SLB

Operational reporting is some of the highest-volume data a drilling operation produces, and some of the least governed. Daily activity remarks are entered by hand under time pressure, in free text, with missing dates, inconsistent depths, merged operations, and activity codes that are frequently wrong or missing. The result is a record that is hard to standardize, hard to compare across wells, and hard to load into a corporate data platform.

We built a set of grounded GenAI agents that turn this raw operational narrative into standardized, OSDU-compliant well operations records, piloted with Equinor FLX across four North Sea wells (a representative well file holds around 175 operational rows spanning roughly a month). The agents reconstruct a coherent chronological daily report from scattered remarks; ground each activity to the correct code within a three-level controlled vocabulary of roughly 3,700 codes (99 main activities, 590 sub-groups) using retrieval and a confidence score; and derive risk-aware performance targets from the statistical distribution of historical execution data instead of subjective benchmarks. The output maps directly to OSDU WellOperationsReport 2.1.0, verified against OSDU reference-data manifests.

The information-management lesson is that text generation is the easy part. The hard parts are binding model output to a controlled vocabulary, quality-assuring records that an AI wrote, and proving they conform to the schema and reference data the business depends on. We share what worked, where output had to be checked and corrected (including a cautionary case of fabricated reference values), and a repeatable path from freeform human reports to governed, queryable records. In the pilot, once the inputs are clean, report generation runs in one to two minutes, against the hours a night supervisor would otherwise spend reconstructing the day by hand.