Everyone’s talking about AI adoption, but behind closed doors, I’m hearing a different story. There’s a bottleneck, and it’s not what you’d expect.
It’s not that AI models lack capability. They don’t. The problem is that AI systems exist in isolation from the actual work we do every day. They’re disconnected from the context that makes decisions meaningful.
Michael Foree, Stack’s Director of Data Science, put it perfectly when he described the email problem. Sure, modern LLMs can read and respond to emails. But they have no idea who sent it, what conversation preceded it, what your relationship with the sender is, or why this email matters right now. To get a decent response, you’re copy-pasting context from Slack, previous emails, meeting notes, and internal documents. Then you’re editing the AI’s output. Then you’re manually sending it. Then you’re probably editing it again.
That’s not AI assistance. That’s a Frankenstein workflow.
The Hidden Cost of Context Fragmentation
The real issue is architectural. Your organization’s context is scattered everywhere. Critical information lives in email, Slack, Jira, Google Drive, wiki pages, and tribal knowledge locked in people’s heads. An LLM has no natural way to access any of this simultaneously. Even when you give it permission to access these tools, you’re asking it to navigate a minefield of irrelevant distractions and missing information.
This is where context engineering becomes critical. It’s not just prompt engineering or fine-tuning. It’s the entire discipline of structuring, validating, and surfacing the right information to AI systems so they can actually reason about your specific situation.
The token cost alone makes naive approaches unworkable. If you dump every email from a person into the context window, plus their Slack history, plus relevant documents, plus meeting transcripts, you’re looking at astronomical context lengths. The AI gets overwhelmed. It hallucinates. It fixates on irrelevant details. It fails to ask clarifying questions because no one trained it to do that.
And that’s just for one email.
Why Enterprise Data is the Real Challenge
Publicly available information is easy by comparison. LLMs trained on internet data have seen countless examples of how-to guides, best practices, and general knowledge. They’ve been trained to ask clarifying questions about ambiguous scenarios like “how do I jump over a log?”
But private proprietary data? That’s where everything breaks. Your company’s specific processes, internal standards, past decisions, and domain expertise can’t be trained into a general-purpose model. The AI labs can’t access your secret sauce, and you wouldn’t want them to. So the AI shows up to your enterprise clueless about how things actually work.
This is why Stack is building tools like Stack Internal to solve the context engineering problem differently. Rather than training LLMs on proprietary data, we’re creating knowledge connectors. The AI can access your internal information, but humans validate the results. An expert in your process can confirm whether the AI’s answer is correct for your specific situation. It’s context engineering with human validation built in.
It’s the holy grail that can’t be bought because it requires understanding of your actual business.
The Setup Burden Nobody Wants
But here’s the thing that’s actually killing adoption: the setup tax is massive. To get any of this working, you need to:
- Connect your email tool
- Grant access to your document repositories
- Integrate your communication platforms
- Set specific permission boundaries
- Define workflows and approval steps
Then you have to maintain it. Update permissions. Add new data sources. Retrain the AI on new procedures.
For a feature you might use once a day? Most people rightfully ask themselves: is it worth it? And in enterprise settings, the burden multiplies across teams, compliance requirements, and security reviews.
I think about this differently than most people. The question isn’t “will this save me time eventually?” It’s “is the friction low enough that I’ll actually start using it today?”
Right now, the answer is usually no.
What This Means for Developers
If you’re building AI-powered tools, this is your real competitive moat. The companies that will win aren’t just training better models. They’re solving the context engineering problem at scale. They’re making it trivially easy to give AI the context it needs without requiring users to understand token costs, context windows, or integration workflows.
They’re also building the validation layer. Because context without validation is just expensive hallucination. Learn more about how to approach llm-integration strategically.
The bottleneck in AI adoption isn’t capability. It’s plumbing. And plumbing problems have different solutions than capability problems. Understanding this distinction could be the difference between a product that frustrates users with setup overhead and one that feels magical because the context engineering happens invisibly.
The future belongs to the teams that can make their AI systems context-aware without making their users context engineers.