Paul BarnhurstFP&A Expert Empowering Finance Professionals to Excel Everday

You're still the last line of defence.

1 Sept 2026 · 1 min

Watch on YouTube

0:45

In this brief YouTube clip, Paul Barnhurst and another speaker discuss the realities of AI model development. They note that no model—human or artificial—comes out perfect on the first try, emphasizing the need for continual instruction, guidance, and iterative refinement. The conversation highlights the critical role of the human modeler as the "last line of defence," using a recent example of training Claude where the model still missed explicit instructions despite extensive refinement. The speakers argue that beyond good prompts, effective model building requires layered checks and balances that enable easy human review, reinforcing why models are constructed with such oversight mechanisms.

Chapters

I haven't seen one model yet that's been perfect. Not that I've quite seen one from a human that's perfect, at least in the first try. But you have to check it. You have to instruct it and guide it and spend time with it. So I would agree it becomes more important.
Yes. And I would also say that you're still the last line of defense, right? As the human modeler. We had an example recently where we were training, teaching Claude a skill, and we had been refining it for a while and we noticed that it had still missed a few. It didn't follow our explicit instructions. So my point is, it's more than just giving good instructions and good prompts. It is also layering in those checks and balances that will make it easy for you as a human to do that review, which comes back to exactly how and why we build models the way we do.