How can AI help oil and gas? Where it works today
AI for oil and gas
Your platform already produces more data than anyone reads: the historian, the maintenance system, the daily reports, the P&IDs, the HSE forms. AI does not need more of it. It needs to know what it means, which pump P-204 is, what feeds it, and who was on shift. AI needs context, and in oil and gas the context is the asset model you already half have. This page says where that pays off today, and where it does not yet, on the terms of the AI and data platform.
Three places it already pays, and none of them is a chatbot
Reading the shift, not the dashboard
A handover today is a person remembering twelve hours. An agent reading the same historian tags, alarms and work orders drafts the summary from what actually happened, with every line traceable to a trend or an event. The outgoing lead corrects it in two minutes instead of writing it in twenty.
From a trend to a work order
A bearing temperature creeping up for three weeks is visible to anyone who looks. Nobody looks at every tag. An agent that knows which tags belong to which pump, and what the spare situation is, drafts the notification with the trend attached. A planner approves it, or does not.
Figures that can show their inputs
Flare volumes, fuel gas, produced water: every reported number is a chain of tags, meters and corrections. When it is questioned two years later, the answer should be a walk back through that chain, not a search for whoever built the spreadsheet.
The data you already have
One model of the platform, fed by the systems you already run
Nothing here is new instrumentation. The historian keeps decades of tags, the ERP holds every notification and work order, the document system has the P&IDs and the inspection reports, and the daily reports say what people saw. What is missing is the layer that ties them to the same equipment. The operational data platform copies them out over standard protocols, without touching the control systems, and lands them against one model: which asset, which system, which train. Getting the data in is the small part of the job, and the technical picture says what the platform is made of.
The historian is not the problem. It is the archive. What is missing sits above it.
First tasks
Work an agent can take this year, with a person signing
The right first task is boring, repeated and checkable: someone does it every shift or every month, the data it needs already exists, and a person can read the draft in minutes. Product Discovery is two weeks spent finding that task on your platform. These three are where it usually lands.
The handover summary
Drafted from the tags, alarms and work orders of the shift, each line linked to the trend or event it came from. The outgoing lead edits and signs. Nothing is sent that a person did not read.
The notification from a trend
Vibration or temperature drifting on a pump the model knows, with the spare and the last overhaul looked up. The agent drafts the notification; the planner decides whether it becomes a work order.
The reconciled tag list
Every platform has a tag list, a maintenance hierarchy and a set of P&IDs that disagree at the edges. The agent proposes the matches, with the evidence, and an engineer confirms them. That reconciliation is the beginning of the graph.
Context, on the platform
A question walks the plant, not the tag list
Ask why the first-stage separator pressure climbed at 03:10 and a person starts from what they know: the separator, the outlet valve, the pump on the oil leg, the level controller. That knowledge is what an agent lacks unless it is written down. AI needs context, and on a platform the context is the graph of what feeds what. With it, the question follows real relationships from the symptom to the cause, and the answer comes with its path. Without it, the agent has forty thousand tags and a hunch.
The path is the reasoning. Every hop is a relationship somebody modelled once.
Deliberately fiction
One postcard from a platform that does not exist
This has not happened. It is written from the pieces above, and it is here because it sounds like a shift report, which is the point. If it bothers you, the useful question is which piece it would fail without.
The night the separator did not trip
02:40. An agent watching the first-stage separator sees the level controller working harder than the flow explains. It walks the graph: the oil-leg outlet valve XV-118 has been commanded open for six hours but the downstream pressure says otherwise; P-204 is drawing less than its curve; the valve's last overhaul was fourteen months ago. It drafts a note for the control room with the three trends side by side, and a notification for the day shift. The operator reads it at 02:46, strokes the valve by hand, watches the pressure fall, and closes the note with one line. Nothing tripped. The notification is approved at 07:15 by someone who slept.
Where AI will not help you yet
It will not run the process; the control system does that, and should keep doing it untouched. It will not replace the historian, which is a good archive and can stay one. It will not know your platform on day one: where the tag list, the maintenance hierarchy and the P&IDs disagree, the agent finds the gaps before it finds anything else, and closing them is engineering work, weeks of it, not a download. And it should not commit anything on its own. On a platform, an agent drafts and a person signs, which is exactly how the paperwork works today. It runs where the data is: on your servers, in a closed network, or on hardware we own in Stavanger, under Norwegian law, and the code is open under AGPL-3.0. The rest of the argument is on the AI and data platform page.
Fair questions
What operations teams ask us
How can AI help oil and gas operations today?
In the repeated, checkable work: shift handover summaries drafted from the historian, alarms and work orders; maintenance notifications drafted from a trend on equipment the model knows; reported figures such as flare and fuel gas that can show every tag and correction behind them; and reconciling the tag list with the maintenance hierarchy and the P&IDs. In each case the agent drafts and a person commits.
Does this require replacing the historian or the control system?
No. The control system is not touched, and the historian stays as the archive it is. The platform copies data out over standard protocols such as OPC-UA into a model of the operation that sits above them, and that layer is additive and reversible.
What data does an AI agent need on a platform?
Less than you think, connected better than it is. Historian tags, the maintenance system, the daily reports and the documents you already have, attached to the same equipment in one model, so the agent knows which tags belong to which pump, what feeds what, and what happened to it last. That context is the knowledge graph, and it is what turns a confident guess into an answer with a path.
Can operational data stay inside our environment?
Yes. The platform is open source under AGPL-3.0 and runs on your own servers, in your cloud or in a closed network. If you would rather not run it, IntelliStream hosts it on hardware it owns in Stavanger, under Norwegian law, and agents reach it through MCP under the permissions of whoever is asking.
Where should an operator start?
With one task that repeats every shift or every month and that a person can check in minutes: the handover summary is the usual first. Confirm the data it needs can be copied out of where it sits, model only the equipment that task touches, and run it with a person signing every draft. Product Discovery is a two-week engagement built to find that first task.
Start with one shift, not the whole platform
Tell us which report your people write from memory today and what systems it should have come from. We will say honestly whether an agent can draft it now, and what the model needs first.