This hands-on ECIM 2026 workshop is designed for data managers, data professionals, and energy domain experts who want to work smarter, faster, and with far more confidence in OSDU. Building on the success of previous OSDU bootcamps, we are raising the bar — combining practical workflows, live demos, AI-assisted “vibe coding,” and real industry use cases that solve everyday challenges.

Over three energetic hours, participants will explore how modern tooling and AI-driven workflows can dramatically improve data quality, accelerate validation processes, and simplify interaction with OSDU services and external APIs.

Data management professionals are fluent in the how of the discipline: catalogs, lineage, quality rules, contracts. We are often less fluent in the why. We tell stories about pipelines and platforms when sponsors need stories about production regularity, reporting confidence, safety, maintenance execution, and trust in operational decisions. The result is familiar: work is treated as technical overhead instead of a business capability. This three-hour, hands-on bootcamp closes that gap. Drawing on best practice in Storytelling, stakeholder management, and visualization, participants learn to translate data work into language that executives, domain leaders, and technical teams can act on. We move from tool-first descriptions to outcome-based narratives that state the operational or reporting problem, the decision at stake, and the measurable result. We map the room with the power-interest grid, identify the three sponsors every initiative needs, and draft a one-paragraph case a sponsor can repeat to a peer. Every module is interactive. Tables and pairs work, pressure-test, and role-play. Participants leave with a drafted Minimum Viable Sponsorship package, a tested pitch, and one concrete next conversation to run in their own context.

The Open Subsurface Data Universe (OSDU) promises a single, open, vendor-neutral home for subsurface and well data — but turning that promise into a working reality inside an operating company is a journey, not a switch you flip. In this presentation, Aker BP shares an honest account of our OSDU adoption so far: the wins worth celebrating, the obstacles we hit, and the work still ahead of us.

We will walk you through where we are today — what we have deployed, which data domains we have onboarded, and how OSDU is starting to change the way our teams find, trust, and use data. We will be open about where we struggled: the effort of data ingestion and mapping, aligning legacy systems and ways of working with OSDU's data model, governance and ownership questions, and the organizational change that technology alone cannot solve.

Looking forward, we will share what is on our roadmap — the capabilities, data types, and use cases we are prioritizing next — and what it will take to get there. Beyond our own four walls, we will reflect on our collaboration with other companies in the OSDU Forum: what works well in working together on a shared platform, what we have learned from others, and where we would like to see the community and the standard improve.

High-quality data products rarely start in the data platform, they start in the operational systems where data is created. Many modern platforms struggle not because of technology, but because of weak upstream data modeling and the lack of of metadata.

Data models and metadata together form the foundation. Models give data its structure and meaning. Metadata gives it context — legal basis, security, quality, ownership, and the full range of properties needed to govern data responsibly. Modeling has its own discipline, but the broader work of metadata across data quality, legal, security, and ethics requires organization through an operational model. "At the Norwegian Tax Administration, we complement the modeling foundation with an operational governance model: data leaders (stewards), legal expertise, and analytical capability are placed close to the domains and their data products, supported by structured collaboration around domain knowledge.

Together, the foundation and the organization turn upstream investment into downstream reuse, automation, and interoperable data products that consumers can actually understand and use.

Key takeaways:

- Information models and metadata together form the foundation for interoperable, well-governed data products

- Modeling discipline gives structure and meaning — semantics are more stable than technology

- Operational governance models matter — data leaders, embedded expertise, and collaboration around domain knowledge

The next competitive advantage in the energy industry will increasingly depend on companies’ ability to collaborate through shared digital infrastructure, common standards and governed data exchange. CoOper8 was established by the industry to provide the digital backbone for this collaboration, delivering shared services that enable trusted information exchange, interoperable workflows and more efficient use of resources across company boundaries.

Drawing on practical examples from equipment information, virtual inventory and project planning, this presentation explores how CoOper8’s industry-wide services are reducing manual work, improving data quality and strengthening operational resilience on the Norwegian Continental Shelf.

Åshild Hanne Larsen is seconded from Equinor into the role as CEO of CoOper8 AS. She is an award-winning technology executive with broad experience as global CIO, VP HR and VP Subsurface Excellence & Digital. She has successfully led corporate-wide IT & digital transformation programs and large-scale organizational change projects, resulting in extensive leadership experience. Åshild holds several board roles and is also widely recognized as a speaker, thought leader and advocate for diversity and inclusion in tech. She holds an MBA from Heriot-Watt University and an MA from Universität Bielefeld.

It wasn’t too long ago when we were working with, or better said: constrained by the adage of: Garbage in, Garbage out. Recently I came across that again, but now in the context of modern day AI: “with all the AI, we still deal with Garbage In: Garbage Out”

We live in an age where economies become circular and a world player like Renewi uses the “Waste No More” in their company strap-line. And with modern AI, that is true for data as well.

In this talk we’ll talk about how, with these days of AI, Data Management has fundamentally changed. How we can actually embrace the data we used to call garbage, but were afraid to throw away “just-in-case” For this purpose we will dive a little bit into the following topics:

• What is data management in this age of generative AI?

• Does the adage of Garbage-in and Garbage-out still hold?

• A little bit of data-architecture to help you out

* A little bit of OSDU Data Standards in the mix as well

• And some cool projects you can download and run for yourself to become the AI enabled data manager

• And finally: how you too can stop worrying and love your data

In this panel, leaders from across the energy, technology, and public sectors will discuss if data quality is no longer the barrier it once was – exploring whether AI has truly solved data quality, or simply scaled it up. Expect strong opinions and different perspectives on trust, governance, and what success looks like in an AI-driven world.

National Data Repositories (NDRs) have long served as trusted systems for preserving, governing, and distributing subsurface data. For decades, success was measured by the ability to ingest, manage, and provide access to growing volumes of information. Today, a new opportunity is emerging. As artificial intelligence, cloud-scale computing, and standards-based data ecosystems mature, the NDR can become more than a repository. It can become the foundation for a new generation of data-driven workflows, analytics, and decision support across the energy industry. This presentation explores the evolution of the Norwegian National Data Repository, DISKOS, and examines how modern data architectures can transform national repositories from systems of record into platforms for intelligence. Drawing on operational experience from managing one of the world's largest subsurface data repositories, the presenters will discuss how trusted data, common standards, and scalable architectures create the conditions for advanced search, automated quality control, digital workflows, AI applications, and future agent-based systems. The session will also examine the strategic value of NDRs beyond regulatory compliance and archival storage. As energy companies seek to accelerate exploration, improve recovery, support carbon storage initiatives, and reduce operational risk, national repositories are uniquely positioned to provide a shared foundation for innovation across the industry. Attendees will gain practical insights into the technologies, governance models, and architectural decisions that can help transform NDRs from passive data stores into active enablers of industry intelligence.

AI has moved from a speculative technology to a strategic priority across the energy industry. Conversations with industry leaders reveal that the debate is no longer about whether to adopt AI, but how aggressively to deploy it. Amidst the shift, two cohorts are emerging. The first group (the AI incrementalists) are applying the technology to improve and optimize existing workflows. The second group (the AI revolutionaries) are using it as a catalyst to fundamentally redesign how their organizations operate. Companies that are best positioned to capitalize on AI’s transformative potential will be those that adopt a portfolio approach: pursuing near-term efficiency gains while simultaneously investing in longer-term operating model reinvention. The self-improvement capabilities of AI mean the traditional “fast follower” strategy will likely not work as it has in the past. Early adopters are already building a competitive advantage that slower-moving companies will find increasingly difficult to overcome.

Are we trying to run before we can walk? While AI offers vast opportunities for workflow efficiencies and enhanced insights, it needs to be built on a sound data foundation. Subsurface data are difficult and expensive to acquire, manage and store. To maximize the return on these investments, we need to ensure that the right data are at our fingertips; whether we’re running traditional subsurface workflows or complex machine learning algorithms, data is currency. AI adds to the jeopardy of poorly managed data, owing to its “black box” nature. We can rapidly build models and make decisions based on vast quantities of data that may never have been curated or verified. As we know it’s “garbage in, garbage out,” so the consequences could be disastrous. To mitigate this risk and safely leverage the power of AI, we need a subsurface data ecosystem that is fit-for-purpose — the right data, in the right place, deduplicated, curated. We can reach this goal by first scanning all data repositories to identify issues, then building a data strategy suited to our aims, and finally using a tech-driven approach to implement this accurately and efficiently. Only then can we fully embrace the value of AI without fear of the consequences.

The SEG-Y revision 2 specification opens up new possibilities for seismic data exchange. This presentation draws on hands-on experience implementing the specification, illustrated with practical examples and tips for adding software support. Along the way, we'll examine the benefits the latest revision brings — and the challenges it introduces.

LUMIN, Kadme's data and insights engine for the oil and gas industry, has expanded significantly over the past year — moving beyond search and data discovery into active quality assurance workflows and AI-native data access. This presentation covers two major additions to the LUMIN product suite.

The first is the Well Log QC module, now in active deployment for a major national oil company in Latin America. The module automates quality control of well log data (in LAS and DLIS formats) by applying a configurable set of business rules that validate completeness, range compliance, curve consistency, and cross-well integrity.

The second highlight is the LUMIN MCP Server. MCP works as a universal connector between AI clients and data sources. The LUMIN MCP Server exposes LUMIN's lexical search, semantic search, and API query capabilities directly to any MCP-compatible AI agent, without custom integrations.

Together, these developments reflect Kadme's direction: a platform where data quality and AI accessibility reinforce each other at every layer of the subsurface workflow.

In this talk we will look at what is required for the digital transformation of seismic data and how this can be achieved and why we have not got there - yet.

The O&G industry generates vast amounts of data across the exploration, development, and production lifecycle. However, transforming this data into actionable insights and business value still remains a significant challenge. Many analytics initiatives never progress beyond the Proof of Concept (PoC) stage. Success requires more than advanced analytics or black-box models; it demands a deep understanding of the underlying physics, engineering constraints, and operational realities that shape decision-making in E&P.

In this presentation, Expert Analytics will showcase a series of real-world customer use cases in production optimization, demonstrating how industrial data science can bridge the gap between raw data and operational impact. By combining domain expertise, physics-based understanding, advanced analytics, and robust software engineering, we help E&P operators tackle complex challenges and accelerate the adoption of data-driven solutions.

While operators embrace OSDU as an open data platform to liberate data for new insight and application interoperability, the journey is taking time and effort from scarce expert resources. Petrabytes have developed an customer-adaptable OSDU data ingestion platform with a Lakehouse Architecture, deployable in any cloud or on-premise environment. The presenters will share lessons learned from projects and how the approach follows a Medallion Design Pattern to provision subsurface data as products. While OSDU is a destination as new System of Records, the approach also allows for synergies though “dual destinations” to analytics platforms, like Databricks and Fabric.

A subsurface data-pipeline is a business-critical element in your data foundation. Reliability may come from design, but availability and serviceability requires management and attention. The presentation will illustrate a concept of Observability which allows for automation of remedial actions if data sources, ETL, enrichment or provisioning is indicating a problem.

* 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.

This project transforms Subsurface Data Management from a manual, service-intensive model into an automated, selfservice environment, balancing strict governance with operational accessibility. The scope includes deploying a frontend visualization platform integrated with a backend architecture to optimize data discovery while maintaining stringent security protocols.

The approach utilizes a visualization tool equipped with AI-driven search and summarization. This interface reduces time spent on data collection during subsurface studies, supported by a backend integration layer ensuring continuous quality assurance. Implementation is phased, beginning with strategic assets selected for data maturity. Subsequent phases incorporate Joint Venture and corporate subsurface data, evolving the system into a unified one-stop channel for all subsurface information.

A commercial platform was selected to ensure long-term sustainability, as internal solutions often struggle to keep pace with rapid technological shifts. Implementing this solution requires balancing standardized features with dynamic user requirements. While the project is ongoing, implementation demonstrates that transitioning data services requires significant investment in personnel, time, and workflow re-engineering; shortcuts create systemic hurdles.

While utilizing data visualization for data services is not new, its implementation within PTTEP’s heterogeneous data ecosystem is innovative. PTTEP’s landscape reflects decades of diverse E&P practices. Integrating and standardizing these varied ecosystems into a unified system governed by strict protocols requires creative problem-solving in a challenging environment—providing valuable lessons for future digital transformation.

This presentation details the implementation of the Onecheckshot methodology within subsurface data management at Equinor, focusing on the principles, governance, and practical challenges of managing checkshot data. The project is a joint effort between two teams within Exploration and Production Norway unit to deliver controlled quality data using an automatic or semi-automatic approach in the cloud, ensuring compliance with technical management criteria.

The Equinor data governance model is highlighted as a federated approach, distributing responsibilities across 13 specialized data offices and one enterprise data management group, with the Subsurface Data Office playing a central role in legal risk assessment, data cleanup, and reporting standardization. Onecheckshot is an example of how we can define systems to deliver products following this high-level model, while providing data in a more efficient way than traditional and error-prone manual methods. It addresses issues such as non-standard file formats and inconsistent quality procedures to provide approval or failure of specific quality control rules, which can be used to pinpoint the error for a specific checkshot file.

Key strategies for successful implementation include effective user onboarding, transparent communication, and continuous feedback mechanisms. The approach has resulted in improved data accuracy, increased efficiency, faster decision-making, and cost savings. By centralizing data in the cloud, advanced algorithms can refine sonic logs and support seismic imaging. The presentation concludes that effective partnerships, tailored solutions, and clear governance documentation are essential for scalable, high-quality data products and data-driven culture in subsurface operations.

To improve the quality and usability of legacy Hydrocarbon Compositional (PVT-HC) data, this project utilized an AI/MLassisted extraction approach to transform fragmented, unstructured datasets into standardized, simulation-ready intelligence for reservoir engineering and analytics. The workflow employs an intelligent cognitive data wrangling platform, beginning with document assessment and the development of data dictionaries to standardize nomenclature. A multi-stage pipeline enhances OCR quality through pre-extraction image processing, while Large Language Models (LLMs) and deep learning algorithms parse complex tabular data from unstructured PDF reports. Crucially, a "Human-inthe-Loop" quality control engine allows Subject Matter Experts (SMEs) to verify outputs, harmonize inconsistent units, and resolve identity mismatches by cross-referencing physical metadata against digital records. Applying this framework to over 200 high-priority reports resulted in 100% extraction and validation completion. The system successfully remediated critical legacy errors, such as mole percentage discrepancies, and resolved complex "one-to-many" mapping issues by correctly re-mapping generalized terms to specific test stages. Furthermore, a novel cylinder mapping strategy links physical samples to digital records even when metadata is missing, recovering previously unusable datasets for practicing engineers. Ultimately, the integration of AI for accelerated bulk processing with rigorous SME validation ensures traceability, demonstrating that upfront standardization significantly minimizes downstream rework and provides highly reliable, audit-ready subsurface intelligence.

As organisations accelerate their adoption of artificial intelligence, many are discovering that the true barrier to value creation is not the technology itself, but the absence of structured, trustworthy and accessible data. While AI tools are increasingly capable of generating high-quality insights, their effectiveness remains constrained when enterprise data lacks governance, context and ownership across organisational domains.

This talk explores how data governance can become a strategic differentiator in the age of AI. Building on the concept of data products, it argues that well-governed, certified data can be treated as an internal knowledge foundation— providing AI systems with reliable, contextualised information comparable to continuously updated, organisation-specific research.

Rather than focusing on technology, the session highlights the importance of establishing cross-functional processes that ensure data quality, clarity of ownership and shared understanding across business domains. The presentation will provide a practical perspective on how organisations can transition from fragmented data practices to a governed, product-oriented approach—enabling scalable and trusted AI adoption. Ultimately, it will demonstrate how the ability to embed data governance across the enterprise may define the next generation of competitive advantage.

Data stewardship is the foundation behind automation and digital twins Large unstructured data requires structured governance and access control Scripted QC ensures consistent, measurable data quality Contractor QC and master/reference data are critical to data integrity.

Stuck pipe events remain among the most critical and costly drilling dysfunctions in modern well construction operations. Early detection is difficult due to the highly dynamic and noisy nature of real-time drilling data, along with operational behaviors that may appear anomalous while remaining physically valid. This work presents a physicsinformed anomaly detection framework for real-time stuck pipe surveillance using multivariate drilling sensor streams and temporal deep learning.

The workflow combines drilling engineering expertise with data-driven analytics through preprocessing and feature engineering pipelines. Real-time drilling data are standardized, resampled, validated, and filtered using physics-aware quality rules designed to distinguish corrupted sensor behavior from legitimate operational phenomena such as negative Weight-on-Bit (WOB), hookload variations caused by frictional effects, and transient torque fluctuations. Additional surrogate features, including rotary energy proxies, torque-speed penetration ratios, drilling resistance indicators, hookload deviation metrics, and transient drilling response features, are generated to better characterize drilling system behavior.

An LSTM Autoencoder is trained exclusively on normal drilling operations to learn healthy temporal drilling patterns without requiring abnormal event labeling. Abnormal drilling conditions, including stuck pipe precursors and drilling dysfunction signatures, are identified using reconstruction-error-based anomaly scoring. Model robustness is evaluated using blind-well validation, where unseen wells are used for threshold calibration and performance assessment.

Initial validation across multiple wells demonstrated promising results, successfully identifying abnormal drilling behavior while maintaining low false-alarm tendencies under varying operational conditions.

Large Language Models (LLMs) have shown strong generative capabilities for question answering, yet they frequently struggle to provide trustworthy responses in highly specialized technical domains where factual precision and relational consistency are critical. In the context of oil and gas, conventional RAG approaches often rely either on purely symbolic graph querying or on dense semantic retrieval, each presenting limitations when addressing semantically underspecified or multi-relational questions. In this work, we propose Cypher-Guided Semantic RAG (CGS-RAG), a two-stage graphenhanced retrieval framework that combines Cypher-based structured candidate retrieval with embedding-driven semantic evidence refinement for trustworthy LLM answer generation over domain knowledge graphs. CGS-RAG first constrains the search space through explicit graph queries generated from the user question, and subsequently performs semantic reranking over the retrieved candidate evidence before constructing the final grounded prompt. To evaluate the proposed framework, we introduce a domain-specific benchmark of 2,046 technical questions over an oil and gas knowledge graph. We compare CGS-RAG against a purely Cypher-based GraphRAG baseline using exact match, F1-score, and RAGAS-based grounding metrics.

The verification of compliance with regulatory standards for well plug location during the abandonment of oil wells is a tedious process prone to human error since it involves matching various configurations of a well against multidimensional and complex regulatory requirements like NORSOK D-010. In this paper, an automated compliance checking mechanism is described which leverages the application of large language models (LLMs) and workflow orchestrations to perform the above task efficiently and accurately.

Our proposed system uses a document intelligence pipeline together with an interactive compliance checker. The pipeline dissects the regulatory standard using extraction nodes to identify well zones, classify well types, define global constraints in terms of fluid mud weight, pressure constraints, fluids, and specific plug placement requirements for each well type and well zone.

The information gathered from the pipeline is then compiled in a structured form in a normalised reference table which includes six well zones, three well types, and three categories of plugs. This table is fed into a compliance checker which takes the structured data of a well, queries an LLM for compliance and returns either COMPLIANT, PARTIAL or NONCOMPLIANT verdicts.

For handling output reliability issues, the solution uses a two-tier assessment approach, where the first layer entails a rules-based approach for assessing facts by extracting numeric statements from the input prompt and validating them with ground truth well data, with suitable tolerances, whereas the second tier involves LLMs acting as judges to assess the consistency of facts and contradictions.

This architecture demonstrates how prompt chaining, structured state management, and self-supervised output evaluation can be combined to operationalise regulatory AI in a safety-critical domain, reducing manual review burden while delivering interpretable, auditable compliance outputs.

Field development planning (FDP) requires coordinated decision-making across multiple interdependent variables, including whether to drill, where to place wells, and which well type best suits prevailing geological and economic conditions. Traditional simulation-optimization workflows are computationally expensive, reliant on expert judgment, and poorly equipped to handle the combinatorial complexity of development scenarios. There is a growing need for intelligent, adaptive frameworks capable of learning robust strategies across diverse operating conditions.

This work presents a deep reinforcement learning (DRL) approach in which an autonomous agent is trained on a synthetic dataset to simultaneously optimize three coupled FDP actions: the decision of whether to drill, spatial selection of drill location, and well type classification. By framing FDP as a multi-action sequential decision problem, the agent learns a unified policy balancing short-term drilling costs against long-term NPV maximization.

Training results over 100 iterations demonstrate consistent reward improvement, with the agent progressively refining its policy from exploratory toward high-confidence, value-maximizing behavior. Evaluation of the converged agent reveals it elects to drill 99% of the time, indicating a learned policy that favors aggressive field development as the dominant value-creation strategy. The synthetic dataset enables controlled benchmarking, isolating algorithmic performance and ensuring reproducible evaluation. Results confirm that DRL can effectively navigate high-dimensional FDP action spaces, offering a scalable alternative to conventional planning workflows

Decline Curve Analysis (DCA) remains one of the most widely applied techniques for forecasting well and field production performance, as it depends on historical production data and is based on formulations introduced by Arps. Despite its theoretical robustness and long-standing industry adoption, conventional DCA suffers from several limitations. It is unreliable for wells with limited or no production history, and the calibration of its parameters typically requires expert judgment, introducing subjectivity and limiting scalability and consistency across large asset portfolios. This paper presents a fully automated hybrid workflow designed to address these challenges and extend the applicability of DCA across all development stages with minimal human intervention. The proposed approach integrates analytical methods, machine learning (ML), and an artificial intelligence (AI) agent to standardize and accelerate DCA parameter estimation and validation.

The workflow consists of four sequential steps. First, decline parameters—including initial production rate, nominal decline rate, and the hyperbolic exponent—are calculated for wells with sufficient production history within the same field or from geologically and operationally analogous fields. Second, these parameters are used to train machine learning models capable of capturing relationships between reservoir, well, and production features. Third, the trained models are applied to estimate DCA parameters for wells with limited or no historical production data. Finally, an AI agent is deployed to automatically assess forecast confidence levels and generate actionable engineering insights and development recommendations without reliance on subject-matter expert review.

The proposed hybrid framework enables consistent, unbiased, and scalable DCA parameter estimation, significantly reducing analysis cycle time and enhancing decision-making efficiency for field development planning and production forecasting.

Scope: Using an AI-enabled, end-to-end Field Development Planning (FDP) framework accelerates scenario generation, simulation and decision-making. A domain-trained Small Language Model(SLM)automates simulation decks, proxy modelling, uncertainty analysis and together with human-in-the-loop decision support provides an AI-assisted FDP framework reducing cycle time and decision latency, enhancing analytical confidence and providing a scalable foundation for multi-asset deployment.

Methodology: The technical approach uses a domain-trained SLM integrated with existing reservoir simulation environment in a pilot phase. Initially static, dynamic and FDP knowledge assets are ingested into the SLM followed by fine-tuning on simulation deck syntax, field history and engineering workflows. AI agents automate simulation deck creation, QA/QC, scenario generation and proxy model training. Uncertainty analysis and AI-guided optimisation are executed iteratively with outputs validated by human-in-the-loop workflows ensuring explainability, traceability and engineering acceptance prior to scaling-up.

Conclusion: An AI-enabled FDP workflow can deliver substantial improvements in speed, scale and decision quality compared to traditional FDP methods. Domain-trained SLM deployment will enable 80% simulation deck preparation time reduction, significantly shortening model setup and QA/QC cycles that typically take weeks. Automated scenario generation and proxy modelling can increase scenario throughput by 10-20 times allowing evaluation of a broader range of development alternatives and uncertainty combinations within practical timelines.

Upstream capital projects often experience extended Front-End Loading (FEL) timelines due to fragmented workflows, sequential decision-making, and high levels of non value added effort across market analysis, concept development, engineering definition, estimation, and governance approvals. This work in progress paper explores the design and early validation of an AI-enabled, end to end capital project delivery framework intended to reduce the FEL-to-FID cycle time while improving decision readiness and predictability.

The study applies a workflow-first, data-driven methodology. Capital project processes spanning FEL0-FEL3 are decomposed to BPW level and analyzed using Process Cycle Efficiency (PCE) to distinguish value-added from non-value-added time. Identified bottlenecks-such as approval latency, manual reporting, data retrieval, and stakeholder handoffs are used to define targeted AI intervention points.

A modular multi-agent architecture is under development, with specialized agents supporting interconnected activities including market signal structuring, concept option generation, early engineering artifact drafting and standards validation, CAPEX/OPEX and schedule estimation, economic scenario screening, automated FEL decision-pack assembly, and KPI-based execution readiness tracking. By enabling greater parallelization and traceability across workflows, the framework is hypothesized to materially reduce FEL cycle time and increase design maturity at decision gates. Initial observations from analogous upstream delivery processes suggest meaningful improvements in process efficiency and lead-time reduction; however, the framework remains under refinement. Ongoing work focuses on agent orchestration, governance integration, and data quality dependencies. The contribution is an integrated blueprint linking process efficiency measurement with AI-assisted, standards-aware project workflows to support faster and more confident investment decisions.