Amrize

The Hidden Cost of Basic AI Prompting

You know them by now: leading AI platforms like Gemini, Claude, and ChatGPT.  If you want to be effective using AI, it all starts with the prompt.  A prompt is the text you type into the AI tool to tell it what to do.  Feel free to ask your friendly AI what the best tips are to write highly effective AI prompts. Typically, a basic response here will be something akin to the RTF Framework. 

The RTF framework is the fastest way to structure your prompt by telling AI the role, the task, and the format. This framework is great for generating quick results.  However, this speed comes with a hidden cost. AI will then quite quickly whip you up some polished false confidence, because AI can be wrong. AI may invent laws, policies, citations and supposed facts. AI may miss source material and the business context. This list goes on.

Recognizing the Limits of AI’s General Knowledge

I recently put AI through its paces with one of my standard interview questions. It failed. When I called it out on OSHA recordkeeping exceptions for stitches, its response was: “Touché, you got me there!” Admittedly, outsmarting AI felt good for a second, but a “Touché?”  It was as if AI just looked into my soul.

AI optimizes answers for the 80 percent use case. Call it ‘generalist mode.’  AI prioritizes high-probability answers and ‘language neutrality.’ 

The response is not guaranteed to be true. Helpful? Yes. Confident? Yes. Wrong? Sometimes. A very simple improvement step you can take here is to ask AI, “are you sure?” Look out AI may confidently double down on the original answer.

Let AI help you draft things faster, but don’t outsource your safety decisions and far-reaching choices to the false confidence in the box.

Making Expertise the Final Check on AI

The problem? Safety isn't found in the 80 percent. It’s found in the 20 percent—the edge cases, the legal exceptions, and the technical nuances. The takeaway: use AI to point the way, but never let it be the final word. You must apply your experience and judgment.

If you don't dig into the "Specialist" weeds, you’re just providing a secondhand, middle-of-the-road guess. The subject matter expert verifies and decides after AI generates.

Keeping Human Judgment at the Center of AI Use

Let AI help you draft things faster, but don’t outsource your safety decisions and far-reaching choices to the false confidence in the box.  Just as risks can be controlled, so can your AI output. A comprehensive review requires verifying the fundamental facts.

Reviewers must also ensure source alignment by confirming that the material matches.  Throughout the process, strict attention must be paid to sensitivities, all while exercising sound judgment that supports your conclusions and decisions.

Be better than the 80 percent use case, because that 20 percent is what makes you irreplaceable.