By Saad Iqbal
Every production engineer has sat in the same Monday meeting: one rig, one wireline unit, and a list of forty wells that all “need something.” Somebody has to decide which three or four actually get worked this month, and for decades that decision leaned on whoever shouted loudest or whichever well a VP happened to drive past. The wells that quietly lost 30% of their rate over six months because of a worn ESP or a gassed-up rod pump often lost out to the one well with a dramatic, visible failure — even when the quiet decliner would have paid back the workover cost twice as fast. AI well intervention candidate selection software exists to replace that gut-feel ranking with a number every well in the stock gets, calculated the same way, every week.
This piece compares four platforms — SLB’s Delfi ProdOps, Prometheus Group, Quorum Software and Cognite Data Fusion — that are actually being used to rank workover and well-intervention candidates in upstream oil and gas, and walks through the mechanics of how a defensible candidate score gets built in the first place.
How does AI well intervention candidate selection actually work?
Strip the dashboards away and every workover optimization software platform is running the same four-stage pipeline underneath:
- Full well stock. Every producing well on the asset starts in the pool — not just the ones someone already flagged, because the point of the exercise is catching the wells nobody has noticed yet.
- Surveillance triggers. Each artificial-lift method throws its own early-warning signals: ESP amperage drift and vibration trends, rod-pump dynamometer card shape changes, gas-lift injection-rate anomalies, declining productivity index on naturally flowing wells.
- AI candidate ranking. A model combines expected production uplift, cost to execute, data confidence and time-to-payback into one score per well, instead of leaving an engineer to mentally juggle four separate spreadsheets.
- Prioritized workover queue. The ranked list gets routed against actual rig, wireline or coiled-tubing unit availability, so the queue reflects what crews can realistically execute this quarter, not an infinite wish list.
The platforms below differ mainly in where they sit in that chain: some are strongest at generating the ranked score itself, others at scheduling the crews once you know which wells matter.
The best AI well intervention and workover candidate selection tools
1. SLB Delfi ProdOps — Well Intervention Opportunity Screening
SLB’s intervention screening tooling inside the Delfi ProdOps environment is built specifically around automated candidate ranking: it scores wells against expected uplift and intervention cost, then feeds a prioritized shortlist directly to the planning team rather than leaving ranking as a side calculation in a spreadsheet. It’s the strongest pure-play candidate-ranking tool of the four, and a natural fit if your production data already lives inside an SLB digital environment.
2. Prometheus Group
Prometheus Group’s AI-driven planning and scheduling tooling (built out further through its Actenum acquisition) focuses less on deciding which well deserves attention and more on the second, equally painful problem: once you know the ranked list, how do you actually sequence rigs, crews and parts against it without creating a scheduling conflict. For an operator whose candidate list is already reasonably solid but whose workover scheduling is the real bottleneck, this is the piece that closes the gap between “we know what to do” and “we actually did it on time.”
3. Quorum Software
Quorum’s upstream production operations suite pulls SCADA, well-test and financial data into one system of record, which is the multi-source data integration layer that a lot of candidate-ranking models quietly depend on and rarely get credit for. Where Quorum earns its place on this list is as the data backbone an operator builds a custom or vendor ranking model on top of, especially across a diverse, multi-vendor artificial-lift fleet.
4. Cognite Data Fusion
Cognite’s industrial DataOps platform contextualizes sensor data, well models and maintenance history into a single data layer, then lets an operator build its own candidate-ranking logic — including hybrid physics-and-ML models — on top of contextualized, trustworthy data rather than raw historian tags. It isn’t a ranking tool out of the box the way SLB’s is, but for an operator with the in-house data science capability to build a custom scoring model, Cognite is frequently the foundation that makes that model reliable at scale.
SLB gives you the fastest path to a working candidate score if your data already sits in its ecosystem. Prometheus Group matters most once ranking is solved and scheduling is the bottleneck. Quorum and Cognite are both infrastructure plays — the difference is that Quorum leans toward an out-of-the-box production system of record, while Cognite leans toward a flexible data layer for teams who want to build their own scoring logic.
What actually goes into a defensible candidate score
A candidate-ranking model that only looks at production decline will happily send a rig to a well that’s declining fast but genuinely uneconomic to fix. The scoring logic that holds up under scrutiny weighs at least four inputs together:
- Production decline relative to type curve — how far the well has fallen below its expected decline trajectory, not just its absolute rate, since a well declining on-curve may not be a candidate at all.
- Expected uplift — a nodal-analysis-informed estimate of incremental rate if the suspected problem (worn ESP, gas interference, scale) is actually fixed, not just “the well will improve.”
- Data confidence — how many independent signals agree. A well flagged by amperage drift, falling productivity index and a dynamometer card change together is a far safer bet than one flagged by a single noisy sensor.
- Cost to execute — rig or wireline day-rate, expected job duration, and parts lead time, since the highest-uplift well in the world isn’t a good candidate if the intervention costs more than two years of incremental revenue.
Score these four independently, normalize them to a common 0–100 scale, and weight them according to what your asset actually cares about this quarter — capital-constrained assets weight cost-to-execute harder, while assets chasing a production target weight expected uplift harder. The wells worth watching are the ones that score high on decline and uplift but low on data confidence: those are your re-surveillance candidates, not your next workover.
Why timing the intervention matters as much as picking the well
The economic case for AI-ranked intervention isn’t really about picking a different well than an experienced engineer would eventually have picked — it’s about picking it earlier. A reactive approach catches the problem at unplanned failure: the well goes down hard, downtime stretches while a crew gets mobilized on short notice, and the operator eats weeks of zero production before the well comes back online. An AI-ranked approach catches the same underlying problem from its early surveillance signature — weeks or months before failure — and schedules a planned workover into an already-booked crew slot, so downtime shrinks to days instead of weeks.
That’s the real return on investment: not a smarter well pick, but a shorter gap between “something is wrong” and “it’s fixed,” repeated across every well on the lease.
Common pitfalls when trusting a candidate-ranking model
- The model doesn’t know about planned shut-ins. A well flagged for a steep decline that’s actually due to a scheduled facility turnaround will rank falsely high — always cross-check the ranked list against the operations calendar before dispatching a crew.
- Uplift estimates assume the diagnosis is right. An uplift number calculated assuming “worn ESP” is worthless if the real problem is a tubing leak — treat expected uplift as conditional on the diagnosis being confirmed in the field, not as a guarantee.
- Static cost assumptions age badly. Rig and wireline day-rates swing with the market; a scoring model running on a stale cost table will misrank the whole queue when rates move even 15–20%.
The wells that quietly cost an asset the most money are rarely the ones making the most noise. A candidate-selection model that scores the full well stock every week, not just the wells someone remembered to look at, is what actually closes that gap — and it pairs naturally with the artificial-lift surveillance tools that feed it. For more on the lift-specific signals that usually trigger these candidate scores, see Best AI Tools for ESP Failure Prediction & Run-Life Optimization, Best AI Gas Lift Optimization Software for 2026, and Best AI Rod Lift & Pump-Off Controller Optimization Tools.
