The reserves engineering gap — three workflow gaps the 2024-26 literature is calling out

If you talk to a reservoir engineer at a super-major today, you'll hear about digital twins, agentic AI orchestrating hundreds of well models, real-time SCADA-to-simulation feedback loops, and audit-grade probabilistic reserves pipelines. If you walk into a small E&P operator's office, a family office managing inherited mineral interests, or a mid-cap fund underwriting a non-op acquisition, the workflow is a spreadsheet, a 14-year-old DCA tool, a Word document with screenshots of decline curves, and a once-a-year PDF "Third-Party Reserves" report stitched together by hand. The gap is real, and the 2024–2026 reservoir engineering literature has started naming it explicitly.

This post walks through the three workflow gaps that recently published research has documented. The diagnosis is consistent across the literature: the gap between leading-edge reservoir engineering and what most operators below the top 30 by enterprise value actually use is infrastructural, not algorithmic. The capability exists. The integration doesn't.

The companion post walks through what the super-majors have actually built — Equinor's Johan Sverdrup digital twin and the SLB/ONGC agentic AI framework. A third post covers the reference architecture for what's now buildable from open-source components at the bottom of the market.

The opening contrast

Two reservoir engineering workflows, both real, both currently in production:

This isn't a critique of the engineers on either side. The engineers at small and mid-cap operators are typically excellent — they know their basins, their geology, their operator behaviors. What they lack is the workflow infrastructure that turns engineering judgment into a repeatable, auditable, scalable pipeline. That's the gap, and it's the same gap the 2024–26 literature has converged on.

Gap 1 — Data and model fragmentation, not algorithm scarcity

J.T. Dewan's 2024 Journal of Petroleum Technology feature, "Beyond the Buzzwords: A Practical Blueprint for Sustainable Reservoir Management," identifies the bottleneck directly. The problem isn't a lack of ML algorithms, a lack of physics-based simulators, or a lack of analytical methods. The problem is that production surveillance lives in one tool, reservoir models live in another, economics live in a third, and the feedback loop between them is a human reservoir engineer copy-pasting between them on a quarterly cadence.

Dewan's framing — "living models," shared data and model repositories, closed-loop feedback — reads less like a research finding and more like a product requirements document. The capability to do the analysis exists in commercial and open-source toolchains. What doesn't exist (for most operators) is the glue layer that keeps everything in sync without a human bridging it manually.

The specific symptoms of fragmentation that show up everywhere:

• Production data, well master data, division order data, and forecast assumptions live in different tools with different update cadences and no shared identifier
• Three different "well ID" conventions — RRC permit, API number, operator-internal — none reconcile cleanly without manual cleanup
• Reservoir model output and the economic model that consumes it are not connected; updating one requires manual re-export to the other
• Type curves, declines, and basin assumptions live in an analyst's notebook that nobody else can edit or audit

Each symptom is fixable individually. None is fixed by adding a smarter algorithm. The fix is integration plumbing — and integration plumbing is the part nobody at most small operators has been resourced to build.

Gap 2 — DCA at scale is still one-well-at-a-time

The 2024 JPT — The Way Ahead survey article "Petroleum Engineers in the Age of AI: Adapt or Become Obsolete?" by H. Al-Mubarak, and the Novi Labs industry reporting it references, both note that even where ML-based production forecasting exists, it tends to exist as siloed prototypes rather than integrated workflow.

Most operators still run decline curve analysis interactively, one well at a time, in tools that are functionally unchanged from twenty years ago. The recurring failure modes:

• No shared, standardized batch DCA pipeline producing P10/P50/P90 distributions across a full portfolio on a refresh schedule
• ML forecasts (where they exist at all) live in a data scientist's Jupyter notebook, not in the reserves report that reaches finance and the auditor
• Type curves are picked manually per well, with no consistent regime segmentation across the book
• Sensitivity analysis is a one-off Excel exercise rather than a routine output of every forecast

For an operator with 50 wells, this is annoying. For an operator with 500 wells, or a fund underwriting a 5,000-well acquisition, it's a structural cap on the speed and quality of every decision. Diligence runs slower, and the decisions that come out of it are based on the wells the analyst had time to model carefully rather than the full set the deal actually covers.

Gap 3 — Reserves reporting is a Word document, not a workflow

The SPE Annual Technical Conference 2025 produced work — referenced in the February–March 2026 JPT technical-paper listing — on probabilistic reserves estimation under PRMS 2018, including Monte Carlo workflows with automated audit trails.

The fact that this is new should tell you everything. PRMS 2018 has been the global reserves classification standard since 2018. SEC reserves reporting for US-listed E&P has had auditable disclosure requirements for over a decade. Yet at the small/mid-cap end of the market, third-party reserves reports (TPRs) are still:

• Generated by an outside firm once a year
• Delivered as a PDF with no underlying queryable data
• Stitched together from Excel models with no version history
• Re-baselined manually every reporting cycle
• Almost impossible to audit when an input assumption is challenged six months later

The published work is finally describing what an integrated, versioned, queryable reserves workflow looks like — change-point detection on water cut and GOR feeding an automated re-classification when a well's behavior diverges from its forecast envelope, with the audit trail attached as queryable data rather than a folder of PDFs. The tools to build it are off-the-shelf. The execution isn't happening at most operators below the top 30 by enterprise value.

The diagnosis the literature has converged on

Read across Dewan, Al-Mubarak, the 2024 Lynch precision water shutoff paper, the 2025 SPE ATCE work on anomaly detection and probabilistic reserves, and the Novi Labs hybrid ML+DCA approach reporting, and a consistent diagnosis emerges. The gap isn't algorithmic. It's the glue between the algorithms.

The specific things that are missing at the small/mid-cap and mineral-owner end of the market:

1. A canonical data layer — a queryable database that reconciles state-agency permit IDs, API numbers, operator-internal references, and division-order entries into a single well registry
2. A batch forecast pipeline — DCA, ML uplift, and Monte Carlo sensitivity running across the full portfolio on a schedule, not interactively one well at a time
3. A versioned reserves classifier — PRMS 2018 logic that produces 1P/2P/3P categorization with the input snapshot, code version, and assumption set attached as audit trail
4. An anomaly surveillance loop — change-point detection on rate, water cut, and GOR that surfaces wells whose behavior diverges from their forecast envelope

Each of these is buildable from open-source components. None of them require new algorithms. All of them require integration engineering — and that's the work nobody at most operators has been resourced to do.

The takeaway

The 2024–26 reservoir engineering literature has converged on a clear diagnosis: the productivity gap between leading practice and small/mid-cap reality is in workflow infrastructure, not in any single algorithm. Dewan named the fragmentation problem in 2024. Al-Mubarak named the siloed-prototype problem the same year. The SPE 2025 work named the reserves-reporting workflow problem. The Novi Labs reporting named the hybrid ML+DCA pattern that's now technically straightforward but operationally rare.

None of these diagnoses are controversial inside the industry. The interesting question — and what the next post in this series covers — is what the super-majors have actually built to close the gap on their side, and what that reference architecture means for the rest of the market.

Sources

• J.T. Dewan, "Beyond the Buzzwords: A Practical Blueprint for Sustainable Reservoir Management," Journal of Petroleum Technology, 2024 — jpt.spe.org
• H. Al-Mubarak, "Petroleum Engineers in the Age of AI: Adapt or Become Obsolete?" JPT — The Way Ahead, 2024
• P. Lynch et al., "Unwanted Water Is Costing You Value. How Precision Shutoff Restores Control," JPT, 2024
• SPE Annual Technical Conference 2025 — probabilistic reserves estimation under PRMS 2018; real-time water-cut and GOR anomaly detection (papers listed in JPT Feb–Mar 2026 technical-paper roundup)
• Novi Labs industry reporting on hybrid ML + DCA at scale (referenced in 2024 TWA piece)
• SPE PRMS 2018 standard

Need to turn this engineering logic into a production-ready tool? Explore built custom engineering workflows for your team.