NOV's mud report automation system now writes its own data-extraction prompts in an average of 8.8 minutes, compared with about 960 minutes for a drilling engineer to build one by hand, according to company testing completed in mid-2025. The AI-generated prompts also beat manually written ones by 2 to 8% in average parsing accuracy.
Daily mud reports track fluid properties, solids control and other drilling parameters that engineers use to monitor rig performance, but the paperwork itself is a problem. Reports arrive from different vendors as multipage PDFs, and the table layouts, parameter names and which subsections appear can vary by vendor, by well and by reporting date. Some data, including vendor names and API well numbers, sits as text inside images rather than in selectable text. NOV has spent the past two years building a cloud-based system, called Mud Report Automation, to pull structured data out of that mess without an engineer parsing each report by hand.
How the system reads a report
The system runs on two modules. An extraction module pulls text out of the PDF reports and converts it into a format a large language model can read, such as Markdown or plain text, while trying to preserve the original table structure. A second module, built on an LLM, interprets that content, spots the structural patterns and reformats unstructured text into a standard data format for storage and analysis. NOV processes an estimated 100 to 200 mud reports a day through the system.
"We're extracting all the useful information out of these documents," said Jay Yoon, NOV's manager of data science and applied AI.
Self-learning added mid-2025
For a vendor the system already knows, it retrieves a stored prompt and applies it directly. For a new vendor, an AI agent drafts an initial prompt by studying reports from existing vendors, then refines and stores it in what NOV calls a dynamic prompt registry. A drilling engineer still checks a small number of that vendor's reports by hand to set a ground-truth baseline before the agent's output is trusted. "We need the human to consume the end result and check everything," said Junzhe Wang, a data scientist at NOV.
This agentic version is NOV's second iteration of the system. The first, developed in late 2024 and rolled out to active rig sites in March 2025, ran on a fixed set of designed prompts with no ability to adapt to new report formats. NOV added the self-learning AI agents in mid-2025 after the earlier model produced false outputs on vendors whose report templates looked unlike anything it had seen before.
NOV tested the agentic system on historical reports from six vendors. One test compared AI-generated prompts against manually written ones for speed and accuracy, producing the 8.8-minute figure. A second test, run on four of the six vendors, checked whether the system kept improving on its own: parsing accuracy rose by an average of 6% as the agents learned from feedback against the human-verified ground-truth database.
NOV published the methodology as IADC/SPE 230772, "Self-Improving Generative AI Agents for Automated Daily Mud Report Parsing."
