Auto-Classify Purchase Requisitions with Google Gemini and Route for Approval

Webhook ingestion → data validation → Gemini classification with confidence scoring → auto or manual approval → Sheets logging → Gmail/Slack notifications

Who

Operations or finance teams automating procurement approval workflows

You get

Automatically categorize purchase requisitions, auto-approve high-confidence requests, escalate low-confidence ones for manual review, and log and notify automatically

You need

n8n · gmail · google-sheets · slack

Steps

1. Configure Webhook and Credentials

   Create a new workflow in n8n, add a Webhook node, and set the POST path to /pr-submission. Configure credentials in this order: Google Gemini API (for AI classification), Gmail OAuth2 (for review and result emails), Slack OAuth2 (for procurement channel alerts), and Google Sheets (for audit logging). Ensure the submission source can send a JSON body containing pr_id, requester, description, and amount.

2. Configure the AI Classification Node

   Add a Basic LLM Chain node (AI: Categorize Purchase Requisition) and connect the Google Gemini Chat Model as the language model. Paste the template's classification prompt into the Prompt field. The prompt instructs Gemini to return strict JSON with final_category, confidence_score, misclassification_risk, suggested_correction, and reasoning. You may modify the Allowed Categories list to match your procurement taxonomy.

   Execution prompt · Source text。Prompt / text field of the 'AI: Categorize Purchase Requisition' node

   Copy prompt

   =You are an enterprise procurement classification assistant.
   
   Analyze the following purchase requisition and determine the best procurement category.
   
   Purchase Requisition Details:
   - PR ID: {{$json.pr_id}}
   - Requester: {{$json.requester}}
   - Department: {{$json.department}}
   - Description: {{$json.description}}
   - Amount: {{$json.amount}} {{$json.currency}}
   - Suggested Category: {{$json.suggested_category}}
   
   Items:
   {{ JSON.stringify($json.items, null, 2) }}
   
   Allowed Categories:
   - IT Equipment
   - Office Supplies
   - Software
   - Facilities
   - Marketing
   - HR & Training
   - Professional Services
   - Logistics
   - Uncategorized
   
   Tasks:
   1. Determine the most appropriate category.
   2. Detect whether the suggested category may be incorrect.
   3. Provide a confidence score between 0 and 100.
   4. Suggest corrected category if needed.
   5. Explain reasoning briefly.
   
   Return ONLY valid JSON in this exact structure:
   
   {
     "final_category": "",
     "confidence_score": 0,
     "misclassification_risk": "",
     "suggested_correction": "",
     "reasoning": ""
   }

3. Set Up Approval Branching Logic

   Add a Code node to parse Gemini's JSON response (Parse AI Classification Result), then another Code node to compute confidence level and approval recommendation (Calculate Classification Confidence). Current logic: auto_approve=true when confidence ≥ 85 and misclassification_risk is not High; manual review for 60–84 or high risk; manual review for <60. Then add an If node (Check Auto Approval Eligibility) branching on auto_approve: the pass branch sets approval_status='Auto Approved' and routing_status='Ready for ERP Update'; the review branch sets approval_status='Pending Manual Review' and routing_status='Escalated to Procurement Team', and triggers a Gmail review ticket.

4. Merge Results and Write to Google Sheets

   Use a Merge node to combine original PR data with AI results by position (Merge PR Data With AI Result), then connect a Google Sheets node (Store Procurement Record) in append mode to the PR_Records sheet. Ensure column names match the node schema: pr_id, requester, department, description, amount, currency, final_category, confidence_level, confidence_score, approval_status, misclassification_risk, suggested_correction, routing_status, review_reason, ai_reasoning, processed_timestamp.

5. Send Gmail and Slack Notifications

   After the Sheets node, connect in parallel a Gmail node (Notify Procurement Approver) and a Slack node (Send Procurement Slack Alert). Gmail sends a processing summary to the approver; Slack posts a channel alert with PR ID, category, confidence, approval status, and risk level. Update recipient addresses and channelId to actual values. [First-test safety] Before the first test, temporarily disconnect the Gmail, Slack, and Google Sheets write nodes, or retarget them to a test label/channel/sheet; enable real targets only after the dry-run looks correct.

   Execution prompt · Source text。text field of the 'Send Procurement Slack Alert' node

   Copy prompt

   =Procurement Request Processed!
   
   PR ID: {{$json.pr_id}}
   
   Requester: {{$json.requester}}
   
   Department: {{$json.department}}
   
   Final Category: {{$json.final_category}}
   
   Approval Status: {{$json.approval_status}}
   
   Confidence Level: {{$json.confidence_level}}
   
   Misclassification Risk: {{$json.misclassification_risk}}
   
   Routing Status: {{$json.routing_status}}

You are an enterprise procurement classification assistant. Based on the provided purchase requisition details, determine the best procurement category from the allowed list, detect if the suggested category may be incorrect, provide a confidence score between 0 and 100, suggest a corrected category if needed, and explain reasoning briefly. Return strict JSON with: final_category, confidence_score, misclassification_risk, suggested_correction, reasoning.