Agents running in production right now.
Client details are kept confidential at their request. What's shown here is representative of actual deployments — the workflows, the approaches, and the real results. No invented scenarios.
Freight document classification and routing
A regional logistics operator was processing 400–600 freight documents daily — bills of lading, CMRs, customs declarations — through a four-person team that spent most of their time on document type identification and initial routing. The workflow existed in email and a shared drive.
We mapped the classification logic over two days of sessions with the team, discovered that 73% of documents fell into six clear categories, and built a classification agent that routed with 94% accuracy on the real document set. The remaining 6% routes to a human review queue.
Invoice reconciliation agent for payment platform
A payment platform was reconciling invoices against transaction records manually at month-end — a process that took two finance team members most of their week and regularly caused delayed closes. The data lived in three systems with inconsistent naming conventions.
We built a reconciliation agent that normalises naming across systems, matches invoices to transactions with a configurable confidence threshold, and flags discrepancies for human review. The threshold was calibrated over three months of test runs against historical data before going live.
Support ticket classification and first-response drafting
A B2B SaaS company with 200 enterprise clients was receiving 150–250 support tickets weekly, triaged manually by a two-person team before assignment. Response time was 8–12 hours for initial acknowledgement.
We built a triage agent that classifies by issue type, urgency, and account tier, then drafts an initial response for agent review and approval. The draft step was deliberate — the team wanted to maintain response quality, not automate it away. Approval takes 30–90 seconds per ticket.
Purchase order anomaly detection and flagging
A mid-size manufacturer was processing 1,200–1,800 purchase orders monthly across four divisions. Pricing anomalies — duplicate line items, price deviations, unit-of-measure mismatches — were being caught inconsistently or not at all before approval.
We built a PO review agent that runs each order through a checklist of anomaly patterns defined with the finance team, attaches a structured findings note, and routes flagged orders to the appropriate approver with the specific issue highlighted. The findings note format was designed for their existing approval workflow.