline)
│   ├── function_app.py      # HTTP router
│   └── extractor/           # 5-stage pipeline + clients (CU, DI, Textract, pdfplumber)
├── copilot-studio-agent/    # Copilot Studio YAML (agent, topics, actions, workflows)
├── code-app/                # Power Apps Code App — React HITL review grid (Dataverse)
└── docs/
    ├── architecture.png             # Architecture diagram
    └── custom-connector-swagger.yml # Swagger 2.0 spec for the custom connector