Build a Play‑Specific LLM: Practical AI for Field Decisions

AI doesn’t need to be a moonshot to matter. In upstream operations, the most useful models are narrow and local—trained on the history of a single play, completion style, and equipment set, then pointed at very specific choices engineers make every day.

Design it around decisions

What needs deciding? Lift changes, choke strategy, pump swaps, chemical setpoints, workover timing.

What data defends that call? Decline curves, pressures, fluid levels, completions, failure logs, and costs.

What guardrails apply? Safety limits, environmental constraints, capex windows, and SLA timelines.

How a play‑specific LLM helps

Surface look‑alikes. “Show me the five most similar wells and what worked.”

Explain the “why.” Engineers need the reasoning, not just a number.

Keep the human in the loop. The model proposes; the engineer disposes.

Build in weeks, not years

Inventory the sources (production, SCADA, work orders, costs) and standardize IDs.

Curate a small, clean training set—quality over quantity.

Write 10–15 decision prompts that mirror your actual field questions.

Publish inside a safe workflow (read‑only at first, with approvals).

Final Thought

Start with the decision, not the algorithm.

Teach the model your field one judgment at a time, and let engineers stay in charge of the call.

Hear how this mindset plays out in operations in our Cougar Energy episode with Robert Wichert on the Wisdom at the Wellhead YouTube channel.