All industry field guides

Applied research 01 · Oil & gas

Make every operating decision carry its evidence.

The field already produces data. The harder problem is connecting live conditions to engineering knowledge, operating envelopes, and the accountable action—without asking a language model to invent what the plant must know.

Decision domainProduction, integrity, maintenance, and process safety
Evidence surfaceHistorian, SCADA, CMMS, inspections, work orders, procedures
Control boundaryRecommend and prepare; authorized operators approve and execute

Our point of view

Oil & gas does not need another dashboard.

It needs a decision layer that can explain why a compressor anomaly matters now, which barrier or production commitment it touches, and what evidence supports the next safe action. That layer must understand equipment hierarchy, failure modes, operating context, permit state, and the difference between correlation and an approved engineering rule.

We build the domain model before the agent. Time-series models detect change; an asset knowledge graph gives the change meaning; symbolic controls enforce operating rules; a grounded agent assembles evidence and coordinates the next step.

Design principlePrediction becomes operational only when it is attached to an asset, a consequence, an approved procedure, and a named decision owner.

These are applied research patterns, not descriptions of completed customer engagements. Public operator programmes in the sources establish sector context; their reported outcomes are not Zustis outcomes.

Applied research outputs

Three decisions worth redesigning.

Each pattern starts with a bounded operational decision and ends with an evidence trail—not an open-ended chatbot.

01

Production intelligence

From scattered shift evidence to one operational narrative.

This research pattern structures the handover around deviations, actions, constraints, and unresolved risk—so the next shift could receive a live account of what changed, not another document to search.

The decision

Which production deviation needs intervention, continued observation, or escalation before the next operating window?

The evidence

Historian tags, alarms, operator logs, lab results, deferment codes, equipment state, maintenance history, and approved operating procedures.

System behavior

Builds a time-bounded event graph, groups causal signals, retrieves governing procedures, flags missing evidence, and drafts a cited handover with ranked actions.

Human boundary

The shift supervisor validates causality and owns the operating instruction. The system cannot change a set point or suppress an alarm.

02

Process safety reasoning

Check the job against the plant, not only the form.

The reference design represents permits, isolations, SIMOPS, equipment status, competencies, and site rules as a machine-checkable safety context before a work pack reaches authorization.

The decision

Is this job ready to enter the permit workflow, and what conflict or missing control must be resolved first?

The evidence

Permit-to-work fields, P&IDs, isolation certificates, gas tests, JSA, SOPs, competency records, nearby work, and live equipment state.

System behavior

Extracts entities and hazards, resolves them against the plant ontology, executes deterministic conflict rules, and returns exceptions with source clauses.

Human boundary

It never issues a permit, declares an isolation safe, or overrides the area authority. Ambiguity blocks progression rather than being guessed away.

03

Asset integrity

Turn an anomaly into a maintenance decision—not just a score.

This research pattern joins condition signals to failure knowledge, inspection evidence, spares, work history, and production consequence to prepare an action an engineer could accept or reject.

The decision

Continue, inspect, derate, or plan intervention—and how urgently?

The evidence

Vibration and process trends, inspection/NDT findings, failure modes, bad-actor history, criticality, spares, maintenance windows, and OEM guidance.

System behavior

Detects drift, retrieves comparable events, estimates evidence quality, evaluates rule constraints, and drafts a CMMS-ready recommendation with alternatives.

Human boundary

Reliability and operations engineers approve the diagnosis, risk treatment, and work order. Low-confidence evidence routes to inspection.

Reference architecture

A compound system, not one model.

Every layer has a different job. Keeping them separate makes the result testable, governable, and replaceable.

01 · Sense

Operational event layer

Streams, documents, inspections, and human observations are time-aligned with provenance and data-quality state.

02 · Understand

Asset knowledge graph

Equipment, tags, locations, failure modes, barriers, work, and procedures become a shared domain model.

03 · Reason

Hybrid decision engine

Statistical detection proposes; retrieval supplies evidence; deterministic rules enforce the operating envelope.

04 · Act

Controlled workflow agent

The agent drafts, routes, and records actions through approved systems with role-based permissions and checkpoints.

The line we do not cross.

  • No silent actionNo safety-critical control change, alarm suppression, permit issue, or equipment release without authorized human action.
  • Source or abstainEvery material recommendation links to live evidence and governing procedure; missing or conflicting evidence is explicit.
  • Context expiryA recommendation expires when plant state, permit state, or evidence freshness moves beyond its validated window.
  • Full replayInputs, retrieved passages, rules fired, model versions, edits, approvals, and downstream actions are auditable.

Validation before autonomy

Prove the decision loop in shadow mode.

We measure against engineering review and real operating history before asking a team to rely on the recommendation.

  1. 01

    Reconstruct

    Replay known events and verify whether the system assembles the correct assets, evidence, and procedure context.

  2. 02

    Challenge

    Test stale tags, conflicting documents, missing isolations, unusual modes, multilingual notes, and adversarial instructions.

  3. 03

    Shadow

    Run alongside the existing process. Compare recommendations, abstentions, and preparation time without operational authority.

  4. 04

    Bound

    Release only approved workflow steps, with thresholds, expiry, escalation, rollback, and continuous drift monitoring.

What earns the right to scale.

Evidence precision
Material claims resolve to the correct tag, document revision, clause, asset, and time window.
Unsafe miss rate
Safety conflicts and mandatory controls are caught; critical false negatives are treated as release blockers.
Abstention quality
The system declines when context is stale, ambiguous, out of domain, or below the agreed confidence floor.
Decision utility
Engineers judge whether the pack reduces search and coordination while preserving or improving decision quality.

Public sector signals

Why this matters now.

  1. ADNOC and SLB launch AiPSO for upstream operations

    Public evidence of AI joining real-time data and field workflows at multi-field scale.

  2. ADNOC and AIQ complete ENERGYai trial phase

    Sector signal for agentic reasoning over proprietary data and decades of operating knowledge.

  3. ADNOC deploys Neuron 5 process optimisation technology

    Public pilot context for predictive maintenance and reduced unplanned interruption.

Continue exploring

Related industry research.

NextEngineering & field servicesExploreShipping & logisticsExploreConstruction & infrastructureResearchOur methods

Choose one operating decision. We will make its evidence and control boundary visible.

Discuss an oil & gas problem All industry field guides