Azure Functions Hosted Skill: Expense Processor

A markdown-first Azure Functions hosted skill for queue-driven expense processing. Its trigger and instructions live in src/expense_processor.agent.md, and Azure Functions handles execution and scale-to-zero.

What it does

• 🧾 Reads any format: text, email, key/value, or JSON, and normalizes amounts expressed with symbols, currency codes, words, or colloquial units.
• 📚 Picks the right policy: lists the documents in Blob Storage and selects the one whose scope matches, then reads and applies it through a read-only Connector Namespace MCP server.
• 🛡️ Governs model calls: sends deployed model traffic through an Azure API Management AI Gateway with managed identity, token limits, prompt shielding, harmful-input filtering, and token metrics.
• 🚦 Routes the decision: approve → expense-approved, review → expense-review, flag / FX → expense-flagged.
• 🔀 Proves it's reasoning: the same normalized 450 USD is auto-approved as travel but sent to review as a client dinner; tighten one policy document and only that category reroutes.

Prerequisites

• An Azure subscription
• uv
• Azure Developer CLI (azd)

The deployment creates an API Management Developer instance. This is the lowest tier that supports the sample's token-limit and Content Safety policies, but it has a recurring cost and no production SLA.

Quickstart

uv sync --project src
azd auth login
azd up

The deployment seeds the policy documents but does not submit expenses, so a fresh deployment's queues remain empty. Verify the empty output queues:

uv run --project src --no-sync python scripts/read_decision.py --queue all --peek --cloud

Submit the three bundled demo expenses, allow up to two minutes for processing, and read the resulting decisions:

uv run --project src --no-sync python scripts/setup_demo.py send-samples
uv run --project src --no-sync python scripts/read_decision.py --queue all --peek --cloud

send-samples skips submission when an output queue already contains a decision. To reset the three output queues for a fresh demo, receive and remove their messages, then run send-samples again:

uv run --project src --no-sync python scripts/read_decision.py --queue all --cloud --max 1000

Omitting --peek deletes the messages it reads. This command clears the three output queues; it does not clear the input or poison queue. The --no-sync commands reuse the environment prepared by the initial uv sync --project src and do not contact PyPI.

You should then see:

Request as received	Amount inferred	Policy	Queue
“four hundred and fifty dollars” flight	450 USD	travel-policy.md	expense-approved
USD 450.00 monitor	450 USD	equipment-software-policy.md	expense-approved
$450 client dinner	450 USD	meals-entertainment-policy.md	expense-review

Open the Application Insights overview, then use Search (formerly Transaction search) or Logs to inspect the hosted skill, model calls, and tool spans:

azd monitor --overview

Look for successful execute_tool azureblob_ListFolder_V4, execute_tool azureblob_GetFileContentByPath_V2, and execute_tool azurequeues_PutMessage_V2 spans, plus AI Gateway requests and token metrics. See Deploy for the portal walkthrough.

Clean up with azd down --purge.

Run it locally

Install Azurite and Azure Functions Core Tools. Copy src/local.settings.json.sample to src/local.settings.json and set the model endpoint and deployment. No API key is needed; the runtime uses DefaultAzureCredential, so sign in with az login or azd auth login. Policy lookup uses the deployed Connector Namespace because Connector Namespace has no local emulator. After azd provision, copy POLICY_MCP_SERVER_URL and QUEUE_MCP_SERVER_URL from azd env get-values into local settings and leave both MCP client IDs empty so your developer credential is used.

azurite --silent --location .azurite               # terminal A
cd src && uv run func start                         # terminal B
uv run --project src --no-sync python scripts/send_expense.py --file samples/travel.txt   # terminal C
uv run --project src --no-sync python scripts/read_decision.py --queue all --peek --cloud

The model call and both Connector Namespace MCP servers still use Azure. The local trigger reads Azurite, but decisions are written to the deployed Azure output queues because Connector Namespace has no local emulator. For setup and Windows help, see Troubleshooting.

How it works

flowchart LR
    msg([raw message<br/>text · JSON · key-value])

    subgraph inbound["Azure Queue Storage · inbound"]
        inq[[expense-requests]]
    end

    subgraph policies["Azure Blob Storage · policies container"]
        pdocs[(travel · meals · equipment<br/>general policy docs)]
    end

    skill{{Expense Processor skill<br/>extract · select · apply · route}}
    gateway{{API Management AI Gateway<br/>identity · token limits · content safety · metrics}}
    model[(Microsoft Foundry<br/>gpt-5.4)]

    subgraph outbound["Azure Queue Storage · outbound"]
        approved[[expense-approved]]
        review[[expense-review]]
        flagged[[expense-flagged]]
    end

    msg --> inq
    inq -->|queue trigger| skill
    pdocs -->|Connector Namespace<br/>Blob MCP tools| skill
    skill -->|Responses API| gateway
    gateway -->|managed identity| model
    skill -->|Queue MCP| approved
    skill -->|Queue MCP| review
    skill -->|Queue MCP| flagged

Loading

The runtime discovers the hosted skill's Markdown definition. Its front matter defines the queue trigger, and its body contains the instructions. A read-only Azure Blob MCP connector lists and reads policy documents, while a write-only Azure Queues MCP connector routes decisions to the three output queues. The Connector Namespace managed identity receives container-scoped Blob Reader and queue-scoped Message Sender roles. In Azure, model requests use the Function identity to enter the AI Gateway; API Management then calls Microsoft Foundry with its own managed identity.

How it works · Use cases · Customize · Deploy · Troubleshooting

Learn more

• Hosted skills in Azure Functions
• Azure Functions Flex Consumption
• Connector Namespace
• Generative AI gateway capabilities
• uv · PEP 723: inline script metadata

License

MIT © Microsoft Corporation.