Just 25% of companies have deployed AI into production, while a further 14% are still in the beta testing phase, according to Vellum’s inaugural State of AI Development Report. 

The AI development platform’s survey of over 1,200 AI developers across industries including technology, healthcare and finance, found that a quarter of firms (25%) are still forming their AI strategy in the pre-building phase, while only 21% have got as far as proof of concept.

More than half of developers reported building both document analysis tools (59%) and customer service chatbots (51%), with just 25% creating AI for code generation and automation, while 15% are building compliance applications.

Looking ahead, the emergence of agentic AI from concept to development is seeing progress, with over half (55%) of developers planning to build complex agentic workflows this year.

Despite companies having had more options last year regarding how they wanted to access AI models, OpenAI continues to dominate with 63% of developers relying on it as an API provider, followed by 34% using Microsoft and Azure, and 32% continuing to use Anthropic.

However, notable mentions in the report also went to model providers hosting open-source models, such as Llama 3.2 70b, including Groq, Fireworks AI, and Together AI. 

While developers are steaming ahead on AI projects that save time and costs, they aren’t placing the same emphasis on their evaluation processes, with only 57% evaluating their AI models and three in four (75%) doing so manually.

Among the biggest challenges developers said they had faced when building AI products was managing Al hallucinations and prompts (57%), prioritising use cases (42%), and a lack of technical expertise (38%).

Slightly less said they were worried about model speed and performance (33%), while a similar number reported issues with data access and security (32%).

Getting past those challenges can bring significant benefits, including advantages over competitors (31%), cost and time savings (27%), and higher user adoption rates (12%), though almost a quarter (24%) of developers said they had seen no measurable impact yet from their AI products.

Recommended reading

• UK Government Launches AI Safety Hub
• Starmer Unveils New 50-point Plan to “Turbocharge” UK AI
• UK Gov’s New AI Sector Study | Key Data & Insights

As to who is actually building these AI tools, Vellum found that while those engineering were the most involved in projects (82.3%), there is a high level of collaboration with other teams, including from those in leadership roles (60.8%), subject matter experts (57.5%), product teams (55.4%) and design departments (38.2%). 

“We’re witnessing a transformation in how businesses leverage AI,” said Akash Sharma, CEO of Vellum.

“The AI industry’s, albeit painstakingly slow, attention to quality control, combined with rapid innovation, shows a sector that’s maturing in the right way. Organisations are balancing ambition with responsibility, creating a sustainable path forward for AI adoption.”