Institutional Knowledge AI Risk: Governing Energy Decisions
When an energy company captures a retiring expert’s knowledge into an AI system, it does not eliminate institutional knowledge risk. It relocates it. The decisions that expert used to own are now made or shaped by a model that is embedded, opaque, and largely unaudited. Capturing knowledge is not the same as governing the decisions that knowledge now drives. That gap is the new institutional knowledge AI risk, and it is most acute in oil and gas, where the “great crew change” is handing decades of judgment to software faster than most operators can govern it.
What is institutional knowledge AI risk?
Institutional knowledge AI risk is the exposure a company takes on when expert judgment that once lived in a person’s head becomes embedded in an AI system that now makes or shapes decisions. The knowledge is retained. The accountability for the decisions it drives is not, unless someone governs it.
This is a different problem than knowledge loss. Knowledge loss is when the expert walks out the door and the judgment leaves with them. Institutional knowledge AI risk is what you get when you successfully prevent that loss by encoding the expert into a model, and then discover the model is making calls no one reviews, on logic no one can see, against standards no one set.
The great crew change meets generative AI
The energy industry has been bracing for a demographic cliff for years. According to the Oil & Gas Journal, more than 50% of the U.S. oil and gas workforce is eligible to retire within the next decade, and a large share of the energy workforce is already 50 or older. The industry has a name for it: the “great crew change.” The stakes are not subtle. As Hart Energy has documented, operators face what amounts to a sudden and irrecoverable loss of institutional knowledge, and the depth of that knowledge is real. ExxonMobil’s average engineer tenure exceeds 30 years.
For most of the last decade, the answer to the great crew change was capture: knowledge management systems, mentorship programs, documentation. Now there is a faster answer. Generative AI can ingest an expert’s decision history, field notes, well logs, and maintenance records, and then produce recommendations that sound like the expert. So companies are doing exactly that.
This is where the timing matters. Recent industry surveys (2025) found that roughly 88% of organizations now use AI in at least one business function, yet only a small minority have a comprehensive AI governance framework in place. The capability arrived first. The governance did not. The great crew change is now being absorbed by AI systems that, in most companies, no one is governing.
Capturing knowledge is not the same as governing the decisions it now drives
Here is the distinction operators keep missing. Capturing knowledge answers the question, “Did we save what the expert knew?” Governing the decision answers a harder set of questions: Is the AI interpreting that knowledge correctly? Who is accountable when it is wrong? Can we see why it recommended what it recommended? Does the recommendation still hold as conditions change?
A captured knowledge base is a library. A governed decision is a controlled process. The retiring reservoir engineer did not just hold facts. They held judgment about when the facts applied, what the exceptions were, and when to override the obvious answer. When you encode that into a model, the facts transfer cleanly. The judgment about when to distrust the facts is exactly what gets lost, flattened, or silently approximated.
| Question | Capturing knowledge | Governing the decision |
|---|---|---|
| What does it preserve? | What the expert knew | How the decision gets made and checked |
| Primary artifact | Documentation, model, knowledge base | Accountable process with review and audit |
| Key failure mode | Knowledge is lost or never recorded | Knowledge is applied wrongly and no one notices |
| Who owns the outcome? | Often unassigned after the expert leaves | A named human, with the AI as input |
| Visible reasoning? | Not required | Required, provenance and rationale on record |
The trap is that capture feels like completion. The project ships, the expert retires, the dashboard goes green. The risk only surfaces later, when a recommendation that no one reviewed turns out to be wrong, and the company realizes the decision had quietly moved from a person to a system without anyone deciding that it should. This is the core of AI decision governance: treating the decision, not the document, as the thing you have to control.
Where embedded expert judgment quietly becomes ungoverned AI judgment
The shift from expert judgment to AI judgment rarely arrives as a formal handoff. It happens recommendation by recommendation, until the model is effectively making the call and the human is rubber-stamping it. In energy operations, four areas are especially exposed.
Reservoir and subsurface calls
A senior reservoir engineer’s intuition about a formation, when to trust the seismic, when the model is lying, which analog wells actually apply, is exactly the kind of judgment companies are now training models to replicate. When a model recommends a completion design or a recovery estimate, the underlying assumptions are buried in weights and training data. If the interpretation is subtly wrong, the error compounds across every well that follows the recommendation.
Maintenance triage and integrity decisions
Predictive maintenance models now decide what gets inspected, what gets deferred, and what counts as an acceptable risk. Those thresholds used to live with a reliability engineer who knew which equipment lied about its own condition. Encode that into a triage model and the deferral decisions become automatic, including the ones that, in a high-consequence facility, should never be automatic.
Drilling parameters and real-time optimization
Automated drilling advisory systems set weight on bit, RPM, and mud parameters in real time, drawing on historical performance that often encodes a specific driller’s hard-won preferences. The model optimizes for what it was rewarded on. If that reward did not include the rare, expensive failure mode the veteran driller was actually guarding against, the system will confidently steer toward it.
Procurement and vendor decisions
AI is increasingly used to score vendors, flag contract terms, and recommend awards. The institutional knowledge here, which suppliers cut corners, which terms have burned the company before, is precisely the tacit judgment that does not survive encoding well. A procurement model that looks objective can quietly bake in the wrong lessons. How that judgment gets exercised is the same problem we examine in how oil and gas companies make buying decisions: the decision is only as good as the governance around the inputs.
A governance model for AI-mediated institutional knowledge
Governing AI-mediated knowledge does not require slowing down the captured-knowledge program. It requires wrapping the decisions that program now drives in a control structure. ModalPoint frames this through DIG, Digital Information Governance (USPTO Reg. No. 8147558), built on four pillars: interpretation accuracy, exposure control, compliance, and signal lifecycle. Two of those pillars carry most of the load for institutional knowledge risk.
Interpretation accuracy asks whether the AI is reading the captured knowledge the way the expert meant it. A model can reproduce an expert’s conclusions while badly misreading the conditions under which those conclusions held. Governing interpretation accuracy means testing the model against cases where the right answer is “it depends,” not just the cases with clean answers, and confirming the system flags ambiguity instead of papering over it.
Signal lifecycle recognizes that a retired expert’s knowledge has a shelf life. Reservoirs deplete, equipment ages, regulations change, the field is not the field it was when the training data was generated. Governing the signal lifecycle means tracking when the knowledge behind a recommendation has gone stale, and forcing a review before a decade-old judgment quietly steers a present-day decision.
Two practices operationalize all four pillars:
- Human-in-the-loop on consequential calls. Not every decision needs a human, but every high-consequence decision needs a named, accountable one. The AI is an input to that person’s judgment, not a replacement for it. The point is to keep accountability attached to a human even after the original expert is gone.
- Provenance on every recommendation. A governed AI decision can answer “why did you recommend this, on what knowledge, from when?” Provenance turns an opaque output into something a reviewer can interrogate, an auditor can trace, and a regulator can accept.
This maps cleanly onto established frameworks. The NIST AI Risk Management Framework organizes AI risk into four functions, Govern, Map, Measure, and Manage, and its Govern function is precisely about assigning accountability for decisions a model influences. ISO/IEC 42001:2023, the first management-system standard for AI, provides the auditable structure to sustain that governance over time. DIG is how we apply both to the specific problem of energy institutional knowledge.
None of this is vendor-specific. ModalPoint is model-agnostic by design, the governance attaches to the decision, not to whoever’s model happens to be running underneath it. Whether your captured-knowledge system runs on a commercial foundation model, an internal build, or a vendor platform, the governance burden is the same, and so is the answer.
Frequently asked questions
What is institutional knowledge AI risk?
It is the exposure created when expert judgment that once lived with a person becomes embedded in an AI system that now makes or shapes decisions, without governance over those decisions. The knowledge is retained, but accountability, interpretation accuracy, and auditability are not, unless they are deliberately built in.
Doesn’t capturing a retiring expert’s knowledge in AI solve the great crew change problem?
It solves the capture problem, not the governance problem. According to the Oil & Gas Journal, more than 50% of the U.S. oil and gas workforce is eligible to retire within the next decade. Encoding their knowledge into AI prevents the knowledge from disappearing, but it relocates the decisions they owned into a system that is opaque and, in most companies, unaudited. That is a new risk, not the absence of one.
How is governing an AI decision different from documenting expert knowledge?
Documentation preserves what the expert knew. Governing the decision controls how that knowledge gets applied, whether the AI interprets it correctly, who is accountable for the outcome, whether the reasoning is visible, and whether the underlying knowledge is still current. Documentation is a library; a governed decision is a controlled, auditable process.
Which energy decisions are most exposed to ungoverned AI judgment?
Reservoir and subsurface interpretation, maintenance triage and integrity decisions, real-time drilling parameter optimization, and procurement and vendor scoring. In each, tacit expert judgment, the knowledge of when to distrust the obvious answer, is exactly what survives encoding least well.
Do we need a new framework, or can we use NIST and ISO?
Use both. The NIST AI Risk Management Framework (Govern, Map, Measure, Manage) and ISO/IEC 42001:2023 give you the structure and the auditability. ModalPoint’s DIG applies them to the specific problem of AI-mediated institutional knowledge in energy, with emphasis on interpretation accuracy and signal lifecycle.
Does this lock us into a specific AI vendor or model?
No. The governance attaches to the decision, not the model. ModalPoint is vendor-agnostic: the same governance applies whether the captured-knowledge system runs on a commercial model, an internal build, or a third-party platform.
About the author
Matthew Bertram is CEO of ModalPoint and EWR Digital, where he leads Decision Intelligence for Energy, governing the decisions AI now influences across oil and gas operations. He works with energy decision-makers to put structure around AI-mediated judgment before it becomes ungoverned risk.
If the great crew change is moving your hardest decisions into AI systems, the question is no longer whether you captured the knowledge. It is whether you are governing the decisions it now drives. Talk to ModalPoint about AI decision governance to put accountability, interpretation accuracy, and provenance around your AI-mediated institutional knowledge.
