AI Workflow Automation Suite: 4h of manual ops work down to 20m

An operations team was keying vendor documents into their systems by hand: invoices by email, vendor EDI feeds and files dropped into Slack. It took about four hours a day, and every vendor formatted things differently.

Intake from email (SES to S3), vendor EDI feeds and Slack uploads, into one pipeline.

Parsing: OCR, then Claude API extraction into structured data. Accuracy went from 85% to 93% with few-shot examples, and to 97% with a validation layer.

Orchestration on Temporal: parse, validate, vendor lookup, price check, route for approval. Every step retries with backoff, and if a service crashes the workflow resumes from the exact step it left off. No lost state, no manual restarts.

Human in the loop: anything below the confidence threshold goes to the ops team in Slack with Approve and Reject buttons. They never open a separate dashboard.

Automated actions: purchase order, invoice and receiving matching, auto-posting to NetSuite, and compliance checks for new vendors.

About 200 documents processed a day

97% fully automated; the other 3% reach a person with context

Manual work down from 4 hours to 20 minutes a day

Temporal over a task queue. A queue would have meant building state management, retries and approval pauses from scratch. Temporal gives all three, so the work went into business logic instead of infrastructure.

97% automation, not 100%. Most AI automation fails by trying to automate everything. Designing a graceful hand-off to people for the edge cases is what made the team trust it. The hard part isn't the LLM; it's the engineering around it.

Stack: Python, FastAPI, Claude API, Temporal, PostgreSQL, AWS. Code is private under NDA; architecture walkthrough on request.