Track employee expense receipts from Telegram with GPT-4o-mini and Google Sheets

Employee sends a photo → AI extracts merchant, amount, date → policy check auto-approves or routes to manager → record results and monthly email report

Who

Small-team finance or ops staff who need employee expense claims with approval trails and a ledger

You get

Employees submit receipt photos via Telegram for automatic entry; the system auto-approves, rejects, or escalates to a manager based on Google Sheets rules, then emails a monthly summary

You need

n8n · telegram · google-sheets · gmail

Steps

1. Prerequisites: create spreadsheet and credentials

   Create a Google Sheets workbook with two tabs: Policy and Expenses. Policy headers: Category | LimitPerReceipt | ApprovalAbove | Allowed | Notes. Expenses headers: SubmittedAt | Employee | TelegramUserId | ChatId | Merchant | Date | Amount | Currency | Tax | Category | Description | Status | Decision | DecisionNote | ReceiptFileId | Month. Configure Telegram, OpenAI, Google Sheets, and Gmail credentials in n8n. Record the manager chat ID and finance chat ID.

2. Trigger: receive receipt photo on Telegram

   Employee sends a photo to the bot (a caption like "team lunch with client" is recommended). The Telegram Trigger node downloads the largest image size. If no photo is attached, the bot auto-replies asking for a photo.

   Execution prompt · Source text。Telegram node Request Receipt Photo text field

   Copy prompt

   =Please send the receipt as a photo (not as text). Add a short caption if you like, for example "team lunch with client" - the amount, date and merchant are read from the picture.

3. AI extraction: GPT-4o-mini reads and structures the receipt

   The OpenAI node sends the image as base64 to gpt-4o-mini with a strict JSON instruction. A Code node parses the response, normalizes dates (e.g. 12.09.2026, 09/12/2026, 12 Sep 2026), amounts (1.250,00 / 1,250.00), and currency symbols, maps the category to the allowed list, and builds a duplicate key. If the call fails, the employee is asked to retry and nothing is recorded.

   Execution prompt · Source text。OpenAI node AI Analyze Receipt Photo text/prompt field

   Copy prompt

   You read receipts for an expense claim. Look at the image and return ONLY a JSON object with these keys:
   {"is_receipt": true or false, "merchant": "shop or restaurant name", "date": "YYYY-MM-DD of the purchase", "total": number (grand total actually paid, including tax and tip), "currency": "3-letter ISO code, infer from symbol or country", "tax": number (VAT / sales tax shown, 0 if none), "category": one of Meals | Travel | Transport | Accommodation | Office | Software | Other, "payment_method": "card | cash | other", "items": "short description of what was bought, max 100 characters", "confidence": 0-100}
   Rules: copy numbers exactly as printed; if a value is not visible use "" or 0; never guess a merchant; if the picture is not a receipt set is_receipt to false. No markdown, no explanations.

4. Policy enforcement: deduplication, limits, and routing

   Load the policy row for the receipt's category from the Policy tab, and check the Expenses tab for an identical claim. A Switch node routes: auto-approve (within limits and no approval needed), duplicate, missing info, not allowed (disabled category or over limit), or needs approval (over threshold, unknown category, older than 60 days, or AI confidence below 70%).

5. Resolution: record, notify, and forward to finance

   All outcomes converge to one branch and are appended to the Expenses tab. The employee receives a Telegram reply with status icon and reason. Approved receipt photos are forwarded to the finance chat. Claims needing approval use Telegram sendAndWait with Approve/Reject buttons; the workflow waits up to 3 days. After manager action, status and timestamp are updated. [First-test safety] Before the first test, point Add Entry to Expenses documentId at a test sheet, or set Inform Employee of Outcome / Send Receipt to Finance / Request Manager Approval chatId to a test account, so production expense rows and real coworkers are not touched.

6. Monthly report: automated email on the 1st

   A Schedule Trigger runs at 8 AM on the 1st of each month. It reads the previous month's approved expenses, totals per employee and per category (by currency), builds an HTML table, and emails it via Gmail. The MONTH constant in the Code node can be set manually to rebuild a past month.

You are an n8n workflow setup assistant. The user wants to replicate the "Telegram receipts → GPT-4o-mini extraction → Google Sheets approval → monthly report" automation. Assist in this order: 1) Confirm the user has an n8n instance with a public URL; 2) Verify the Google Sheets Policy/Expenses tab structure matches requirements; 3) Guide importing the workflow JSON and replacing placeholders like YOUR_SPREADSHEET_ID, YOUR_MANAGER_CHAT_ID, YOUR_FINANCE_CHAT_ID; 4) Suggest sending a test receipt to validate AI extraction and policy routing. Do not perform sensitive operations on the user's behalf.