The new State of AI survey from Vercel confirms what many of us building in the AI space have known for some time: accuracy remains the #1 technical challenge for developers building AI products.

The survey, which targeted app builders who are working with AI daily, provides a revealing snapshot of the current AI development landscape. Here's what we found most striking about the results, and how these insights align with Infactory's approach to solving enterprise AI challenges.

AI's Top Technical Challenge: Accuracy (aka Hallucinations)

According to Vercel's survey of 656 application builders, a staggering 60% identified accuracy/hallucinations as their top technical challenge when building AI features. This towers over other concerns like latency/performance and cost management, which were each cited by only 23% of respondents.

This reinforces what we've heard consistently from our customers and partners: traditional approaches to AI integration simply aren't reliable enough for business use cases. When nearly two-thirds of AI practitioners identify the same fundamental problem, it signals a market-wide issue that demands a new approach.

How Developers Are Addressing the Accuracy Problem

The survey also reveals interesting patterns in how developers are trying to solve the accuracy challenge:

• 86% don't train their own models (likely due to the prohibitive cost and complexity)

• 60% use RAG/vector databases as their primary model customization strategy

• 70% rely on manual testing to evaluate model outputs

These statistics paint a picture of teams doing their best with available tools but still struggling with fundamental reliability issues. RAG (Retrieval-Augmented Generation) has emerged as the default solution for improving accuracy, but as the survey and our lived-experience shows, it's not solving the core problem for most teams.

Data Sources: The Foundation of AI Accuracy

The survey also examined the data sources developers use to enhance their AI models:

• 48% use their own proprietary data

• 44% use customer data

• 40% use public datasets

• 40% use web scraping

• 23% use synthetic data

What's particularly interesting is how developers are balancing the use of proprietary and public data. While nearly half are leveraging their proprietary data, they're still heavily reliant on public datasets and web scraping, approaches that can introduce inconsistency and limit the competitive advantage of their AI applications. It’s the fundamental principle of “garbage in, garbage out.”

Why Current Solutions Fall Short

The survey results highlight a fundamental disconnect: while teams recognize accuracy as their biggest challenge, they're primarily addressing it through techniques like RAG that can reduce but not eliminate the problem.

RAG offers improved accuracy by providing relevant context to language models, but it still suffers from several limitations:

1. Text chunking breaks data relationships and loses critical context

2. Semantic similarity searches are inherently probabilistic, not deterministic, so there’s still risk of hallucinations and inaccuracies 

3. Inconsistent results for identical queries undermine trust

While RAG was a step in the right direction, these survey results show that “close enough” isn’t good enough and the accuracy issue still needs to be solved.  

Beyond RAG: Infactory's Unique Query Methodology™

At Infactory, we've taken a different approach to the accuracy problem. Instead of trying to incrementally improve probabilistic methods, our Unique Query Methodology™ (UQM) delivers deterministic results with complete traceability and repeatability, making your AI solution more accurate and reliable.

Unlike traditional LLMs or RAG solutions, Infactory:

1. Maintains complete data structure and relationships—no chunking, no loss of context

2. Delivers consistent, accurate results every time for identical queries thanks to deterministic processing

3. Derives answers only from your selected sources; no scraping or searching the open web

4. Provides complete lineage for every output

This approach transforms how enterprises can leverage their proprietary data, creating a foundation for AI applications that don't just reduce hallucinations but eliminate them entirely for data-driven, business-critical queries.

No Training Required: The "Use Your Data" Advantage

Perhaps most significantly for enterprise teams, Infactory requires no model training whatsoever. While 86% of survey respondents aren't training their own models (likely due to cost and complexity), most are still using suboptimal methods to integrate their proprietary data.

Infactory eliminates this tradeoff. Our platform allows you to:

• Connect directly to your data sources, whether databases, APIs, or document repositories

• Automatically generate queries that leverage your data's full context and relationships

• Deploy these queries as APIs that deliver consistent, accurate results for your end solution (chatbot, assistant, agent - whatever!)

This means you can transform your proprietary data into a competitive advantage without the overhead of model training or the security risks of using a LLM. 

A New Path Forward

The State of AI survey confirms what we've built our company around: accuracy is the fundamental challenge holding back enterprise AI adoption, and current approaches aren't solving the problem.

As teams continue building increasingly sophisticated AI applications, the foundation of accuracy becomes even more critical. Infactory's solution offers a new path forward, one that aligns with the demands of enterprise applications where "maybe" simply isn't good enough.

Source: Vercel State of AI Survey 2025