Applied AI & Workflow Automation

Agentic AI Purchase Order Processing System

Portfolio Case Study

Status:  In production

Stack:  n8n  |  Anthropic LLM  |  Gmail API  |  Google Drive API  |  Google Sheets  |  Salesforce  |  Coefficient  |  Coda

Built:  Independently, outside formal role

Impact:  10+ hours per week recovered across PO intake, file routing, and CRM reconciliation

Context

Zeitview is an enterprise B2B SaaS company providing drone inspection and AI analytics services to Fortune 500 energy clients. The operations team processes purchase orders daily from a portfolio of 20+ enterprise accounts, each with distinct workflows, procurement systems, and document formats.

Problem

Purchase orders arrived across multiple email inboxes with no consistent structure. Every client formatted their PO documents differently: the PO number might be in the header, footer, a reference field, or buried in a line item table. Emails frequently contained multiple attachments, only one of which was the actual PO. Senders varied by client domain, individual contact, and in some cases, third-party procurement platforms that obscured the origin entirely.

Processing each order required a human to open the email, identify the correct attachment, extract the PO number, rename and file the document in the appropriate Google Drive folder, log it manually in a spreadsheet, and enter the relevant data into a Coda work order form. The process was repeated for every order, every day. It was reliable only because someone showed up and did it manually each time.

Opportunity

The process had no steps that required human judgment at the intake stage. Every decision (which attachment is the PO, what client sent it, where it should be filed, what data needs to be logged) was rule-based and pattern-driven. That meant it was automatable. The opportunity was not just to save time, but to make the intake process faster, more consistent, and less dependent on any one person executing it correctly.

A secondary opportunity emerged once the core pipeline was stable: closing the gap between PO data and Salesforce Opportunity data. Operations needed both to execute jobs effectively, but pulling them together required manual cross-referencing across systems. Automating that connection would give Ops a complete picture in one place.

Solution Architecture

I built an agentic AI workflow in n8n that handles the full intake process from email receipt to Coda work order population, with a human approval gate before final submission.

•    Gmail API monitors designated inboxes and triggers the workflow on new message receipt.

•    The LLM (Anthropic) reads the email body and all attachments to identify which file is the PO, extract the PO number, and classify the order by client. Client-specific extraction logic handles the highest-variance accounts where a single generic prompt was not sufficient.

•    Google Drive API renames the confirmed PO document using a standardized naming convention and routes it to the correct client folder.

•    Google Sheets logs each processed PO with structured metadata: PO number, file link, client, and date. This creates an audit trail and a clean data layer for downstream reference.

•    Coda receives the structured output and pre-populates the work order entry form. A human-in-the-loop approval gate holds submission until a team member confirms the entry is accurate.

•    Coefficient exports Salesforce Opportunity and quote data on a sync schedule, feeding the corresponding account context into the pipeline so the Coda form is populated with both PO data and the relevant Salesforce record. This works around a Salesforce write-back feature that requires a paid upgrade not currently in scope.

n8n was selected for its no-code entry point, low cost relative to alternatives like Zapier or Make, and self-hostable architecture. The modular node structure means individual components can be swapped or extended without rebuilding the surrounding pipeline.

Impact

•    10+ hours per week recovered across PO intake, file routing, and CRM reconciliation.

•    Intake process is now consistent regardless of which team member is on deck, eliminating single-person dependency.

•    Operations receives a pre-populated work order form with both PO data and Salesforce Opportunity context, reducing manual cross-referencing before job execution.

•    Full audit trail of every processed PO, with structured metadata available for reporting and downstream integration.

•    Human approval gate preserves accuracy without restoring the manual bottleneck.

Product Capabilities Demonstrated

•    Opportunity identification: spotted a high-frequency manual process with no steps requiring human judgment and scoped a solution independently.

•    Agentic AI design: architected a multi-step LLM-orchestrated pipeline with conditional logic, client-specific extraction rules, and structured output formatting.

•    Production deployment: built, tested, and shipped a system that runs daily in a live operational environment, not a prototype or proof of concept.

•    Real-world edge case handling: iterated on classification logic to handle variant PO formats, multi-attachment emails, and inconsistent sender patterns across enterprise clients.

•    Architectural tradeoff reasoning: selected n8n for speed and cost, designed human-in-the-loop gates deliberately, and engineered a Salesforce workaround within real budget constraints.

•    Bias for action: initiated and delivered the entire project outside my formal role because the process was broken and the fix was visible.