These are anecdotal observations on how AI, specifically LLMs, was used in software engineering in 2025, particularly within the context of an existing complex system with a large user base.

Good and High-Impact Use Cases

1. Automation This was one of the most successful use cases. AI was used to automatically group customer support requests to detect larger systemic issues and to enrich incident alerts by summarizing relevant context from different sources. However, skepticism remained regarding AI-driven resolutions. In several instances, the AI made recommendations that would have been detrimental to the resolution effort.

2. Prototyping and “Throwaway” Work AI was useful for creating prototypes, mocks, and disposable assets to demonstrate an idea or prove its viability. It was especially helpful for creating UI mocks and quickly iterating on them. While not a replacement for diligent, human-centered UI design, a picture is worth a thousand words. These mocks were helpful in getting humans to agree on a design quickly.

3. The Personal Editor It was used as a personal editor and reviewer for better writing. It acted as a second pair of eyes for providing feedback on both documentation and code, helping engineers refine their communication. This is particularly helpful when engineers need to communicate with different audiences, or if English is their second language as in my case.

Good but Low-Impact Use Cases

1. Marketing and Engagement AI was good for quickly generating social posts, videos, and blog content. It requires human review, but it is good at summarizing existing technical material and mapping it to specific marketing domains. From the perspective of an engineering organization this is a time saver, but it is not impactful enough to change engineering roadmaps.

2. Low-Priority Code Generation It was used for generating low-stakes code, writing boilerplate, running tests, and iteratively improving scripts for side tasks. However, it isn’t trusted for complex or production-critical code paths yet.

3. Just for Fun Generating images for internal slide decks, memes, chats, and team culture assets. Simple and fun.

Poor Use Cases

1. Enterprise Support AI is not a good replacement for enterprise support. In this context, customers are paying for expert-level human interaction. Using AI here often frustrated users who expected a person.

2. The “Fluff” Generator One of the poor uses was the volume of long, AI-generated emails and documents shared “as is.” It wasted people’s time sifting through bloated text to find the actual point. This frustration led to one of my favorite quotes: “if you are going to share an AI generated document, please kindly share the prompt instead”. I think that will remain to be great advice in the coming years as well.