End-To-End Petrophysical Operations with Agentic AI: A Horizontal Carbonate Reservoir Well Case Study from a Mature Onshore Field, Abu Dhabi

Abstract

Abstract Petrophysical work follows a well from planning through drilling and into completion, and at each step it usually passes between different specialists, vendor software, and email threads. This paper describes an in-house agentic artificial-intelligence assistant that brings those steps together behind one interface driven by text and voice. The system was validated end to end on a horizontal development well, denoted Well A here for confidentiality, drilled through a heterogeneous carbonate reservoir in a mature onshore field in Abu Dhabi. The assistant runs a fleet of nine configuration-driven skills grouped into two co-equal suites. A planning suite covers a well-data registry, document compilation, logging-program generation, cement-bond evaluation, and a logging-activity dashboard. An operations suite covers a vendor-aware quick-look interpretation pipeline, a production-log checklist automator, a drilling-correspondence router, and an inbox-triage dashboard. The skills chain deterministically: the daily geosteering report is parsed automatically, and the picked subzone tops seed per-zone Archie parameters straight into the log computation. On the case-study well, the planning package was assembled in about five minutes against a typical half-day of manual effort. During drilling, the cycle from a raw real-time log email to a delivered evaluation deck fell from roughly ninety minutes to under four minutes on a cached, scheduled run; the cold-start cycle on a log arriving from the rig for the first time, which is the one that governs how quickly the steering team sees an answer, ran between five and ten minutes. Automated subzone tagging let porosity and saturation cutoffs track a lithology change across adjacent subzones within the same shift, which supported steering decisions. Cement-bond and production-log preparation were completed within the shift the source data arrived. Reliability came from deterministic engineering beneath the language model: the evaluated interval was assembled from separate logging runs across two contractors, and porosity was computed per sample through a two-tier cascade of density-neutron porosity with a sonic fallback. The pipeline output was checked against an independent computer-based interpretation built in commercial petrophysical software on the same well. The petrophysical result is reported as computed, with the well name, reservoir zonation, and depths withheld for confidentiality.