It was a seductive promise, right? AI tools would become the universal answer engine for software development (and a lot else besides). Even with zero coding knowledge, you could prompt your way to a solution. Within a few years, the thinking went, developers would scarcely need to talk to another human being to do their jobs.
The data tell a different story.
Despite the proliferation of AI coding assistants, reasoning models, and LLM-powered documentation tools, more than 80% of developers still visit Stack Overflow on a regular basis. And when developers don't trust an AI-generated answer—which happens more often than software vendors would like to admit—75% of them turn to another human for clarity.
Don’t get us wrong: The story here isn’t that AI has failed to deliver on its promise for enterprise software. The story is that developers need more than AI to solve the hard problems they encounter every day. Enterprise SaaS buyers should pay close attention to developers’ concerns around trustworthiness before assuming that AI features will carry the day.
In this post, we’ll explain why developers continue to rely on human expertise to solve the hardest problems, how comments can teach developers more than the accepted answers alone, and how enterprise organizations should be thinking about AI-powered software in light of these insights.
Humans are asking more advanced questions
Stack Overflow's parent company, Prosus, uses an LLM internally to categorize questions on the platform as either “basic” or “advanced.” That’s how we learned that the number of advanced technical questions on Stack Overflow has doubled since 2023.
In other words, over the same years in which AI coding assistants have become dramatically more capable, the volume of hard questions developers are bringing to a human community has doubled.
How should we interpret this? AI tools are handling the easier, more straightforward stuff. Boilerplate generation, syntax lookups, standard library usage, common patterns—all of this is increasingly offloaded to AI, and largely successfully. But the residual questions, the ones developers can't resolve on their own even with AI assistance, are harder than ever. Developers arrive at Stack Overflow when AI tools can’t deliver reliable answers.
This has significant implications for enterprise SaaS buyers. If the question you’re asking to assess an AI tool is, “Can it answer developers’ coding questions?” you’re looking at the easiest part of a problem to solve. Every AI tool worth a sliver of market share can do that. The more important and relevant thing to ask is: Can it answer developers’ hard questions—the ones they still look to other humans to solve?
Answers alone aren’t the answer
When we asked our community why they use our platform, the top answer was something of a surprise: Developers come to Stack Overflow to read the comments. Sure, they’re interested in the accepted answer, but that’s not all they’re after.
This behavior is worth dwelling on because it reveals something fundamental about how developers—and knowledge workers more broadly—evaluate technical information. The accepted answer tells you what works. The comments tell you why it works, when it might not work, what the edge cases are, whether the solution is relevant for your particular use case, and how other people have modified it for their own contexts.
Developers aren't looking for answers alone. They're in search of knowledge—and answers aren't knowledge. Developers understand that to understand something at a deep level, they need to immerse themselves in the discourse around it: the sometimes-contentious, always-contextual conversation that emerges when various practitioners try tackling the same problem from different angles.
This is what AI tools can’t replicate. A language model can synthesize patterns from existing text, but it can’t engage in meaningful debate, acknowledge and cope with uncertainty, or surface the most revealing conversations. Think of a Stack Overflow thread with a dozen comments debating the pros, cons, and best practices of a particular technical approach. The knowledge in that thread isn’t restricted to the approved answer; the conversation is the knowledge. Flattening that back-and-forth into a confidently vapid paragraph captures only a fraction of its value.
The validation gap
Enterprise software buyers are right to be optimistic about AI's productivity benefits. Code generation is faster. Documentation search is more natural. Onboarding new developers to unfamiliar codebases is less painful. All these gains are real, but there are still gaps AI needs to close. One of these we’ve been going deep on is the trust gap, a top-of-mind concern for enterprise SaaS. Another is the validation gap.
When a developer isn't sure whether to trust an answer, they need recourse to human judgment. The 75% figure—the share of developers who turn to another person when they don't trust AI output—represents the size of that gap in practical terms.
The validation gap has real costs for the enterprise, as we’ve written. A developer who can't validate an AI-generated solution might waste time second-guessing it, abandon the approach entirely, or deploy something unproven and untrustworthy. As an enterprise SaaS buyer, those aren’t the outcomes you’re looking for.
This is why the most valuable AI-adjacent tools in the enterprise stack are those that do more than generate answers. They help developers determine which answers to trust. A knowledge intelligence layer that connects internal expertise with open questions, surfaces relevant community discussion, and makes institutional knowledge searchable makes AI tools more useful and more valuable by giving users the all-important context they need to confidently evaluate AI output.
Evaluating AI-enabled SaaS: What to look for
When you’re assessing AI features on an enterprise software platform, a few questions are worth asking:
- Does the tool acknowledge uncertainty? Confidently delivered wrong answers are much worse than acknowledged uncertainty. Tools that surface confidence levels, flag edge cases, or indicate when a question falls outside their reliable knowledge base are more trustworthy in practice than those optimized for fluency.
- Where does it route hard questions? For complex problems, the right answer is often “I'm not sure—here's where you should look.” A tool that has a credible answer for the 20% of hard questions, or one that connects users to human expertise for those questions, is more valuable than one that provides fast, confident, and low-quality answers to everything.
- Does it preserve context and discourse? Raw answers are less valuable than answers with context. Platforms that surface discussion, tradeoffs, and dissenting perspectives enable better decision-making than those that collapse knowledge into a single authoritative output.
- How does it integrate with human expertise? The goal is not to supersede expert communities but to make the invaluable knowledge they contain more accessible to more people. Tools that bridge AI capabilities with structured human knowledge, whether in the form of internal institutional expertise or external developer communities, will outperform those treating AI as a standalone oracle.
The bottom line
The doubling of advanced questions on Stack Overflow since 2023 is a sign that while AI has succeeded at solving the easy problems, the remaining problems are genuinely hard.
AI tools are game-changers in many ways. But for the questions that really get your developers stuck in the mud, human expertise (and the platforms that enable it) is how they get unstuck. In a SaaS market saturated with AI features, human knowledge remains the gold standard. That’s why the wisest approach to your enterprise stack isn’t choosing between AI features and stress-tested human experience. It’s choosing platforms that will let the two work together.