How Geometry Solution Automated 40+ Hours of Weekly Invoicing

The Problem: Drowning in Repetitive Logistics Work

Geometry Solution is an e-commerce logistics company that sources products and manages fulfillment for Amazon sellers. Their business model depends on speed and accuracy—processing hundreds of orders, invoices, and warehouse receipts every day. But as CEO Ryan Li describes it, the team was "drowning in paperwork."

Every single day involved manually entering invoice data from suppliers into their accounting system—line by line, product by product. They had to cross-reference warehouse receipts with purchase orders to verify shipments arrived correctly, track inventory across multiple warehouses and update spreadsheets with stock counts, generate customer invoices for Amazon sellers by pulling data from multiple sources, and print and file physical receipts for compliance and auditing purposes.

Ryan's team was spending over 40 hours per week just on these repetitive tasks. The work was mind-numbing, error-prone, and impossible to scale. Hiring more people wasn't a solution—they'd just burn out on the same repetitive workflows. The team needed automation that could handle the complexity of their legacy systems without breaking every time a supplier changed their invoice format or a warehouse portal updated its interface.

The Solution: Building a Software Layer for Legacy Systems

Geometry Solution partnered with Agent School to build a software layer on top of their legacy logistics systems. Instead of trying to automate tasks with brittle scripts, Agent School wired up deterministic workflows that could execute at software speed. The key difference? These workflows were built to be self-healing—when supplier portals changed their UI or invoice formats shifted, the system automatically detected the changes, adapted the workflows, and recertified them without human intervention.

Here's how it worked in practice. Ryan and his team provided their standard operating procedures for each workflow—step-by-step instructions they'd been following manually. Agent School's team analyzed these SOPs and identified which steps could be converted into operation nodes (deterministic actions like clicking buttons, reading fields, or making API calls) and which steps required AI nodes (tasks needing reasoning, like validating invoice line items against purchase orders). Each operation node was stored as a reusable building block, so once it was built for one workflow, it could be reused across others.

The workflows were then structured as JSON files that chained these operation nodes together. When a workflow ran, it executed the stored sequence of actions deterministically—no heavy LLM reasoning on every run, just fast, predictable execution. This meant workflows that used to take 45 seconds with traditional AI approaches now ran in under 2 seconds. And because the workflows were certified before deployment—tested across hundreds of edge cases in isolated environments—Ryan's team could trust them to run in production without constant monitoring.

Four Workflows That Transformed Operations

Invoice Processing Workflow: Ryan walked through his invoice entry process once while Agent School recorded every step. The resulting workflow learned to extract data from supplier invoices (whether they came in as PDFs, emails, or web portals), validate line items against purchase orders, and enter clean data into their accounting system. The workflow now handles over 200 invoices per day with 99.2% accuracy—higher than the team's previous manual error rate. After four weeks of operation, this workflow alone is consistently saving the team 27 hours per week.

Warehouse Receipt Verification Workflow: Every shipment that arrives at their warehouse generates a receipt that needs to be matched with the original purchase order. Agent School built a workflow that logs into warehouse portals, downloads receipts, cross-references SKUs and quantities, and flags discrepancies for human review. What used to take 10 hours per week now happens automatically overnight, with the team only stepping in when the workflow detects an anomaly that requires judgment.

Inventory Tracking Workflow: Managing stock across four different warehouses meant constant spreadsheet updates and manual syncing. The new workflow monitors inventory levels in real-time, syncs data across systems, and triggers reorder alerts when stock runs low. Ryan's team gets a consolidated dashboard without lifting a finger, and the workflow self-heals when warehouse systems change their data formats.

Print & File Automation Workflow: For compliance, certain documents need to be printed and physically filed. The workflow automatically identifies which documents require printing, queues them to the network printer, and logs them in the filing system. Even the most mundane task is now hands-off, freeing the team to focus on higher-value work.

Why It Worked: The Agent School Difference

Traditional automation would have failed here for three fundamental reasons. First, suppliers constantly change their systems. Invoice formats evolve, portal layouts get redesigned, and API endpoints shift without warning. Traditional RPA scripts would break weekly, requiring constant manual fixes. Agent School's self-healing architecture detects these changes through continuous evaluation tests, automatically updates the operation nodes, and recertifies the workflows—no human intervention required.

Second, edge cases happen daily in logistics. Missing line items, duplicate invoices, quantity mismatches, partial shipments—the list goes on. Agent School's certification process tested hundreds of edge cases in isolated environments before deployment, so the workflows knew exactly how to handle each scenario. When a workflow encounters a situation it hasn't seen before, it flags it for human review and learns from the resolution for next time.

Third, speed matters for competitive advantage. Geometry Solution's workflows run at software speed—executing in under 2 seconds with zero LLM calls in production. Traditional AI approaches that reason through every step from scratch take 45+ seconds per task and cost 10x more per execution. By caching effective action sequences as deterministic workflows and only invoking reasoning for novel situations, Agent School delivers both reliability and efficiency. The workflows process hundreds of transactions daily without the latency or costs that make traditional AI automation prohibitively expensive for repetitive work.