Prompt Studio
A collaborative environment for designing, testing, and optimizing AI prompts across LLM providers.
Private Manufacturing Company
Manufacturing / Distribution
80–150 employees across operations, warehouse, and admin
A mid-market manufacturing company supplying industrial components to OEMs across the Midwest. They process 40–60 supplier invoices per week, manage 200+ SKUs across two warehouses, and operate on thin margins where procurement efficiency directly impacts profitability. The procurement team consisted of two buyers, one warehouse manager, and an operations director — all spending significant time on manual coordination instead of strategic sourcing.
Procurement was entirely manual — supplier invoices arrived by email, inventory was tracked on paper, and purchase orders required 3–5 days of back-and-forth.
Finding the Signal in the Noise
Process audit over 3 weeks: shadowed buyers, reviewed 4 weeks of procurement email threads, mapped every touchpoint from invoice arrival to PO completion.
Invoices arriving as PDF attachments in personal email inboxes — no central repository
Vendor verification required manual lookup in a shared spreadsheet with 400+ outdated entries
Inventory checks meant walking to the warehouse floor and physically counting stock
Duplicate purchases occurred 2–3 times per month because nobody had visibility into pending POs
Approval chain stopped completely when the manager was out of office — no escalation path
The business grew from 20 to 100+ employees without ever redesigning its procurement process. Every department built its own workaround, and those workarounds became the de facto system.
Impact: Average procurement cycle: 5–7 days from invoice receipt to PO issue. Estimated 20+ hours per week of combined manual work across the procurement team.
Before: The Manual Workflow
A week in procurement looked like this. Every step was manual, every transfer was an email, and nobody had a complete view of the pipeline.
- No central system of record — information was scattered across inboxes, spreadsheets, and notebooks
- Every invoice required walking to the warehouse for inventory checks — 15–20 minutes per invoice
- Vendor spreadsheet had 30% outdated entries — no validation at point of entry
- Multi-day approval delays when managers were traveling or out of office
- Duplicate POs slipped through 2–3 times monthly, causing double orders and restocking fees
Questions Before Building
Before writing a single workflow node, every assumption gets challenged. These are the questions that shaped the architecture.
Should approval happen before or after inventory validation?
Inventory validation first. If stock is sufficient, no purchase is needed. Approval should only be invoked when procurement is actually required — this avoids unnecessary manager overhead.
What happens when OCR confidence is below 80%?
Route to human review with highlighted low-confidence fields. The AI extracts what it can; the operator only corrects specific values rather than re-entering everything.
How should duplicate invoices be detected?
Hash-based fingerprinting: vendor + invoice number + amount + date produce a deterministic ID. In-flight matching checks pending POs, not just completed ones.
What if the vendor does not exist in the approved list?
Auto-route to procurement lead for vendor onboarding. The system pre-fills vendor details from the invoice to accelerate the process.
Can AI approve purchases directly?
Only for low-value, low-risk categories with deterministic rules (office supplies, recurring subscriptions). All other approvals require human sign-off.
When should accounting be notified?
At PO generation, not at goods receipt. This lets accounting plan cash flow in advance. A second notification fires when goods are received and matched.
Understanding the System
Methodology: Value-stream mapping over 3 weeks. Every procurement step was timed, every handoff documented, and every delay root-caused. We measured cycle time, touch time, wait time, and error rate across 40 real invoices.
68% of cycle time was waiting — approvals, inventory checks, information requests
12% of cycle time was value-adding work (decisions, vendor communication)
20% was rework — fixing errors from manual data entry, correcting wrong POs
5 distinct systems touched per invoice, none integrated
- Eliminate inventory walk by digitizing stock levels with a simple check-in/check-out log
- Auto-validate vendors against an approved list at point of entry — block unknown vendors
- Generate PO drafts from invoice data instead of re-typing everything
- Structured approval with auto-escalation after 24 hours
- Single dashboard showing all active POs, pending approvals, and recent completions
How the System Is Built
A modular event-driven procurement system connecting supplier communications, document AI, inventory tracking, approval routing, and accounting synchronization through a central workflow engine.
Email Ingestion
Monitors purchasing inbox, extracts invoice attachments, and initiates workflow
Document AI (OCR)
Extracts vendor, amount, line items, and PO number from invoice PDFs
Vendor Database
Approved vendor list with status, payment terms, and contact information
Inventory Tracker
Simple stock level database for real-time availability checks
PO Generator
Auto-populates PO template from extracted invoice data and approved vendor info
Approval Router
Routes POs to appropriate approver with escalation after 24h
Accounting Sync
Creates PO record in accounting system and matches on goods receipt
Dashboard
Real-time view of all procurement activity, pending items, and cycle metrics
Live Workflow Simulation
Watch the automated procurement system process a supplier invoice from arrival to completion — AI extracts data, business rules validate, and the system orchestrates every step.
Procurement Operations Dashboard
Step into the operations manager role. Select a sample invoice, review the AI-extracted data, and approve or reject — all within the automated workflow.
Supplier Inbox
3 pendingSelect an invoice to process through the automated workflow.
No activity yet. Select an invoice to begin.
Edge Cases & Exceptions
Supplier accounting error sends the same invoice twice within 3 days. The second invoice has a different email subject line but identical content.
Hash-based fingerprinting detects the match against a pending PO. The duplicate is automatically rejected with a message: "This invoice appears to be a duplicate of PO-2024-0142. If this is a correction, please reference the existing PO number."
A new supplier sends an invoice before completing vendor onboarding. Their company does not appear in the vendor database.
Invoice is routed to procurement lead with pre-filled vendor details extracted by OCR. The lead reviews, approves the vendor, and the workflow resumes automatically. Vendor onboarding time: under 5 minutes.
A supplier sends a handwritten invoice that was scanned at low resolution. OCR confidence drops to 34%.
Invoice is flagged for manual review. The AI highlights which fields it could read (total amount) and which it could not (line items, vendor name). The operator corrects only the unreadable fields.
The inventory system shows 50 units available, but 40 are already reserved for an existing production order that has not been picked yet.
Inventory check now queries available-to-promise (ATP) quantity, not physical stock. Reserved but unpicked stock is excluded. Workflow holds until actual ATP is sufficient.
A PO requiring manager approval sits for 36 hours because the manager is at a conference with limited email access.
Escalation timer fires at 24 hours — sends reminder. At 36 hours, auto-escalates to department head. The system also notifies the original approver when they return: "You have 3 pending approvals that were escalated in your absence."
A supplier submits an invoice under a different legal entity name than the one in the approved vendor list (e.g., "Acme Corp" vs "Acme Manufacturing Inc.").
Fuzzy matching on business name + tax ID catches the mismatch. Invoice is routed to procurement lead with both names displayed for comparison. Once confirmed, the alias is added to the vendor record.
Engineering Breakdown
Event-driven architecture using n8n as the central workflow orchestrator. Each procurement step is a discrete node in a directed graph — parallelizable where possible (vendor validation + inventory check), sequential where required (approval after PO generation). AI document processing runs as a serverless function triggered by workflow events. The data layer combines Airtable (vendor database, PO records) with a lightweight SQLite inventory tracker. Accounting sync uses QuickBooks Online API via OAuth2.
Business Impact
The manual procurement process was systematically replaced with an automated workflow. Cycle time dropped from days to minutes. The team shifted from data entry to exception handling. No new headcount was needed as the company grew.
Cycle Time
5 days → 30 min
Manual Steps
12 → 3
Duplicate POs
~0 per month
Process Visibility
Real-time dashboard
Notes from the Engineer
What I initially considered — and why I rejected it
I initially wanted to build a custom procurement platform with a React frontend, backed by a PostgreSQL database. The reasoning was: "If this is going to be their core procurement system, it deserves a proper application." I rejected this after the first week of discovery. The client did not need another application — they needed their existing tools to talk to each other. A custom platform would have meant retraining, migration, and ongoing maintenance. n8n let me connect what they already had: email, spreadsheets, QuickBooks. The lesson: build the thinnest possible integration layer, not the most impressive application.
The hardest part was not the automation — it was the data
Writing the n8n workflows took about a week. Cleaning the vendor database took three weeks. The spreadsheet had 400+ vendors, 30% of which were outdated — duplicates, renamed companies, vendors that had been inactive for years. Before you can automate a process, you have to trust the data that feeds it. We ended up building a vendor verification workflow as a prerequisite step: email each vendor, confirm their details, mark active/inactive. This was tedious, but it directly prevented the worst failure mode: the system auto-generating a PO for a defunct vendor.
The biggest trade-off: AI vs. deterministic rules
There is a temptation to use AI for everything because it looks impressive. I deliberately chose deterministic rules for vendor validation, duplicate detection, and approval routing. AI is used only where it genuinely adds value: document understanding (OCR) and fuzzy vendor matching. Why? Because procurement spend is real money. A false-positive duplicate detection means a double order. A false-negative vendor validation means ordering from an unapproved supplier. Deterministic rules are auditable, predictable, and explainable. AI is probabilistic. Use it where probabilities are acceptable (data extraction) but not where they are not (financial validation).
What I would change if rebuilding today
I would add idempotency keys from day one. We had incidents where a network timeout caused the workflow to retry, creating duplicate entries in the accounting system. The fix was straightforward — generate a unique idempotency key at the email ingestion step and check it before any side-effect operation. This should have been in the initial architecture, not added as a patch. Second, I would use a proper queue (RabbitMQ or Redis) instead of n8n's built-in execution queue for high-volume periods. On Monday mornings when suppliers send 20+ invoices simultaneously, execution order matters.
Lessons Learned
Automation reveals data quality issues before it solves process issues
The hardest part of this project was not building the workflows — it was cleaning the vendor database. Every automation initiative should budget 2x the expected time for data preparation. If the data feeding your workflow is unreliable, the workflow itself becomes unreliable.
Deterministic rules should always run before AI
There is a temptation to use AI as a hammer for every nail. But most errors in the procurement process were not comprehension errors (AI territory) — they were validation errors (rules territory). Checking a vendor against an approved list does not require GPT-4. Running deterministic checks first catches the majority of issues cheaply and predictably.
Humans are not the problem — bad tools are
Before this project, the procurement team was blamed for being slow. After automation, the same people processed 3x more POs in less time. They were not the bottleneck — the tools forced them to be. The correct framing is not "replace humans with automation" but "give humans tools that let them do the work that matters."
Approval workflows must design for absence, not presence
The original process assumed managers were always available. The automated system assumes they are not — and escalates accordingly. Designing for the failure case (manager on leave, sick, quit) made the system resilient to reality rather than fragile to an ideal.