I recently came across market research data on AI Agent engineering applications, which I found quite interesting. Here are some key findings:

In terms of practical applications, customer service has become the most popular scenario, accounting for 26.5%. Next are research and data analysis (24.4%), followed by internal collaboration and productivity tools (17.7%). Code generation ranks fourth, making up only 9.8%. This may differ from many people's expectations.

However, the biggest obstacle remains quality issues. 32% of practitioners list it as the primary barrier, covering accuracy, relevance, consistency, and other dimensions—consistent with pain points from last year. Cost concerns, on the other hand, are less prominent this year.

Interestingly, 89% of companies have installed some form of observability tools on their Agents, with 62% achieving relatively fine-grained tracking. This indicates that everyone recognizes the importance of tracking multi-step reasoning and tool invocation capabilities; it’s no longer optional.

Regarding model selection, although OpenAI’s models are used by over two-thirds of organizations, multi-model ensembles are mainstream—more than 75% of organizations run multiple different models simultaneously. Interestingly, one-third of organizations are still investing in building their own model infrastructure, showing that beyond API convenience, on-premises deployment remains attractive to enterprises.

As for fine-tuning, it’s still somewhat niche. 57% of organizations do not perform fine-tuning at all; instead, most meet their needs through base models combined with prompt engineering and Retrieval-Augmented Generation (RAG).

This data mainly comes from B2B enterprise users and should reflect the current practical situation.