Key takeaways: Developer trust in AI output is declining. Over 75% of developers still want human validation when they don’t trust AI answers. Debugging AI-generated code takes more time than expected, with “almost right but not quite” solutions being the top frustration. Advanced questions on Stack Overflow doubled since 2023, indicating that LLMs may struggle with complex reasoning problems. Agentic AI adoption is split: More than half of developers are still sticking to simpler AI tools, but 70% of adopters report reduced time on tasks thanks to agentic workflows. Small language models and MCP servers are emerging as cost-effective solutions for enterprise and domain-specific tasks. The 2025 Stack Overflow Developer Survey gives us a nuanced look at AI adoption among enterprise development teams. AI tools are widely used, but as adoption rises and developers bump into the real-world limits of their shiny new tools, trust declines accordingly. At the same time, the survey underscores the value developers continue to place on human knowledge and experience, especially as AI tools become more unavoidable. On a recent episode of Leaders of Code, Stack Overflow Senior Product Marketing Manager Natalie Rotnov highlighted what enterprises should take away from these findings, especially around AI adoption and implementation. Here, we’ve distilled Natalie’s take on the survey findings, laid out some action items for leadership, and dug a little deeper into her recommendations around agentic AI for the enterprise. Spoiler alert: It all comes back to data quality. The AI trust decline: why developer skepticism is healthy Stack Overflow’s 2025 survey of nearly 50, 000 developers around the world revealed that developer trust in AI tools is declining. This probably doesn’t surprise any developers out there, but it might come as a surprise to the C-suite if they’ve been bullish on AI tools but not necessarily attuned to how their teams work. According to Rotnov, developers’ skepticism of AI is healthy. “Developers are skeptics by trade,” she explains. “They have to be critical thinkers, and they’re on the front lines intimately familiar with the nuances of coding, debugging, and problem-solving.” Aren’t those exactly the people you want working with brand-new AI coding tools? What’s behind the AI distrust? The survey identified developers’ biggest frustrations with AI: “Almost right, but not quite” solutions. AI produces code that appears correct but contains subtle errors. These create pitfalls, especially for less seasoned developers, who may not have the experience to identify and correct these issues. Time-consuming debugging. Fixing AI-generated code often takes longer than expected, especially without proper context. Lack of complex reasoning. Current AI models struggle with advanced problem-solving and higher-order work. These concerns align with research findings. Research from Apple suggests that LLMs primarily engage in pattern matching and memorization rather than true reasoning. The paper showed that as tasks grew more complex, model performance deteriorated-evidence that reasoning models are still relatively immature. Key term: Reasoning models are AI models designed to break down problems and think through solutions step-by-step, mimicking human cognitive processes. OpenAI’s o1 is one example. Do developers still rely on human knowledge? Despite AI’s constantly expanding capabilities, our survey revealed that human knowledge still ranks supreme when it comes to complicated technical problems. More than 80% of developers still visit Stack Overflow regularly, while 75% turn to another person when they don’t trust AI-generated answers. Even more telling: Despite developers tinkering with reasoning models, advanced questions on Stack Overflow. com have doubled since 2023. Stack Overflow’s parent company, Prosus, uses an LLM to categorize questions as “basic” or “advanced.” The dramatic increase in questions tagged “advanced” suggests that developers are encountering problems AI tools can’t help them with. What does human validation mean for enterprises? Rotnov emphasizes two important conclusions that enterprises should draw from this data: LLMs haven’t mastered complex reasoning problems. Instead, developers turn to human-centered knowledge communities for help. AI is creating new problems that communities have never encountered before. Not only are human expertise and validation still essential, then, but the new problems cropping up because of AI use, misuse, or overuse require human-driven solutions. Example: A developer using an AI coding assistant might generate a working application quickly, but when they need to optimize performance, handle edge cases, or integrate with legacy systems, they require human expertise and collaborative problem-solving. What are the key action items for leaders driving enterprise AI projects? Rotnov outlined two high-level action items business leaders can take to make their AI projects successful while supporting technical teams’ preferred tools and workflows: investing in spaces for knowledge curation/validation and doubling down on retrieval augmented generation (RAG). Invest in spaces for knowledge curation and validation What to do: Create internal platforms where developers can document, discuss, and validate new problems and solutions emerging from AI-assisted workflows. Why it matters: As AI changes how developers work, they need structured spaces to build consensus around new patterns and best practices. Best practices: Choose platforms that support structured formats with metadata (tags, categories, labels). Implement quality signals like voting, accepted answers, and expert verification. Ensure the format is AI-friendly so this knowledge can feed back into your internal LLMs and agents. Key term: Metadata refers to information about data (like tags, categories, or timestamps) that helps organize and contextualize content, making it easier for both humans and AI systems to understand and retrieve relevant information. Double down on RAG systems RAG (retrieval augmented generation) is still “having a moment,” Rotnov says, and for good reason. The survey showed: 36% of professional developers are learning RAG. Searching for answers is where AI adoption is highest in development workflows. The “RAG” tag has become one of the most popular new tags on Stack Overflow. What RAG does: RAG systems summarize internal knowledge sources into concise, relevant answers that surface wherever developers work, whether that’s within IDEs, chat platforms, or documentation. Critical consideration: RAG is only as good as the underlying data. If you’re summarizing poorly structured or outdated information, you’ll get poor results. Example: A developer troubleshooting a deployment issue could query an internal RAG system that pulls from documentation, past incident reports, and team wikis to provide a comprehensive answer without manually searching multiple sources. Future-proofing AI models through reasoning and human validation For organizations building their own models (whether internal tools or products), Rotnov emphasizes two priorities: improving reasoning capabilities and implementing human validation loops. Improve reasoning capabilities The challenge: Current reasoning models are immature and struggle with complex tasks. The solution: Train models on data that demonstrates human thought processes, not just final answers. Important data types include: Comment threads showing how humans discuss and evaluate solutions. Curated knowledge that reveals how understanding evolves over time. Decision-making processes that expose the “why” behind conclusions. Survey insight: For the first time, Stack Overflow asked how people use the platform. The #1 answer? They look at comments. This reveals that developers are looking for more than just the accepted solution. They want to see the discussion, the relevant context, and the diverse perspectives surrounding a question. Implement human validation loops The issue: Model drift: when AI outputs become less accurate as real-world conditions change. The fix: Build continuous feedback mechanisms where humans evaluate and correct AI outputs to ensure accuracy and alignment with human values. Example: Stack Overflow is piloting integrations where AI models appear on leaderboards and users can vote on responses from different models, providing real-time feedback on performance. The developer tool sprawl problem Here’s a surprising finding: Over a third of developers use 6-10 different tools in the course of their work, but contrary to popular assumptions, tool sprawl doesn’t correlate with job dissatisfaction. “It surprised me because everyone’s been trying to solve this tool sprawl problem for years,” Rotnov notes. “But it seems like developers accept that each tool serves a specific use case and that they need them to do their job.” In deciding which AI tools and technologies to invest in, enterprises should bear in mind that developers can tolerate a fair amount of tool sprawl-as long as each one is serving a distinct function within their workflows. Agentic AI: The promised solution? And speaking of workflows. Agentic AI refers to autonomous systems that can perform complex tasks across multiple tools and platforms to achieve specific goals without constant human guidance. In theory, agentic AI promises to solve tool sprawl. But adoption of agentic AI systems is limited: 52% of developers either don’t use agents or stick to simpler AI tools. Security and privacy concerns remain significant barriers to agent adoption. Reasoning model immaturity limits agents’ capabilities. However, among developers who have started using agentic AI in their workflows, the results are promising: 70% report that agents reduced the time they spent on specific tasks. 69% agree that agents increased their productivity. Younger/less experienced developers are more likely to adopt agents. As we’ve seen with the adoption curve of AI tools generally, developers will embrace agentic workflows when they see proof positive that those systems work. Recommendations for navigating agentic AI On that note, Rotnov had some recommendations for enterprises rolling out agentic AI systems. Start small and iterate As with any new tool or technology, Rotnov recommends that enterprises pilot low-risk agentic use cases before rolling out broader implementations. Demonstrate value, build consensus, and then roll it out to more users once you understand how things work on a micro scale. Consider piloting with interns or newer developers on onboarding tasks, where mistakes have lower consequences and feedback loops are clear. Embrace MCP servers MCP (model context protocol) is a standardized way for LLMs to access and learn from data sources. It’s analogous to the International Image Interoperability Framework (IIIF), which standardizes how images are delivered and described over the web. What MCP servers do: Help AI learn implicit knowledge: your organization’s language, culture, and way of working. Enable faster familiarization with internal systems. Provide read-write access and pre-built prompts for dynamic knowledge sharing. Connect to existing AI tools and agents for less context switching. Real-world application: Stack Overflow recently released a bi-directional MCP server. A developer building an internal app in Cursor can connect to the MCP server and immediately access enterprise knowledge-complete with structure, quality signals (votes, accepted answers), and metadata (tags) to inform their application’s outputs. Consider small language models Why the trend: Small language models (SLMs) are gaining popularity because they’re: Task-specific: Smaller models can be fine-tuned for particular domains or use cases. Cost-effective: As you’d expect, small models are cheaper to build and maintain than large models. Better for the environment: Unsurprisingly, SLMs require less computational power. Ideal for agents: Smaller models are ideal for specialized agentic tasks. Example: A healthcare company might deploy an SLM specifically trained on medical coding standards and their internal protocols for processing insurance claims, rather than relying on a general-purpose LLM. Don’t sleep on APIs While MCP servers and agents get attention, APIs remain crucial for reducing context switching and the overall cognitive load on developers. In fact, developers are more likely to endorse and become fans of a technology if it has an easy-to-use and robust API. What to evaluate: Is the API well-documented and supported? Does it use a REST architecture or other AI-friendly format? Is pricing transparent? Is there an SDK available for easier integration? Example: Stack Overflow recently launched a TypeScript SDK for Stack Internal, making it easier for developers to build integrations and custom workflows. Data quality is the key to enterprise AI success Rotnov was very clear about the number-one recommendation she has for enterprises contemplating AI projects: “You really need to be looking long and hard about what internal data sources you have that LLMs and AI can learn from and provide accurate answers to your teams.” Key questions to ask: Are you giving developers spaces to create new knowledge and problem-solve collaboratively? Is that knowledge well-structured with good metadata and quality signals? If you’re using third-party data, does it meet the same quality criteria? Is your data conducive to AI, i. e., organized in ways that LLMs can effectively learn from? No matter what you’re building-agentic systems, RAG implementations, or custom models-the underlying data quality determines success. Even synthetic data generation requires high-quality source material. Final thoughts For their AI initiatives to succeed, enterprises must balance the productive potential of AI tools against the need for continuous human validation and community-driven knowledge infrastructure. Thriving developers aren’t using AI to replace human judgement or stand in for human experience. They’re using them as force magnifiers. In the same way, thriving enterprises are combining AI capabilities with human expertise, leveraging well-structured knowledge systems and thoughtful implementation strategies to make sure AI adds value at every level of the business.
https://stackoverflow.blog/2025/11/25/essential-ingredients-for-enterprise-ai-success/
