GROUNDWELL is the intelligence layer that sits between incoming field data and the financial system of record. We build job-cost models for heavy civil infrastructure contractors. We call our approach costing cognition: our AI agent, Arnon, reasons about how a heavy civil job turns crews, equipment, and material quantities into costs and predicts what each field entry will do to a contractor’s job cost before the entry posts. A general-purpose LLM like Claude or ChatGPT can summarize data, such as a foreman’s Daily Report, but it does not validate whether the hours, quantities, and cost codes are correctly reported. In one widely cited global survey, 30% of construction professionals said more than half of their project data are bad (Autodesk & FMI, 2021). Our system knows what’s behind each line of a daily report: all day, every day, our system ingests data from the back office and from the field signal network, detecting anomalies and verifying actual activities performed. And when it detects a deviation, Arnon — our agent — signals that the job may be veering off scope, off schedule, or off budget. Over repeat runs, Arnon creates a unique, machine-readable model of how a contractor’s operational and financial worlds fit together. In fact, with increased precision and reliability, future bids are improved and margin expands. Each customer runs a growing share of its projects through GROUNDWELL, from 10% in Y1 to 90% in Y3.