The 2025 state of AI Development

The survey sorts teams by how far they've taken AI, from talking to users about needs up through a strategy and proof of concept, a beta, and finally a live production deployment. Roughly a quarter of companies had something in production, with the rest spread across earlier stages. The breakdown matters because it shows the field is not one wave but a spread, so a team that feels behind is often just at the same stage as most peers.

Document parsing and analysis came out on top as the most common use case, followed by customer service and chat assistants, then natural language analysis. Content generation, recommendation systems, code generation, and research automation filled out the list. The pattern points to work that involves reading and sorting existing text, not just producing new text, which is a useful signal about where the demand sits.

OpenAI held the largest share of usage, largely through GPT-4o and GPT-4o Mini, with Microsoft Azure, Anthropic, and AWS Bedrock close behind. Larger companies leaned more on cloud platforms, and Azure's share climbed with company size. Hosted open models from providers such as Groq, Fireworks AI, and Together AI picked up ground, which the report reads as a sign that teams want more than one option in the mix.

Retrieval-augmented generation dominated the architectures teams described, with a majority using a vector database. Only about a third fine-tuned a model at all. The report ties the low fine-tuning figure to falling costs and rising model quality, since a good base model plus retrieval often covers the job without a custom training run.

Most developers said they evaluate their AI, but the methods skew manual, with test-and-review far outpacing automated evaluation tools. Monitoring split similarly, with in-house dashboards ahead of third-party tools. Managing hallucinations and prompt issues ranked as the top challenge, which lines up with how thin the automated checks still are.

The report closes with a forward look, predicting that 2025 turns into a year of tooling and that agentic workflows move from experiments to shipped features. It also flags rising multi-model setups and continued cost pressure as forces that shape what teams build next.