Why the next era of enterprise AI needs context engineering [Q&A]
The adoption rate of artificial intelligence in the enterprise shows no signs of slowing down, but while organizations have been focused on which of the many models to use, the playing field is beginning to level out.
We talked to Saket Saurabh, CEO and co-founder of Nexla, to discuss ‘context engineering’ and why he believes it will be key to gaining competitive advantage.
BN: What is the context layer and why is it important?
SS: If prompt engineering was about asking better questions, context engineering is about giving AI the foundation to answer them well. The context layer is everything that shapes how AI systems reason and act -- proprietary data, documents, workflows, domain knowledge, policies, and even memory of previous interactions.
Why it matters: as models converge in quality, the real competitive advantage won't come from choosing the right model. It comes from building the right context. Think about onboarding a new employee at a complex organization. It takes months for them to understand the systems, policies, and nuances needed to make good decisions. AI faces the same learning curve. The better the context you provide, the better AI performs, whether that's summarizing reports, flagging anomalies, or optimizing workflows.
BN: How can this be combined with agentic AI systems?
SS: Agentic AI systems can reason, plan, and execute tasks -- but they require more than clever prompts. They need a deeper understanding of where they are, what they know, and the constraints that apply. That's what context engineering provides.
The shift is from instructing AI to equipping it. Instead of programming every rule, you define boundaries, supply context, and let AI reason within those boundaries. This is contextual augmentation, not full automation. You're not giving AI control of critical systems like the general ledger. You're enabling it to handle the connective tissue -- the work that humans would do with the same organizational context.
BN: What are the major challenges of managing context within large, dynamic systems?
SS: The scale problem isn't about computing power anymore -- that's become abundant. It's about orchestration. When you stitch together dozens of data products across distributed systems and feed them into AI models, reliability and observability become paramount.
Context comes from everywhere now: chat transcripts, support tickets, sensor feeds, video, contracts, PDFs. This explosion of data variety has outpaced traditional data engineering methods. AI doesn't consume static tables, it needs live, dynamic information flows. You have to design systems that can retrieve, transform, and deliver the right context to the right process, all in real time.
The other challenge is governance. For a financial institution feeding underwriting models with historical claims data, you need to mask sensitive information, maintain lineage, and ensure compliance. For healthcare providers integrating on-premises patient data with cloud-based AI, you must ensure privacy while delivering relevant context. These aren't just technical problems -- they're governance problems that require automation.
BN: In multi-agent systems, how can context be shared and maintained across different agents to prevent conflicting assumptions and duplicate work?
SS: You need orchestration that understands dependencies, monitors flows, and ensures the context being fed to AI is always complete, consistent, and compliant. When agents have isolated views of the world, you get exactly what you described -- duplication and conflicting assumptions.
The solution is designing a shared context environment. Agents access the same knowledge bases, respect the same business rules, and recognize the same permissions. They coordinate rather than collide. This requires moving beyond static workflows into adaptive systems that can flex as data and requirements change.
This is really about building composable systems -- where you can package data into well-documented, secure data products with embedded metadata about how they can be used. When agents pull from these products, they're all working from the same source of truth.
BN: How will the increasing size of context windows and the development of new model architectures impact the practice of context engineering in the future?
SS: Larger context windows change the game, but not because you can stuff more data into a prompt. They change how we architect enterprise AI systems.
Right now, we spend enormous energy on retrieval -- finding the perfect subset of documents or data to include. With 200,000-token windows instead of 8,000, we can shift from retrieval engineering to true context engineering. You can bring in broader knowledge bases, richer memory, more complete business rules, and deeper user context all at once.
New architectures that process context more efficiently matter just as much. Models that can reason over structured data, understand temporal relationships, and maintain consistency across longer interactions unlock enterprise use cases where AI operates over days or weeks, not just single conversations.
But here's what people miss: bigger context windows mean bigger governance challenges. If you're feeding more data into agents, you need stronger lineage tracking, better versioning, and more sophisticated orchestration. The discipline of context engineering becomes more important, not less.
The future isn't about giving AI infinite memory. It's about giving it the right memory, structured in ways that scale with enterprise complexity. Organizations that master this -- making context dynamic, scalable, and durable over time -- will have agents that are more aware, more reliable, and more aligned with how their business actually works. That's where the competitive advantage lives.
Image credit: Tongsupatman/Dreamstime.com
