Engineering for the Enterprise: When AI Meets Real-World Complexity
Sierra Ventures recently hosted an invitation-only discussion with enterprise founders and technology leaders to unpack a timely question: What happens when the promise of AI meets the complexity of real-world enterprise systems?
The conversation focused on the practical challenges of deploying AI at scale—from data infrastructure and security to governance and go-to-market. Leaders shared what’s working, where things are breaking down, and what it will take to build enterprise-ready AI in 2025 and beyond.
The language models are ready. The infrastructure is not.
Most enterprise teams agree that large language models are mature enough for many use cases. The real blockers are tooling, infrastructure, and compliance. Security, observability, and runtime environments are often missing or immature. As one speaker noted, spinning up agents is easy, but running them safely and contextually in production remains challenging.
The conversation is shifting from model selection to application strategy.
In 2023, much of the enterprise discussion focused on choosing the right foundation model. Now, the focus is on solving real business problems. Leaders are asking how to integrate AI into workflows, how to secure it, and how to ensure ROI. The attention is moving from which model to use to how to make AI actually useful.
Rethinking work matters more than replicating it.
There is growing consensus that building AI agents to mimic human roles misses the point. The real opportunity is to redesign processes entirely. Instead of thinking, “What would a person do,” leaders are exploring what only an agent can do. AI is not just automation, it is a new way of delivering outcomes.
Legacy data stacks are holding enterprises back.
Today’s enterprise data stack is optimized for structured data and BI dashboards, not unstructured inputs like documents, chats, or emails. The breakthrough of large language models lies in their ability to process unstructured information. Unlocking their full value will require rethinking where enterprise data lives and how it is accessed.
Agents need context, not just instructions.
Many enterprise agents are underpowered because they lack the data awareness and domain context to be effective. Some companies are layering in semantic models and knowledge graphs to give agents a deeper understanding of enterprise information. The future of multi-agent systems depends on this kind of context-rich architecture.
Enterprise adoption begins with developer enablement.
Security and compliance concerns often slow down AI adoption. But locking down experimentation can backfire. Companies that empower developers to test new tools in a structured way are learning faster and building internal capability. Guardrails are necessary, but so is space to explore.
Use case clarity beats model comparisons.
In today’s enterprise environment, choosing the right use case is more important than choosing the right model. Whether teams use open-source, fine-tuned, or commercial models, success starts with understanding what business problem they are solving and what data the agent needs to solve it.
A new AI-native stack is emerging inside the enterprise.
Forward-looking teams are beginning to imagine a stack where structured systems remain intact, but new unstructured systems, including agents, LLM interfaces, and semantic layers, play a central role. The winners will be those who adapt their infrastructure to match how AI actually works.
Leaders must understand AI directly, not through proxies.
One of the most actionable takeaways: Executives need direct exposure to AI tools, risks, and workflows. Strategic decisions around investment, risk management, and team structure depend on first-hand understanding. You cannot outsource learning.
Execution, not hype, will define the next wave of AI.
While many teams are excited about building with AI, the gap between prototype and production remains large. Leaders emphasized the importance of picking high-impact, narrow use cases that can run end-to-end and move a clear business metric. Success depends less on technical novelty and more on operational execution.
Where enterprise goes next
As one speaker said, “You can write agents in your sleep. The hard part is making sure they actually work and that someone wants to use them.”
Enterprise AI is moving from experimentation to execution. The path forward requires rethinking the stack, designing for real outcomes, and enabling teams to build and learn. Sierra Ventures is excited to partner with the builders who are turning these challenges into opportunities.