AI In Business Workflow (Tuesday) and Regulatory & Compliance (Wednesday)

Decision-ready is a property you compute, not a label you assign.

Here is the gap. Data quality is measured against completeness, consistency, and standards. Decisions are made against physics and operating constraints. Data that passes every catalog check can still carry an assumed operating range, a silently substituted default, or a stale calibration that no dashboard flags. When AI turns that data into a recommended setpoint, the failure surfaces in production, and the digital solutions team gets the call and wears the blame.

Most companies built the common data foundation: data is findable, standardized, and shared. Yet applications on that foundation still stumble in operations. What we can show, measured on live assets, is the standard applied to one decision at a time: recommendations verified against the physics before they reach the field, with their validity and their uncertainty stated rather than assumed. This talk walks two deployments, each with a different unfit source, then states plainly what must become platform infrastructure before the organization follows.

On the first deployment, a production network of several hundred ESP-lifted wells, the unfit source was the field data itself. The optimizer and its audit pipeline surfaced what nobody knew was broken, because searching for the best operating point and searching for what is wrong are the same search. Wells assumed to run no lower than 40 Hz were running at 38. A missing gas-oil ratio was filled with a default of 500 against a true value of 81, a substitution that misled our own engineers twice. Requests for nineteen wells were answered from outside their trained ranges with no signal to anyone. We fixed data the client did not know needed fixing, and the fixes hardened into rules: per-well operating envelopes from the production data system before training, no defaults, and refusals that name the well, the input, and the valid range.

On the second deployment, a live water injection network, the unfit source was the simulator itself. A simulator is a data source, and almost no data team treats it as one. It generates the training data and carries the calibration, yet it gets no completeness score, no freshness check, no fitness assessment. On this asset it could not reliably converge and pinned outputs at one atmosphere, a floor a model would have learned as physics, and a solve reported converged while its mass balance was off by 36 percent. It failed the fitness check any governed source would face, so we replaced it with a purpose-built physics solver: 0.12 seconds per batched solve, 1.9 percent rate error against the converged reference. In a controlled comparison on this network, the same optimizer with the same seed broke an erosion limit by 2.8 percent against a trained model and shipped a feasible optimum through the physics solver.

Both deployments share the runtime discipline. Every answer declares whether any input sits outside its trained range. Conservation is checked at run time by substituting the answer back into the balance equations, gated at 2 percent. Audit gates run at every pipeline stage, and a model set that fails cannot go live. And uncertainty is quantified, not asserted: on the injection network, Bayesian inversion cut median pressure error from 20.1 to 5.7 percent with no refit, stated what wellhead data cannot identify, and ranked the next measurements by the decision they improve.

Here is the distance that remains. In these deployments decision-ready AI models held because our engineers enforced it by hand. It becomes an organizational property only when two capabilities move from people into platform. First, evergreen models: regeneration triggered by evidence from new well tests, run by the operator's own engineers through audit-gated, versioned pipelines. Second, governance by design: a generated application inherits the refusal semantics, validity declarations, and lineage instead of reimplementing them. That layer is in build, not yet deployed. We will report what breaks.

Decision-ready is a number you check. An organization is decision-ready when those numbers are checked by default.

Plug & Abandonment (P&A) represents a major technical, operational, and regulatory challenge for mature oil and gas provinces such as the Norwegian Continental Shelf, where large inventories of aging wells must be safely decommissioned under stringent environmental and regulatory requirements. A key obstacle to efficient P&A execution is the limited accessibility and quality of historical well information, which predominantly exists in unstructured legacy formats, including scanned reports and handwritten notes stored in heterogeneous file types (PDF, Word, PowerPoint). These limitations extend planning timelines, introduce uncertainty, and increase the risk of costly reabandonment. This paper presents a Generative AI driven solution that transforms legacy well documentation into a structured, engineering grade digital data foundation built for P&A workflows. The approach enables automated extraction and standardization of P&A relevant data and supports generation of consistent digital well schematics to inform risk based P&A design and regulatory compliance.

Extracted data is curated, quality controlled, standardized, and ingested into subsurface and engineering data platforms, ensuring traceability to source documents and enabling automated generation of P&A ready digital well schematics. The methodology was validated through a P&A case study involving more than 12,000 pages of multilingual legacy documentation across 9 wells from the Netherlands and the Gulf of America. Tasks that previously required 3 to 4 weeks of manual effort per well were completed in less than 2 days while maintaining high extraction accuracy. For operators and regulators, this GenAI enabled approach supports faster historical well assessments, earlier identification of integrity risks, standardized P&A decision making, and increased confidence in regulatory compliance, while converting legacy well data into a reliable, engineering grade foundation for risk based P&A execution.

Using Google Cloud's Gemini models with the Open Subsurface Data Universe (OSDU) platform.

* Multi-Agent Offset Well Analysis: This AI architecture is used to predict hydrocarbon production and run optimization scenarios for future field development based on multi-disciplinary datasets (geoscience, well design, operations). It uses a multi-agent architecture to orchestrate sub-agents: an Analytics Agent for data analysis and summarization, a Plotting Agent for generating Python data visualizations, an ML Agent for making predictions and evaluating ML models, and an Optimizer Agent for optimizing against constraints.

* Strata Scanner: An AI-driven multimodal intelligence solution designed to digitize dormant, unstructured historical data—such as well log images in .TIFF format—into modern OSDU formats in minutes.

* Automatic Schema Transformation: An interactive, conversational AI tool that helps users pipeline semi-structured database data into OSDU formats.

* Agents for Search & Automation: The integration deploys agents grounded in OSDU enterprise data to unlock conversational data access, allowing users to efficiently summarize complex technical documents, analyze large datasets, and automate workflows across the OSDU ecosystem.

Critical information for the evaluation of carbon storage prospects is held in the well abandonment documentation for historic wellbores. But extracting that information is a slow, and laborious exercise. Can LLMs do it better, faster, and cheaper? The NSTA and Flare Solutions have been working together to find out. We will share the outcome of work across over 1,000 well abandonment documents to extract cement plug and casing cutting parameters critical to understanding store integrity risk and the possibility of wellbore remediation. We'll also share what we've learned along the way about using commodity LLMs to answer domain-specific questions quickly, cheaply, and reliably.

Offset well analysis is one of the most valuable yet time-consuming workflows in drilling. Engineers often spend days or weeks selecting relevant analog wells, reviewing daily drilling reports, completion reports, end-of-well reports, and other semi-structured records, then manually translating that history into planning actions. This paper presents an AIaccelerated offset well analysis workflow that moves beyond data extraction to analysis and operational decision support.

The approach combines clustering-based well similarity methods with generative AI applied to unstructured and semistructured drilling data. Structured well attributes are first used to identify the most relevant offsets for a new plan. Large language models with OCR and document parsing then analyze drilling narratives to extract and interpret hazards, operational sequences, lessons learned, formation behavior, NPT drivers, losses, instability indicators, and mitigation actions. Rather than only producing structured fields, the workflow builds a customizable driller’s roadmap: a practical, well-specific summary of expected risks, depth-linked events, recommended mitigations, and execution guidance for the next well.

This reduces review time from weeks to days, improves consistency, and unlocks learning from large historical well populations that are too large for manual review. The paper will also present case studies showing how the workflow helps identify stuck-pipe risk, loss-zone patterns, and offset performance trends, demonstrating how generative AI can turn legacy well records into operational intelligence for faster, smarter, and safer well planning.

The study presents the development of an AI-based agent and software system designed to extract data from geothermal well reports and perform well nodal analysis. The main goal of the project was to create a tool capable of processing technical documentation (containing text, images, tables) and converting them into structured data used for optimizing the production of geothermal wells.

The system utilizes a Retrieval-Augmented Generation (RAG) approach combined with a Chroma vector database to efficiently retrieve relevant technical information. To ensure high-quality data extraction, a pre-trained YOLO model is used to filter images, focusing OCR efforts on relevant schematics while excluding unnecessary noise such as company logos. The AI agent identifies and extracts key parameters, including well trajectories and casing details, transforming them into a structured JSON format.

A dedicated nodal analysis module allows users to evaluate inflow and pressure conditions (IPR/VLP). The workflow is semi-automated, allowing engineers to review and adjust the extracted data before running calculations. This humanin-the-loop design preserves QA/QC reliability while significantly accelerating the transition from raw documentation to actionable insights.

The solution was developed during the SPE Europe Energy Geohackathon 2025, where it was awarded second place. The results demonstrate that the system effectively reduces manual workloads and supports more consistent analysis of geothermal assets, proving the practical value of combining Agentic AI tools with traditional engineering workflows.

This presentation introduces the Data Delivery Plan (DDP), a tool developed by Equinor to improve planning and reporting of wellbore data. The solution provides a structured overview of planned data acquisition aligned with regulatory requirements (Blue Book Table A-1) and the drilling programme, enabling consistent and compliant reporting.

The DDP application allows users to define, review, and export a complete data acquisition plan through a guided workflow covering wellbore information, dataset selection, and final validation. It ensures visibility of mandatory datasets and supports improved completeness, transparency, and communication across operators, service and log QC providers, Diskos and authorities.

Two implementation options are presented: a lightweight version for creating and exporting plans locally, and a more advanced version with user login, storage, and dashboard functionality for managing multiple plans. Overall, the DDP aims to standardise data planning, reduce errors and gaps in reporting, and enable more efficient, consistent, and compliant data delivery to authorities and internal systems.

Over the last 2 years, we have partnered with the NSTA to audit, reclassify, and re-tag nearly 1 million well data items. We will share what we learned about running a mega-scale classification project, including approaches to automation, quality control, and standardising approaches to classification across a diverse project team. During the project, the AI revolution started. We'll also share our thoughts about the impact AI has had, and will have in classification and taxonomy management.

Is Diskos an enabler for new developments and analysis of the Norwegian Continental Shelf, or are we just more of the same... Status of the now more than 31 years old colaboration project. And, maybe some interesting news on new ways of utilizing the vast amout of data and information that has been collected and shared for decades. Everybody wants to create value - can we help?

This presentation shows how large language models (LLMs) can be used to navigate complex data on the Norwegian Continental Shelf (NCS). Central to this is the NOD FactPages, available through REST APIs/Data serivce, which provide structured data.

The data service offers rich metadata describing the data itself and rich metadata about the relationships between tables and data is provided. This makes it possible for LLMs to act as an interface, where users can ask complex questions and get answers based on AI combining relevant data through "navigating" the APIs.

Sodir, together with FabriqAI, has applied this approach in the Knowledge Navigator. This solution uses LLM agents to retrieve and combine information from all public NOD data sources, both structured and unstructured, Tthe FactMaps API and data service forming a key backbone to navigating . To the right unstructured data/information, if the Factpages is not itself sufficient to provide an answer.

Good metadata provided in a format that is AI-native through well-described APIs is grounds for value creation by navigating and receiving relevant information using LLMs.