By Shaun Baker, CTO and Head of Software Engineering, Landmark
When customers describe a data problem, they usually mean they lack information. Repeated conversations reveal a different issue. Data almost always exists, but teams cannot find it, do not trust it, cannot connect it to anything else, or must wait for data clean-up or validation activities that block actionable insights needed today.
That distinction matters because it changes what teams build. Decades of subsurface data sits in silos, locked in vendor formats and departmental systems that never communicate. Some data still lives on paper. Even when organizations store data, they often lose the context that gives it significance. Teams find a well log but cannot identify who produced it, under what conditions, or which decision it supported. Remove that context and the log loses value for serious analysis by human or machine. The gap widens in older assets, which creates tension because legacy assets still produce much of the world's oil and gas. New wells arrive as structured digital records. The wells that matter most for today's production often arrive as fragments.
The industry tends to skip this part: Every measurement, interpretation, and forecast carries uncertainty, and a workflow that presents a single confident number hides the most critical information.
Take a well log. Parts of a curve may contain imputed values or gaps that someone filled in years ago to make the data usable. In better cases, labels flag the imputation, and teams can evaluate it. In harder cases, especially with legacy datasets, nothing marks the imputation. Systems treat decades-old guesses as measured values, and all downstream calculations inherit false confidence without detection.
The central engineering question does not focus on the removal of uncertainty before action. That condition rarely exists. The question centers on how systems represent uncertainty, quantify it, and carry it through every workflow step, so it reaches the recommendation intact. If imputed data supports a recommendation, decision makers must see the discount and understand why. A recommendation without its uncertainty does not accelerate decisions; it introduces hidden risk.
Handled well, uncertainty delivers two benefits. First, it discounts the answer honestly, so decision makers understand its weight. Second, it directs effort. When teams identify which uncertain inputs influence a recommendation, they know which gaps matter and which do not. Uncertainty becomes a prioritized worklist instead of a disclaimer.
The hardest version of this remains unrecorded uncertainty. AI must surface imputations that lack labels, detect changes in units, and identify measurements that deviate from expected behavior. The easier task remains the representation of known uncertainty. The recovery of lost separates robust systems from ones that only appear confident.
Waiting for clean data stalls progress. Many AI projects fail because teams expect pristine data before they start work. That condition rarely arrives.
A different approach builds for uncertainty instead of an attempt to eliminate it. Systems attach confidence measures to every workflow so teams act on imperfect data without misjudgment of its quality. AI also allows practical improvement of legacy data through analysis of large historical datasets, identification of the high-value gaps, and correction of issues that affect decisions.
One principle sets priorities: fix the future first. Teams must govern new data at the point of creation to prevent new legacy problems, then work backward through historical records based on business value. Data quality does not conclude as a project. It operates as a continuous workflow embedded in software.
Automated seismic interpretation reduces analysis cycles from weeks to hours while it maintains transparency. Predictive maintenance reduces unplanned equipment downtime. Automated classification lowers the cost of digitization of legacy archives. Analytics detect production anomalies faster than manual review cycles. Generative models extract insights from engineering reports that sat unused for decades.
AI cannot define bad data alone; domain experts must set that standard. AI can identify inconsistencies, reconcile historical measurements, and correct errors to a defensible confidence level.
Automated interpretations already function as recommendations. Systems now generate more complex recommendations, which include well placement and reserve estimation. Models perform analytical work and present answers with attached confidence. Humans retain responsibility for decisions and outcomes.
Transparency makes this model effective. Systems present recommendations with clear visualization, show the full rationale, and allow experts to inspect each component. When experts change inputs, systems rebuild workflows and recompute results. Experts do not evaluate black boxes; they work within the reasoning and start from quantified recommendations rather than blank pages.
Many organizations modernize data environments while they recreate old issues. Teams store AI-generated output without a record of where it originated, ingest sensor data without consistent tags or units, and migrate systems to the cloud without improvement to the structure.
The solution begins at data creation. Data contracts and governance must live within software so systems enforce quality at ingestion. Systems with governance embedded in the data layer provide consistency.
Value does not come from data, and it does not come from models. It comes from decisions. Organizations must define high-impact decisions and work backward to the data quality each one requires. Teams must prioritize datasets based on value and feasibility instead of an attempt to clean everything.
Small improvements often yield high returns. Consistent naming, units, and tagging unlock downstream capabilities with minimal effort.
Forecasting in E&P now shifts from static estimates to dynamic, probabilistic models, and that shift aligns with the role of uncertainty. As new data arrives, AI updates a reservoir model in near real time and carries revised uncertainty forward. Experts now spend less time on large-scale pattern recognition and more time on judgement.
The prize remains large, but distribution will not occur evenly. Organizations that close data gaps, represent uncertainty honestly, and value proof over promise will capture the advantage.
If this work aligns with your priorities, contact the Halliburton Landmark team to compare what works today and what will deliver results tomorrow.
Halliburton Landmark, provides E&P professionals with software-driven lifecycle insights that generate new ideas, actions, and results to maximize asset value
