Not Faster, but Better: Experience and Guidelines from a Year of AI Use in Groundwater Modeling

Abstract

This article describes my use of artificial intelligence (AI) tools over 1 year of consulting work in groundwater modeling, from August 2025 through August 2026. I expected that AI would make me more efficient with my existing process—that I would simply get twice as much done. That is not what happened. A task that took 8 h before AI still takes 8 h. What has changed is how I use those hours. I used to spend 2 h planning and 6 implementing, and this balance of time has reversed. In my opinion, the result is a better product, not a faster one. Here I describe the process I have settled upon: a staged workflow in which different models, or one model under different assigned roles, help plan an approach, critique it, and then implement it. The workflow is built around three behaviors of large language models—they perform better when given a role, they are biased toward agreeable answers, and they degrade as context fills. I offer guidelines for working within those limits and recommendations for practitioners starting out. I also briefly discuss the costs, both direct and indirect, and where our experience with water and its allocation can be brought to bear.