In these times when AI promises "transformation" at every corner, it might sound a bit audacious to claim that my place at the table, as a 20-year system architect, remains the same. But believe me, that's exactly the situation. Our tools are evolving, our skill sets are expanding, yet the fundamental problems we're trying to solve remain surprisingly familiar.
My first shock when integrating AI into production planning within a manufacturing ERP came not from the power of the technology, but from the obstacles posed by data quality and the complexity of the business process. No matter how smart AI is, it's impossible for it to produce accurate output with dirty or incomplete data. It was at this very point that I once again understood how critical the system architect's role is.
AI's Promise, On-the-Ground Realities, and My Experience
AI's potential is undeniable; I see it too. In the financial calculators of my side product, the ability to analyze large datasets and provide personalized recommendations to users is truly impressive. On one hand, shaping the model's output with prompt engineering, and on the other, feeding it the right information source with RAG (retrieval-augmented generation) patterns, have become new methods of "API design."
However, situations I've encountered in a client project showed that AI is still not a "magic box." For example, I set up Gemini Flash, Groq, Cerebras from different providers via OpenRouter using an agent pattern with fallback mechanisms. But the question of who oversees the system when a model hallucinates or produces an unexpected output is still on my desk. In AI-powered operations, we still need traditional observability tools to ensure the reliability of pipelines.
ℹ️ Experience Speaks
Without a correct problem definition and a robust data infrastructure, AI can be just an expensive toy. The system architect's job is to adapt this technology to real-world problems and manage expectations.
The Unchanging Iron Rules of Architecture
No matter how much artificial intelligence advances, the cornerstones of system architecture remain in place. Network segmentation is still the backbone of security in corporate networks. Issues like VLAN tagging chaos or routing flap continue to be complex situations that need human intervention, despite AI's promises of automated network configuration. Last year, finding that VPN tunnels weren't working properly at a client due to MTU/MSS mismatches once again showed how important fundamental network principles still are, not AI.
Topics like data integrity, performance, and reliability also maintain their independence from AI. To understand PostgreSQL WAL bloat, you still need to look at the pg_wal directory and perform connection pool tuning. Redis OOM eviction policy choices are critical decisions that directly affect how the system behaves under load. While AI can optimize these things, they won't work correctly without initial architectural design and human intervention. Linux system administration details like Journald rate limit or cgroup memory.high are also still critically important for a system to run healthily.
The System Architect's Role: Evolution or Expansion?
In my opinion, the system architect's role is experiencing an expansion rather than an evolution. We are no longer solely responsible for the system's availability and performance. Issues like AI models' reliability, explainability, and ethical implications are also on our table. In the AI-powered risk analysis we developed for a bank's internal platform, the final decision was still made by a human. AI merely provided faster and more comprehensive data.
Observability layers must also monitor the predictions and actions generated by AI. Our CI/CD pipelines must now manage not only code but also model version control, testing, and deployment. This means we need to consider blue-green or canary deploy strategies for AI models as well. Approaches like feature flag and dark launch have become even more important for controlled rollout of AI-powered features. These new layers build upon our existing capabilities; they don't replace them.
⚠️ Risk of Misunderstanding
AI is not a magic wand that can instantly solve complex system problems. Ignoring fundamental engineering principles can lead to greater disasters in AI projects.
The Architecture of the Future: Human + AI Partnership
In the content creation for my own site, even though AI prepared the drafts, it was still me who added the final touches and the "Mustafa Erbay" voice. This is a strong example of AI acting as a "middle brain." It collects data, analyzes it, makes predictions, but we are still the ones who make the final decision and take responsibility. When developing an ERP for a manufacturing company, we brought AI-optimized inventory level suggestions. However, the supply chain manager evaluated these suggestions and made the final decision, because there were factors AI didn't see, such as market conditions or unexpected events.
As system architects, we need to understand very well what AI does well and what it cannot do. Automation is great, but critical thinking, the ability to understand complex systems end-to-end, and evaluating trade-offs are still our most valuable assets. In security measures like SQL injection mitigation or DDoS mitigation layers, AI can be a great tool for anomaly detection, but the design and implementation of these measures should still be done by humans.
Being a system architect in the age of AI means adding new and powerful tools to our technological instrument box. However, skillfully using these tools, playing the right symphony, still depends on our interpretation and experience.
So, for you, what has been or will be the most critical skill for a system architect in the age of AI?
Top comments (0)