Field Signal · Upstream AI · Published October 2026
AI in upstream oil & gas: where intelligence becomes operational control
The important upstream AI boundary is no longer between “AI” and “no AI.” It is between systems that interpret operations and systems that can change them. That boundary changes the architecture, assurance and human authority required.
Upstream AI is becoming a control-system problem
Upstream operations have used numerical models, optimization, automation and machine learning for years. Reservoir simulation, seismic processing, drilling optimization, predictive maintenance and production surveillance did not begin with generative AI. What is changing is the degree to which multiple forms of intelligence can now be connected to real-time data, operational workflows and execution systems.
In March 2026 Halliburton and ExxonMobil reported fully automated geological well placement in Guyana, combining automated subsurface interpretation, well placement, real-time hydraulics and rig automation. In July, Halliburton and Eni reported integrated closed-loop drilling and managed-pressure drilling in deepwater Indonesia. SLB is simultaneously pushing AI toward the edge across wells, facilities and production systems, where latency, connectivity and operational continuity matter. These are not chat interfaces placed beside operations. They are examples of digital intelligence moving into the operating loop.
As AI moves from interpretation toward action, model quality remains necessary but becomes insufficient. Safe operating envelopes, state estimation, data quality, latency, deterministic controls, fallback modes, authority and management of change become part of the AI system.
One upstream stack, five different kinds of intelligence
Reservoir, wellbore, flow, pressure, equipment and process models constrain what is physically plausible.
ML detects patterns, estimates states, forecasts behavior and identifies anomalies from operational data.
Foundation models interpret documents, logs, procedures and mixed context; they synthesize and explain.
Agents retrieve context, call tools, coordinate tasks and prepare or execute bounded actions.
Deterministic and automated systems execute within defined operating constraints and interlocks.
The architecture becomes powerful when these layers complement rather than impersonate one another. A language model should not substitute for a hydraulics model. A predictive model should not silently become the source of record. An agent should not bypass a safety interlock because it has access to an execution tool.
The upstream decision chain
Each step has a different evidence burden. An anomaly detector can be useful with uncertainty. A system changing choke position, mud parameters or well trajectory needs a much stronger combination of validated state, operating envelope, execution authority and recoverability.
Where the architecture changes across upstream
| Domain | High-value intelligence | Execution boundary | Architecture that matters |
|---|---|---|---|
| Subsurface | Seismic interpretation, reservoir characterization, uncertainty, development scenarios | Technical interpretation and investment decisions remain accountable engineering decisions | HPC, domain models, provenance, uncertainty, versioned interpretations |
| Drilling & wells | Geosteering, dysfunction detection, hydraulics, parameter optimization, well planning | Closed-loop action is possible inside validated envelopes and interlocks | Real-time data, downhole models, edge compute, rig controls, fallback states |
| Production | Surveillance, allocation, optimization, lift/injection, flow assurance, anomaly detection | Recommendations can progress to bounded control where process safety permits | Historian/SCADA context, digital twins, edge AI, constraints, control integration |
| Integrity | Degradation detection, inspection prioritization, failure prediction, barrier surveillance | AI can prioritize evidence; acceptance, deferral and fitness decisions need defined engineering authority | Inspection history, sensor quality, degradation models, standards, consequence data |
| Operations | Alarm context, procedures, troubleshooting, work coordination, operating knowledge | Retrieval and recommendation are broad; process changes require explicit authorization | OT/IT separation, permissions, procedures, live state, audit, cyber controls |
Subsurface: AI does not remove uncertainty
Aramco’s 2026 upstream supercomputing investment is instructive. The planned system expands seismic and reservoir-modeling compute capacity sevenfold. The implication is not that physics and simulation are being replaced by general AI; it is that upstream intelligence continues to depend on very large domain-specific computational systems. AI can accelerate interpretation, pattern recognition and scenario generation, but reservoir decisions still depend on assumptions, uncertainty and model lineage.
A useful subsurface AI architecture therefore preserves the chain from raw and processed data through interpretation, earth model, reservoir model, scenario and decision. Generative interfaces can make this chain easier to interrogate, but they should make provenance more visible rather than flatten it into a persuasive answer.
Drilling: autonomy is already real
The strongest evidence of operational autonomy is currently in well construction. Halliburton’s LOGIX environment integrates steering, drilling dynamics and collision avoidance. Its 2026 Guyana deployment connected automated interpretation and geosteering with rig automation; the Indonesia deployment integrated drilling execution with managed-pressure drilling. SLB’s Neuro geosteering likewise uses real-time subsurface measurements and cloud/edge AI to guide trajectory.
This changes the governance question. “Should AI be allowed to drill?” is too broad. The engineering questions are: which variables may the system manipulate; inside what limits; from which validated state; with which interlocks; under what degraded-data conditions; with what human override; and what happens when models disagree?
Production: intelligence has to live close enough to the process
Production optimization often has the strongest economic case for continuous intelligence: high-frequency data, repeated decisions and measurable outcomes. It also exposes architectural constraints that enterprise AI discussions frequently miss. Remote assets may have constrained connectivity. OT networks are intentionally separated. Latency matters. Equipment state changes continuously. A cloud-only conversational layer is therefore an incomplete architecture.
SLB’s June 2026 collaboration with Qualcomm specifically targets low-power edge AI across wells, facilities and production systems. SLB’s Intelligent Production Studio goes further: a physical multi-well environment integrates flow measurement, injection systems, connected equipment, digital twins and autonomous workflows. The direction is toward distributed intelligence in which cloud, edge and control layers each do the work appropriate to them.
Fleet learning, heavy analytics, model management, cross-asset optimization, enterprise knowledge.
Low-latency inference, local context, resilience during connectivity loss, equipment-adjacent intelligence.
Deterministic execution, interlocks, process constraints, safe states and validated automation.
Integrity: prediction is not disposition
Integrity is a useful test of Human–AI work design. AI can combine sensor histories, inspection results and failure patterns to identify degradation or prioritize inspections. But a prediction does not itself establish fitness for service, acceptable risk or permission to defer intervention. Those are engineering and operating decisions made within standards, consequence models and organizational authority.
The design objective is therefore not maximum autonomy. It is a faster, better evidenced chain from signal to engineering disposition: detect → contextualize → estimate degradation → rank consequence → recommend inspection/action → engineer disposition → record outcome. The outcome should then become evidence for future model performance.
Authority should rise more slowly than capability
Read state
Interpret evidence
Estimate future state
Propose action
Stage bounded action
Act inside envelope
Capability can progress quickly across this sequence. Authority should progress only when evidence, operating constraints, assurance and recovery mechanisms justify it. High consequence does not prohibit automation; drilling automation already demonstrates that. It raises the standard of the surrounding system.
From AI pilot to operational system
- Choose one recurring operational decision, not an abstract AI use case.
- Map the evidence chain: sensors, logs, models, procedures, systems of record and human observations.
- Define the operating envelope: constraints, unavailable actions, degraded modes and stop conditions.
- Establish a baseline: cycle time, production/reliability outcome, NPT, false alarms, intervention quality or other relevant measure.
- Start below the authority ceiling: observe, explain or recommend before granting action where consequence warrants staged evidence.
- Run in parallel against existing decisions long enough to expose disagreement and edge cases.
- Integrate only what earned integration: data, models and workflow before execution rights.
- Record decisions and outcomes so operational evidence becomes the next evaluation dataset.
- Use management of change when the AI system changes operating practice, responsibilities or control logic.
- Scale by repeatable outcome, not by number of deployed models or agents.
The emerging upstream architecture
The next upstream AI system is unlikely to be one model or one agent. It is a governed assembly: domain data and models; live operational state; enterprise and field knowledge; predictive and generative intelligence; tools and agents; edge execution; deterministic control; and human operational authority.
The useful design question is therefore not “where can we add AI?” It is:
Which recurring upstream decisions can be made materially better, faster or more consistently—and what complete operating system is required for intelligence to influence them safely?
Sources
- Halliburton, 16 Mar 2026 — ExxonMobil/Halliburton fully closed-loop automated geological well placement, Guyana.
- Halliburton, 9 Jul 2026 — Eni/Halliburton closed-loop rig automation and MPD, Indonesia.
- SLB, 9 Jun 2026 — SLB and Qualcomm edge AI for wells, facilities and production systems.
- SLB, Jun 2026 — Intelligent Production Studio and autonomous multi-well production environment.
- SLB, 8 Oct 2026 — TotalEnergies global deployment of DrillPlan coherent well planning.
- SLB, 2024–2026 — Neuro autonomous geosteering and DrillOps automation.
- Aramco, 6 May 2026 — next-generation upstream supercomputer for seismic processing and reservoir modeling.
- Aramco — Building the AI future; 2025 AI deployment and realized-value summary.
Related research: AI in oil & gas: a value-chain map of credible use cases