Offshore operations are moving from AI that analyzes information toward systems that can take a sequence of actions with limited supervision. Petronas is adding agentic AI capabilities to myPROdata, its web-based portal for Malaysian exploration and production data. Agentic AI means software that can pursue a goal through multiple steps rather than simply return one answer.

In August, ADNOC and SLB deployed an AI-enabled real-time operations platform across more than 120 rigs. The shift promises substantial operational gains. It also raises a basic engineering question: how much authority should an AI system receive before a human must approve what happens next?

The benefits deserve serious weight. ADNOC says its platform reduces engineering effort by 30 to 40%, allows engineers to support two to three times more rigs, and can help avoid one to two days of downtime. Woodside’s remote subsea inspection program at Shenzi sends live inspection data to onshore engineers instead of requiring a full team of inspectors, engineers, and support personnel aboard the inspection vessel. That reduces personnel exposure offshore and speeds response to anomalies.

ExxonMobil has also deployed closed-loop autonomous drilling, where software can adjust drilling parameters without continuous human intervention. Used well, these systems can reduce repetitive work, non-productive time, and personnel on board while improving consistency and situational awareness.

Those advantages make the control question more important, because the consequences change when AI moves from recommendation to execution. A poor recommendation can be reviewed and rejected. A system with credentials, access to operational technology, or authority to trigger a workflow can turn an error into an action before anyone intervenes.

Offshore operators should therefore give each AI system an authority budget: the maximum delegated power it receives before human approval is required. The budget should be defined before deployment, just as operators define equipment limits, alarm thresholds, and operating envelopes.

A practical way to set that budget is by consequence. Low-risk systems can receive broad read-only access and freedom to analyze. Medium-risk systems can prepare work orders, maintenance plans, or operational recommendations but hold them for approval. High-risk actions, such as changing setpoints, modifying control logic, initiating shutdowns, deploying software, spending material amounts of money, or sending external instructions, should cross a clear human-approval gate. Operators can widen those limits later as evidence of reliability accumulates. That risk-based approach also fits the broader NIST AI Risk Management Framework, which emphasizes managing AI risks across deployment and use.

For a reservoir or drilling-data system, the authority budget might allow analysis and recommendations but prohibit changes to drilling parameters. For an inspection robot, it might permit autonomous routing and data collection while requiring approval before it interacts with equipment or initiates maintenance. For production systems, operators can separate permission to detect an abnormal condition from permission to change a setpoint, silence an alarm, modify control logic, or issue a work order. SLB has already demonstrated offshore drilling sections controlled almost entirely autonomously, which shows why the boundary between observation, recommendation, and action needs to be explicit.

Access should follow the long-established cybersecurity principle of least privilege: give a user or process only the resources and authorizations needed for its assigned task. An AI agent analyzing maintenance records does not automatically need permission to alter them. A system helping plan a shutdown does not automatically need access to execute one. Credentials can be task-specific and time-limited, with stronger approval requirements as consequences increase.

Operators also need an audit trail. Material AI decisions and tool use should be logged so teams can reconstruct what happened. Unusual behavior should trigger review. Safety-critical functions should have reliable pause, fallback, and shutdown mechanisms that do not depend on the AI system cooperating. NIST is now developing a trustworthy AI profile for critical infrastructure that explicitly addresses AI in operational technology, industrial control systems, autonomous robots, fail-safe operation, and human oversight. Offshore operators do not need to wait for that work to finish before applying the same logic.

Clear limits can also improve responsible AI adoption. Workers are more likely to use powerful systems confidently when they know which decisions remain theirs and where the technology must stop. The goal should be to capture AI’s efficiency and safety benefits while keeping delegated authority proportional to demonstrated reliability and operational consequence.

As offshore AI moves closer to real-time action, authority can no longer remain an informal assumption. Before an AI system receives credentials to act, operators should decide exactly what it may do on its own, what requires human approval, and how people can stop it when conditions move outside the expected envelope.