The AI Solution Layer
Nine capability classes organized by the constraint each relieves. Six of the nine verdicts are not "build" — and that is the argument.
Faster parts, same system
A use-case catalogue would make this proposal worse, and the evidence for why is already published.
One advisory firm’s upstream digital use case library counts more than 160 digital use cases across the upstream and midstream value chain — thirty in exploration, thirty-five in drilling, sixty-six in production — and the taxonomy contains no land or regulatory category at all; the closest entry is six shared-services cases internal .
A Digital Acceleration Index study of upstream operators found nine in ten had some form of digital vision, and only 15% had multiple high-value use cases that reinforced each other enough to support end-to-end decision making internal . A VP of Upstream Technology at a supermajor, quoted in that study: "Digital in upstream O&G has in the past decades delivered a lot of promises, but not a lot of value."
The compliance-function perspective names the failure mode precisely — use-case AI "improves a slice but preserves the broken assembly line," delivering "faster parts, same system."
The most useful thing this research found is an absence
Not one vendor in this market publishes an accuracy figure for extraction on oil and gas land instruments — not the data majors, not the land system incumbents, not the legal-research franchises, not the cloud document-AI providers certain . The category being sold is an unmeasured capability. That is a negotiating asset rather than merely an observation.
Ask any vendor for a held-out accuracy score on pre-1950 handwritten assignments in a named county, and watch what happens.
Nine capability classes
Nine capability classes follow, organized by the constraint each one relieves rather than by the technology. The verdict is the point of the section.
Six of the nine verdicts are not "build", and that is the argument
The standard build-or-buy framework for AI places a capability on two axes: value potential relative to competitors, and differentiated data access relative to the AI vendor. Extraction sits in the commodities quadrant — optimize turnkey services, share the data, take the cost reduction.
The identity graph, deficiency prediction and the obligation engine sit in the gold-mine quadrant, because the operator holds filing history, lease records and well identity crosswalks that no vendor has and cannot obtain.
AI defensibility work reaches the same conclusion in different language: generative, predictive and conversational AI are table stakes, and agentic capability is where a company wins and differentiates. A program that builds everything arrives late and over budget. A program that builds nothing has bought a license and called it a transformation.
A sequence, not a portfolio
The order is not a preference — three of the four depend on the second, and attempting them in any other sequence produces exactly the reinforcing-use-case failure the sector has already measured.
Deficiency prediction
Quarter 1The operator owns the data, the benchmark is its own historic first-pass rate, and it can be evidenced inside a quarter.
Canonical identity graph
Months 1–18Runs in parallel, funded from the first win and defended as infrastructure rather than as an AI use case.
Cross-domain obligation engine
Months 12–36Depends on: Canonical identity graphThird, and only once the graph is real. Start with the regulatory deadline corpus, not the contract corpus.
Production and emissions lineage
Months 18–42Depends on: Canonical identity graphFourth. It rides on the same graph. Do not open a second front before the graph is real.
One caution on the headline number
The 183 operator days are not all recoverable, and it is better to say so before a client does. 43 CFR 3171.12 and BLM Instruction Memorandum 2013-104 give the agency ten days to notify an operator that a package is incomplete and give the operator forty-five days to cure certain .
Some share of "waiting on operator" is therefore a statutory cure window rather than operator slack, and the decomposition should be obtained before anyone promises the days back. The useful nuance is that the remedy is unchanged either way: a package that is complete on first pass never enters the cure clock at all.
Calibration, from the buyers themselves
A 2025 survey of 283 agentic AI buyers found 21% reporting that return on investment came in below expectations and a further 29% unable to tell either way, against 17% who saw it exceed. Seventy-one percent associate agentic AI with "requires more expertise" and 57% associate the non-agentic alternative with "lower risk" internal . This is not a market in which a list of use cases gets a program funded. It is a market in which the buyer has probably been disappointed once already and will test the second proposal considerably harder.