The challenge
A mid-sized North Sea operator faced a recurring problem: their well cost estimates were systematically optimistic, and the optimism only became visible late in the drilling campaign when it was too late to adjust.
The root cause was not engineering error. It was information fragmentation. Well planning teams were working with different analogue sets, different cost databases, and different interpretations of formation difficulty — without a shared framework for reconciling those differences or capturing what had been learned from previous campaigns.
The result was a planning process that looked rigorous (detailed spreadsheets, experienced engineers, regular reviews) but produced estimates with wide, unacknowledged uncertainty.
What we built
We deployed Strata as the central well planning intelligence layer for their 2024-2025 drilling campaign. The implementation had three phases.
Phase 1: Data consolidation and analogue library construction
The first challenge was data. The operator had well data spread across multiple systems — a legacy well database, engineering files in SharePoint, petrophysical data in a separate system — and no unified view.
We built a data pipeline that consolidated this into Strata's well database, applied quality filters to flag wells with incomplete or unreliable records, and built an initial analogue library covering 600+ wells across the relevant basin.
This phase took six weeks and revealed something important: about 30% of wells the team had been using as analogues had data quality issues that made them unreliable for difficulty estimation. The team knew some of these wells were problematic, but there was no systematic way to exclude them from the planning process.
Phase 2: Difficulty modelling and calibration
With clean analogue data, we built formation-specific difficulty models for the key intervals in their target geology: a shallow reactive shale interval, a high-pressure carbonate section, and a deep HPHT completion interval.
Each model was calibrated against historical performance, with explicit out-of-sample validation to check that the models were predictive rather than just explanatory. The calibration process involved the engineering team directly — not just reviewing outputs, but reviewing the analogue selections and the feature weights that drove the model.
This mattered. Two difficulty model versions were rejected during calibration because the engineering team identified that the model was using variables that were confounded with rig capability (older wells drilled with less capable rigs appeared easier, not because the geology was easier, but because the data encoded operational limits as geological behaviour).
Phase 3: Campaign-level integration
The final phase connected well-level difficulty and cost estimates into a campaign-level uncertainty model.
This is where the work paid off most clearly. Individual well estimates can look acceptable in isolation; campaign-level aggregation reveals correlation structures that are invisible at the well level. If difficulty estimates for wells in the same formation interval are correlated (as they will be, because they share the same geological uncertainty), campaign outcomes are more variable than naively summing individual estimates would suggest.
The campaign model made this correlation explicit, produced P10/P50/P90 cost estimates for the full programme, and identified which wells drove the most uncertainty — allowing the team to focus their planning efforts where they would have the most impact.
Outcomes
The 40% reduction in planning cycle time came primarily from eliminating the manual analogue selection process. Before Strata, engineers spent significant time searching for comparable wells across disconnected systems. With Strata, analogue retrieval was structured and fast, with explicit documentation of the selection rationale.
The improvement in cost estimate accuracy — from ±28% to ±14% at P50 — was the more consequential outcome. This had direct financial implications: better cost estimates meant better project economics screening and more accurate budget commitments to the organisation.
What we learned
The technical implementation was the straightforward part. The harder work was changing how the planning team related to uncertainty.
The previous process had produced point estimates because the team felt that expressing ranges would be interpreted as lack of confidence. Strata's calibration history — showing that the previous estimates had been systematically optimistic at a specific scale — gave the team the evidence they needed to make the case internally for probabilistic estimates.
The shift was not from qualitative to quantitative. The team was already quantitative. It was from false precision to honest uncertainty. That shift requires organisational trust as much as technical capability.