§ Field NotesWorkflow AutomationMay 11, 20265 min read

    Automating Invoice Processing With AI.

    The workflow is simple, but the pattern behind it is worth understanding because it applies to a lot of common business problems.

    Dan Storms
    CEO · Jinka

    We recently put together an automation for our own accounts payable workflow. The mechanics are not complicated, but the underlying pattern is one we keep coming back to across very different problem areas.

    § 01

    The problem

    Vendors send invoices as PDF attachments by email. Someone has to open each one, pull out the relevant data, match it to the right vendor in the accounting system, and create a journal entry.

    It is not difficult work, but it is time-consuming and easy to fall behind on. The kind of task that quietly absorbs hours every week without anyone noticing where the time went.

    § 02

    How it works

    We use n8n, a workflow automation tool, to monitor an inbox for incoming emails. When a PDF attachment arrives, n8n sends it to Google Gemini, which reads the document and extracts the accounting data into a structured format.

    The system then matches the vendor information against our vendor list in ERPNext and creates the journal entry automatically. Once that is done, our bookkeeper gets a notification to review and approve the transaction.

    § Fig. 01 — invoice processing flow
    § Trigger01
    Inbox
    n8n watches for incoming emails with PDF attachments.
    § Extract02
    Gemini
    Google Gemini reads the PDF and returns structured accounting data.
    § Sync03
    ERPNext
    Vendor matched against the list. Journal entry created automatically.
    § Review04
    Bookkeeper
    Notification sent. A human approves before anything is final.
    The whole thing runs without anyone touching it until the final review step.
    The whole thing runs without anyone touching it until the final review step.
    § 03

    Where else this pattern applies

    The underlying logic is the same across a lot of workflows: something arrives, AI extracts or processes the relevant information, that information gets pushed into the right system, and a person is notified to review.

    Lead intake is a good example. Someone books a meeting, provides their details, and rather than just logging that in a CRM, the system does a quick background summary using something like Perplexity and sends it to you before the call. From there you can layer in follow-up sequences, CRM updates, and outreach testing without any of it requiring manual input.

    • 01 —
      Trigger
      Something arrives. An email, a form submission, a calendar booking, a webhook from another system.
    • 02 —
      Extract
      AI reads unstructured input and returns structured data. PDFs, images, free-text fields, transcripts.
    • 03 —
      Sync
      The structured output is pushed into the system of record. ERP, CRM, ticketing, billing.
    • 04 —
      Review
      A human is notified to confirm or correct before anything irreversible happens.
    § 04

    The tools

    n8n for workflow orchestration, Google Gemini for document extraction, ERPNext as the accounting system.

    n8n is self-hostable, which matters if you are processing documents with sensitive financial or client data. Keeping the orchestration layer on infrastructure you control means the only thing leaving your environment is the specific call to the model, on terms you have already vetted.

    § Work with us

    Have a workflow that should be automated?

    Jinka builds AI-driven automations that quietly remove the repetitive work from your team's day.

    Let's talk
    § Continue reading

    More from the field

    All resources →