Gains:
- Ability to distinguish where artificial intelligence saves real time and cost in the oil and gas workflow and which decisions should remain the responsibility of the authorized engineer
- Ability to implement a discipline that cross-validates each AI output with order of magnitude, physical plausibility, and field/well evidence
- Recognizing the concepts of uncertainty, hallucination, commercial confidentiality and security-critical risk and acquiring the habit of setting up secure prompts by anonymizing the context.
Petroleum and natural gas engineering is the art of bringing fluid from a reservoir (a porous rock mass containing oil or gas) many kilometers underground to the surface in a safe, efficient and environmentally friendly manner. This business is woven with data from start to finish: seismic sections, well logs, core measurements, production flows, pressure logs, equipment sensors, pipeline SCADA flows. Artificial intelligence comes into play precisely in this abundance of data; It speeds it up, makes patterns visible, and reduces hours of work to minutes. But let's reiterate the first and most important sentence of this module: artificial intelligence is an accelerator in this profession, not a signatory. The cost of a well blowout, a pipeline explosion or a misdeclared reserve is life, the environment and big capital; The authorized engineer bears this cost, not the AI.
In this unit, we map out how to embed AI into the oil and gas workflow: where there is real gain, where there is danger, and with what validation discipline we will test each output.
Where Artificial Intelligence Brings Real Profit
The areas where AI is strongest in the industry are where extracting patterns from large numbers of samples is valuable and errors can be caught quickly:
- Data structuring and cleaning: Making dispersed well logs, production tables in different formats, free text drilling reports queryable.
- Pattern recognition: Fault traces in seismic section, facies (rock type specific to the depositional environment) boundaries in log curves, fault signatures in vibration spectrum.
- Prediction and prediction: Porosity from logs, degradation curve from production history, remaining useful life from sensor.
- Draft production: Report skeleton, literature summary, code snippet, procedure outline.
- Anomaly detection: Pressure drop in the pipeline, unusual vibration in equipment, unexpected deviation in production.
What these fields have in common: AI produces a first draft or noteworthy, the engineer verifies it and takes responsibility.
Where Artificial Intelligence Cannot Take Over Responsibility
Some decisions should inherently remain with the engineer because in case of error the damage is irreversible:
- Well control and pressure management (blowout prevention) decisions.
- Safety-critical alarm thresholds (gas leak, H2S, high pressure).
- Reserve declaration and official figures presented to the investor.
- Pipeline and pressure equipment integrity/decommissioning decisions.
- Environmental discharge and site safety approvals.
Caution: "AI suggested" is not a defence. The question to ask in an engineering error investigation is: “With what independent evidence did the authorized engineer verify this output?” If you don't have an answer, you shouldn't have used the output at all.
Validation Discipline: Three-Layer Filter
Pass each AI output through three filters based on importance.
1. Order of magnitude control. Is the result roughly within the expected power of 10 range? The permeability of a sand reservoir is typically in the millidarcy–darcy range; If the model says "80,000 darcy" there is a unit or editing error without going into detail. Porosity is between 0%–40%; 65% is physically impossible.
2. Physical reasonableness check. Does the output obey known laws of physics? Water saturation should be between 0–1; production flow rate cannot be negative; material balance (produced + remaining = initial) should be closed; The drawdown curve cannot show increasing production over time (unless there are new wells/interventions).
3. Independent evidence check. Compare the output with a source that does not produce it: core measurement, well testing (pressure–flow data), adjacent well, field inspection, a second method. In the hierarchy of evidence, direct measurement (core, well test) is always superior to model prediction.
Tip: Scale verification intensity with the risk of the output. For a literature summary, a quick read is enough; Full independent calculation and second engineer control is essential for a well pressure limit.
Secure Prompt: Privacy and Context
Oil-gas data is extremely sensitive from a commercial perspective. Well coordinates, production flows and reserve figures often belong to the license holder and the partnership; Engaging in an uncontrolled external AI service creates loss of competition, breach of contract and legal risk.
- Use pseudonyms ("Well-A", "Field-X") instead of actual well names/coordinates.
- Replace actual production and reserve figures with representative (scaled) values; The ratio and pattern are preserved, the secret number is not leaked.
- Only use tools approved by the institution with a data processing agreement.
Three Mini Cases: By the Numbers
Case 1 — Time savings, correct use. LAS log files of 240 wells in one field were kept with different naming and units. AI-powered standardization has reduced unit conversion and curve name matching from two days to three hours. Random 10% wells were manually checked; Mismatching of the neutron log in two files was detected and corrected. The gain was real because the verification layer was in place.
Case 2 — Contrived reserve risk. An engineer fed production data into the model to speed up drawdown curve analysis; The model took the b prime of 1.4 in the hyperbolic drawdown and extrapolated it to 30 years, making the EUR (final predicted production) twice as high as expected. The material balance check showed that the figure contradicted the reservoir volume; The b base was moved to the physical limit and the result was corrected. If blind were considered, the reserve would be exaggerated.
Case 3 — Safety-critical limit. One team asked AI for suggestions to “optimize” the gas detection alarm threshold; The model suggested increasing the threshold to reduce false alarms. The HSE engineer denied this: a false negative in gas detection (missing the actual leak) is fatal and the threshold cannot be set below the legal/standard limit. AI recommendation alone cannot be a layer of security.
Weak Prompt / Strong Prompt
Weak prompt:
Look at the production data of this well and tell us its reserve. [data]
Powerful prompt:
DRAFT drawdown curve analysis from the following (representative/anonymised) production data. Follow these rules:- Apply the Arps pattern; Keep the b exponent in the physical range 0-1, write your justification. - Give the EUR estimate as a low/medium/high scenario, not a SINGLE number. - List each assumption separately; indicate additional data required for cross-checking with material balance. - Do not present the result as "exact"; mark steps that need to be verified.Data: [anonymised production table]
A strong prompt imposes physical boundaries, scenario thinking, and the need for verification.
Four Copiable Templates
1) Output risk classification:
Determine the risk class before using the following AI output: (a) what will be damaged/what will be damaged, (b) is the damage reversible, (c) what independent evidence should be verified. Output: [summary]
2) Order of magnitude filter:
Check the order of magnitude of the following result: what is the typical expected range, is the result in this range, if not, where is the possible unit/construction error? Result and its unit: [value]
3) Anonymization assistant:
Identify commercially sensitive information (well name, coordinates, actual flow rate, reserve, partnership) from the text below and give pseudonym/representative value suggestion.Text: [text]
4) Verification plan:
Write a verification plan for the following engineering output: which hand calculation, which physical boundary, which independent data (core/well test/adjacent well), and which second-party confirmation is required? Output: [summary]
The Role of AI: At a Glance
Quest
The role of AI
Engineer's role
verification anchor
Log standardization
Configuration, transformation
Sample control
Raw file, core
Porosity estimation
account draft
Approval
Core, Archi equation
Deduction/EUR
Model fitting, scenario
decision
Material balance
Leak/alarm
anomaly sign
decision, intervention
independent detector
reserve declaration
draft, account
Liability, signature
PRMS/SEC, evaluator
Common mistakes
- Mistaking fluency for accuracy. The confident language of the model does not mean that the number is correct.
- Not scaling verification according to risk. Passing a security-critical output as quickly as a literature summary.
- Entering confidential data into an uncontrolled vehicle. Sharing real well and production data without anonymization.
- Avoiding responsibility with "AI suggested". Moving the printout to signature without independent evidence.
- Loading the security function into a single model. Replacing independent, certified layers with AI.
In summary
- AI is a powerful accelerator in the oil-gas workflow; But the decision and signature responsibility remains with the authorized engineer.
- Filter each output through three layers: order of magnitude, physical plausibility, independent evidence.
- Scale verification intensity with the risk (loss on error) of the output.
- Anonymize business sensitive data and use only approved tools.
- Safety-critical functions should not be based on a single AI model, but on multi-layered independent systems.
Application task
Choose an AI use case from your own business (or a representative one): porosity estimation from logs or production drawdown, for example. Write a three-layer verification plan for this output: (1) what the expected order of magnitude is, (2) what physical limits apply, (3) what independent evidence you will confirm, and who will do the second check. Then anonymize your prompt so that it does not contain the actual well name/coordinate/number.
checklist
- [ ] I can distinguish between areas where AI profits and decisions where the responsibility should remain with the engineer.
- [ ] I filter each output through orders of magnitude, physical plausibility, and independent evidence.
- [ ] I scale verification intensity with the risk of the output.
- [ ] I anonymize commercially sensitive data and use only approved tools.
- [ ] I do not rely on a single model for security-critical functions, but on multi-layered independent security systems.