Early warning, not alarm fatigue: what drilling-risk AI can learn from alarm management

By Serhii Honcharuk · Published November 3, 2026

Stuck pipe is one of the oldest problems in drilling and one of the costliest. Published estimates, many of them dated, attribute roughly a quarter of non-productive drilling time to stuck-pipe events and put the cost to the industry in the hundreds of millions of dollars a year. Treat those numbers as orders of magnitude rather than precise figures. The direction is not in doubt: catching early signs of trouble, in torque, drag, pressure or returns, is worth real money and real safety.

That makes it an obvious target for AI. It also makes it easy to build the wrong tool.

The cry-wolf problem

A model that flags every deviation will produce a stream of alerts. After a shift of mostly false ones, people stop looking, and the one alert that mattered is lost among the rest. Process industries met this problem long ago. When control rooms let alarms multiply, operators are overwhelmed, important alarms get missed and responses go wrong. The response was a body of practice, captured in standards such as ISA-18.2 (IEC 62682) and the EEMUA 191 guidance, built on a simple discipline: an alarm should exist only if someone can and should act on it.

Four questions for any drilling-risk tool

Those habits translate directly:

  1. How many alerts per rig per day, and how many led to action? Precision matters as much as detection.
  2. How early? An alert that arrives when the pipe is already stuck is a report, not a warning. Lead time is the product.
  3. Why this alert? It should say which indicators moved and how they compare with offset wells, so the driller or engineer can judge it in seconds.
  4. Who decides? The tool advises; the engineer acts. Every alert and every response should leave an audit trail.

Prove it before it interrupts anyone

The safest way to introduce a risk model is to run it quietly. Replay it against recorded wells, including the events that occurred, and count what it would have flagged and when. Then run it live in the background and compare against what the rig team actually did. Only after its alerts have earned trust do they belong in front of the people on the rig.

This is how we are approaching our own drilling-risk and well-design work: fewer alerts, earlier, with reasons, and the decision left with the expert.

Sources

← All insights