AI & Automation

AI for Horizontal Well Placement: What Works Today and What Does Not

AI in well placement covers everything from chat assistants to automatic correlation. A plain guide to what is useful on a live well today, what is still being proven, and six questions to ask before you trust any of it.

8 November 2026

"AI-powered" now appears on almost every geosteering product page. It can mean very different things: a chat window that explains terms, an assistant that reads your project, a score on how well an interpretation fits, or software that proposes the interpretation itself. Each has a different track record, and each deserves a different level of trust on a live well.

This guide sorts them by what is useful today and what is still being proven, and ends with six questions to ask before any of it gets near your steering decisions.

Four Kinds of AI in Well Placement

Kind What it does Ready for a live well?
General chat assistants Explain concepts, draft reports, help with formulas Yes, for knowledge work; not for numbers about your well
Assistants connected to your project Read your wells, logs and interpretations; answer with calculated numbers Yes, when the numbers come from a real engine and changes need approval
Interpretation checks Score how well each interval of a correlation fits the type well Yes, as a second opinion
Automatic correlation and steering Propose the interpretation, or the steering, on their own Still being proven; a proposal at most

1. General Chat Assistants

Tools like Claude and ChatGPT are good at explaining how apparent dip relates to true dip, drafting an end-of-well report, or checking a spreadsheet formula. They have not seen your well, though. Ask one for the stratigraphic position of your last survey and it can only guess, or calculate from numbers you typed in, in text, with no way for you to check the method.

Use them for knowledge, not for positions, distances or dogleg numbers on your well.

2. Assistants Connected to Your Project

The useful step up is an assistant that works inside your geosteering software, or connects to it. It reads the project you have open, runs the calculation with the same engine the software uses on screen, and answers with the numbers it used. The language model understands the question; it does not do the arithmetic.

On a live well, this saves real time. "Where is the last survey point relative to the sweet spot?" "Which intervals of this interpretation are weakest?" "How far are we off plan, and what would it take to get back?" Each of these used to mean opening several screens or exporting to a spreadsheet.

What separates a safe assistant from a risky one is what happens when the answer leads to a change. A safe one prepares the change and shows it to you exactly as it will be saved. Nothing is written until you approve it.

3. Interpretation Checks

Correlating a lateral against a type well is judgement, and two experienced geosteerers can disagree. Software can measure how well each interval of an interpretation fits: how closely the lateral's gamma matches the type well at the stratigraphic depth the interpretation claims. Low-scoring intervals are where to look again.

This works today as a second opinion. A score is not a verdict: a flat-gamma interval can score well under several different interpretations, and a correct one can score poorly where the logs are noisy. Treat it as a flag for attention, not as an answer.

4. Automatic Correlation and Steering

This is where most of the headlines are, and where the record is thinnest. Software that proposes the whole interpretation, or the steering, has to be right not on average but on the interval that decides whether the well exits the zone. Experts reject these systems for good reasons, which our CEO set out in why autosteering is never accepted by industry experts.

The more promising direction is software that learns from the engineer instead of replacing them: it takes the expert's own verdicts on past interpretations, and its proposals are judged against that judgement and approved by a person before they count. Even then, the output should be a proposal you can inspect, never a change made on its own.

A related area that needs no machine learning at all is planning ahead of the bit: working out which targets the bit can still reach under a maximum dogleg, and turning the chosen one into a plan. That is geometry, and it can be checked to the decimal. We describe our approach in Introducing LookAhead.

Six Questions to Ask Before You Trust It

  1. Where do the numbers come from? From the software's own calculation engine, or from the language model? Only the first can be checked.
  2. What can it change without you? The right answer is nothing. Every change to an interpretation or a plan should wait for a named person's approval, and show exactly what will be written.
  3. Is there a record? Can you see later who approved which change, and when?
  4. How is it tested? Is every release run against a fixed set of cases, including safety cases, before it ships? Or is it judged on demos?
  5. Does it use your data to train? For whom? Your data should improve your results only, never another customer's model.
  6. Does it know where the bit is? Most positions are calculated at the last survey station, which sits behind the bit. A tool that talks about "the bit" without saying so is projecting, and should say how.

Where GeoMaster Stands

In GeoMaster®, registered in the United Kingdom (UK00004424070), AI sits in the second and third categories today. GeoEngine, the AI assistant in every project, reads your wells, logs and GeoMatch interpretations and answers with numbers from GeoMaster's own geosteering engine. It reports where the last survey point sits in your Sweet Spot, scores each interval of an interpretation, explains DrillTracker, and offers to save a version after a significant edit. Every change it prepares waits for your approval, and each approval is recorded. Every release must pass a fixed set of geosteering test cases, including safety cases, before it ships.

Autosteer, which learns from your own verdicts on intervals and proposes an interpretation for you to approve, and LookAhead are in development and not available yet. See what GeoEngine does today.

GeoEngine is included in every plan, including the free one. Start free, or request a 14-day Pro trial; we approve requests within 1 working day.

Frequently asked questions

Written by the GeoMaster team at Greenstone Energy Ltd

Geosteerers and drilling engineers building geosteering software in the browser. Meet the founders →

Work With Us

Run your next well in GeoMaster. Start on the free plan, or request a 14-day Pro trial: we approve requests within 1 working day and you won't need a card.

Read Next

AI & Automation

Using Claude and ChatGPT for Geosteering: What an MCP Connector Can Do

Claude and ChatGPT know a lot about geosteering but nothing about your well. An MCP connector changes that. What a geosteering connector can do, what it cannot, and what to check before you connect one.

8 November 2026Read →
AI & Automation

We Asked 20+ Interpretation Experts and None of Them Trusts Autosteer

We asked more than 20 interpretation experts. None trusts autosteer, because it invents faults to match the log. What we presented at the LPS New Technology Seminar 2026 on agentic AI for LWD interpretation in horizontal wells.

6 November 2026Read →
AI & Automation

Closed-Loop Execution vs Open-Loop Advisory: Integrating Geosteering into a Drilling Automation System

How integrating subsurface interpretation with drilling automation systems improves wellbore placement and reduces drilling hazards.

28 August 2026Read →