Ask an energy executive where the company uses AI and you will usually hear about the visible projects: the chatbot pilot, the copilot licenses. Those are real, but they are rarely where AI is making decisions. The decisions happen further down, inside software that ranks, flags, forecasts and scores, where a model’s output quietly became the default somebody acts on.
Across upstream, midstream and downstream, the AI carrying the most operational weight is usually the AI nobody announced. You cannot govern it until you know where it sits. And one of the most consequential decision points sits outside the company entirely, in the AI tools your buyers use to build a shortlist.
What counts as an AI decision if a person still signs off?
Most operational AI in energy does not act on its own. It recommends, and a person approves. That arrangement looks safe on an org chart and often is not in practice. When a production engineer works through a long list of wells each morning and the model has already sorted them by expected decline, the sort order is the decision. The engineer starts at the top. Whatever sits at the bottom gets looked at later, or not at all.
NIST’s Generative AI Profile names this pattern. One of the twelve risks it lists is “human-AI configuration,” which covers automation bias and over-reliance on what a system produces. The profile targets generative AI, but the failure travels: a human in the loop who has stopped questioning the loop is a signature, not a control.
So the useful test is not whether a person approves the output. It is whether anyone would notice if the model’s recommendation were wrong for a month. Wherever the honest answer is no, AI is making the decision.
Where does AI make decisions upstream?
Upstream, the quiet decisions cluster around subsurface interpretation and production. Machine-learning tools that pick faults and horizons in seismic volumes shape which prospects look attractive long before anyone debates them in a capital meeting. Production forecasting models, including the ones that fit decline curves automatically, feed the numbers behind workover priorities and, eventually, reserves conversations. Failure prediction on artificial lift decides which pumps get attention this week.
None of these feel like decisions to the people using them. They feel like tools, which is how the model’s assumptions about the reservoir and the equipment end up embedded in choices the company will later have to explain.
Where does AI make decisions in midstream operations?
Midstream is where AI decisions are most likely to carry safety and regulatory consequences. Anomaly classification on in-line inspection data helps decide which features get excavated and which get monitored. Leak detection logic decides what counts as an alarm worth a controller’s attention, and the threshold setting is a decision about acceptable risk whether or not anyone frames it that way.
The thing to watch here is drift. A model tuned on last year’s operating conditions keeps producing confident output after the throughput, the product mix or the sensor fleet has changed. Nobody re-decides the threshold, because nobody remembers deciding it the first time.
Where does AI make decisions downstream?
Downstream, AI tends to sit closer to money. Process optimization layers with machine-learning components adjust refinery operating targets. Demand forecasts shape run plans and inventory positions. In fuels and lubricants marketing, pricing tools can move posted prices with little human review, and trading desks use models that rank opportunities before a trader ever sees them.
Because that output touches revenue and customers directly, this is where “the system recommended it” is the weakest answer available.
Where should an operator look first in each segment?
| Segment | Where the model sits | What it is quietly deciding | The question to ask its owner |
|---|---|---|---|
| Upstream | Seismic interpretation assist | Which features and prospects get a closer look | What did the model miss on the last prospect we drilled? |
| Upstream | Production forecasting | Workover priority and the forecast behind it | When was the forecast last checked against actuals, and by whom? |
| Upstream | Artificial-lift failure prediction | Which equipment gets attention first | What happens to the wells the model ranks lowest? |
| Midstream | Inspection anomaly classification | Excavate or monitor | Who can overrule the classification, and is the override recorded? |
| Midstream | Leak detection alarm logic | What counts as an alarm | Who set the threshold, and when was it last reviewed? |
| Downstream | Process optimization | Operating targets | Which constraints can the model move without asking? |
| Downstream | Pricing and trading models | Price and position | Could we reconstruct why it recommended what it did on a given day? |
| Outside the company | Your buyers’ AI assistants | Whether you make the shortlist | What do these tools say about us today? |
Which AI decision happens outside your company entirely?
The last row of that table is the one most energy companies have never examined. Engineers, procurement teams and plant managers are starting to ask AI assistants which service companies, suppliers and contractors to consider. The answer they get is a decision about your company, made by a system you do not control, using whatever public information it can find and trust.
A 2023 research paper on generative engine optimization showed that the way a source presents its content changes how often generative engines surface it; in the authors’ tests, some methods raised a source’s visibility in AI answers by up to 40%. If your capabilities, certifications and project history are not described clearly in places these systems read, you are not being considered, and nobody will tell you. Fixing that is what EWR Digital does as an oil and gas marketing agency, and it is the same problem as the rest of this map, facing outward: a decision about you is being made on evidence you have not reviewed.
Who should own each decision point once it is mapped?
The NIST AI Risk Management Framework puts this under its Govern function, where one subcategory calls for mechanisms to inventory AI systems, resourced according to the organization’s risk priorities. The framework is voluntary, but its logic is hard to argue with. You cannot assign accountability for a decision nobody has written down.
The harder part is that each entry needs a named owner who is neither the vendor nor the data science team, but the business leader whose results depend on the output. The asset manager owns the forecast. The control room manager owns the alarm threshold. The commercial lead owns the pricing model. That assignment is what AI decision governance, the work ModalPoint does, is built around: moving accountability from “the system” to a person with a title.
What changes once the decision points have names?
Something small, usually. A threshold gets a review date. A forecast gets compared with actuals on a schedule instead of after a surprise. A pricing model gets a log that would let someone reconstruct last Tuesday. None of it slows the operation down, and most of it uses tools already in place.
What it changes is the answer to a question every energy company will eventually hear from a regulator, an auditor or a partner: who decided this? For each row in the table above, someone at your company should be able to answer with a name. If a row has no name yet, that is the one to start with.
Matthew Bertram is Chief Marketing Officer of the Oil & Gas Global Network, President of ModalPoint, an AI decision-governance advisory, and CEO of EWR Digital, a Houston digital marketing agency in business for more than 25 years. He hosts The Best SEO Podcast, created the LLM Visibility™ methodology for getting brands cited in AI search, and is a contributor to NIST’s Cyber-AI Profile.