Understanding Observability in AI Systems

Imagine you’re driving a car with no dashboard — no speedometer, no fuel gauge, not even a warning light. In this scenario, you’re blind to essential information that indicates the car’s performance and health. You wouldn’t know if you’re speeding, running out of fuel, or if your engine is overheating until it’s potentially too late to address the issue without significant inconvenience or danger. Now think about your infrastructure and applications, particularly when you’re dealing with microservices architecture. That's when monitoring comes into play. Monitoring serves as the dashboard for your applications. It helps you keep track of various metrics such as response times, error rates, and system uptime across your microservices. This information is crucial for detecting problems early and ensuring a smooth operation. Monitoring tools can alert you when a service goes down or when performance degrades, much like a warning light or gauge on your car dashboard. Now observability comes into play. Observability allows you to understand why things are happening. If monitoring alerts you to an issue, like a warning light on your dashboard, observability tools help you diagnose the problem. They provide deep insights into your systems through logs (detailed records of events), metrics (quantitative data on the performance), and traces (the path that requests take through your microservices). Just as you wouldn’t drive a car without a dashboard, you shouldn’t deploy and manage applications without monitoring and observability tools. They are essential for ensuring your applications run smoothly, efficiently, and without unexpected downtime. By keeping a close eye on the performance of your microservices, and understanding the root causes of any issues that arise, you can maintain the health and reliability of your services — keeping your “car” on the road and your users happy.

We don’t trust what we don’t understand. This isn’t just about whether AI can make good decisions. It’s about whether we can see how it makes them. In business, we don’t approve budgets without context. We don’t hire talent without interviews. We don’t trust people who can’t explain their choices. So why are we so quick to deploy AI we can’t interrogate? Here’s the truth: → You can’t govern what you can’t observe. → You can’t align what you don’t understand. → And you definitely can’t scale uncertainty. The future doesn’t belong to those who just ship faster. It belongs to those who build transparency into the core of their systems. This week, that future just got a little closer. OpenAI launched a new generation of ChatGPT agents designed for exactly this: → A replay feature that lets you inspect every step an agent takes → Real-time approvals before any action is executed You’re not watching a black box anymore. You’re witnessing the reasoning behind the result. With embedded tracing tools in the new Responses API and Agents SDK, businesses can now answer a question that’s eluded us for years: Why did the AI do that? It’s not a perfect system. But it’s a meaningful shift. From opaque automation → to observable intelligence. From guessing → to governing. From hype → to "almost" trust. What’s one area of your workflow where visible reasoning from AI would change the game for you? --------------- Follow me for more on the AI for SMEs Journey.

Trustworthy AI in production demands a fundamentally different approach to classical software. Unlike deterministic systems, AI applications – especially those built on LLMs and RAG – face constantly shifting data inputs, probabilistic outputs, and complex pipelines that span data, systems, code, and models. My colleague Shane Murray recently spoke on this same topic at the University of Arizona for IEEE International Congress on Intelligent and Service-Oriented Systems Engineering (CISOSE) alongside Vrushali C. (Dir Eng, Data & AI at Okta) Sharoon Srivastava (Principal Product Manager AI at Microsoft) Stephanie Kirmer (Senior MLE at DataGrail) Anusha Dwivedula (Director of PM at Morningstar) Vibe-coding a new AI tool might seem easy enough, but making it reliable is anything but. As Shane states in his position, to ensure reliability and trust, organizations must continuously observe every layer of their data + AI stack—not only in a secure testing environment, but live in production—by combining automated, scalable monitoring with human-in-the-loop oversight and a repeatable operational practice to rapidly root-cause and resolve issues. Only by pairing these approaches can we detect failures, mitigate risks, and sustain trust as AI systems evolve in the real world. You can see the full abstract from the session in the doc below. And if you want more from Shane, you can read his full thoughts in his latest article - or check out his feature in this week’s Alt Data Weekly (shout-out to John Farrall). Reliability isn’t a new challenge. But in the milieu of AI-everything, we need to define a different approach. The wheels are turning. Are you on board? Resources: https://lnkd.in/gZ_Nta3H https://lnkd.in/g8g2U3qs

Menlo Ventures released a new report detailing the modern AI stack. Here's what I found most interesting: 🏗️ Data pre-processing You need specialized solutions to extract and prepare the data used by these modern AI systems. Whether it comes from a dedicated provider like unstructured.io, or if it's just a set of regular expressions, consider taking another look at how you feed data to your models. 💡 Context fragmentation When building RAG systems, answers often require context from multiple sources or multiple locations in the same source. Hrishi Olickel developed what he calls "walking RAG", and this is a great example of how to navigate across multiple contexts. 📊 Observability Implement solutions that assist in the development and monitoring of your AI applications. Anecdotal feedback regarding a system's performance just doesn't scale. Simple logging is a good start, but consider investing in a tool like Braintrust Data or Patronus AI for a more robust solution. Thanks to the authors - Matt Murphy, Tim Tully, Grace Ge, Derek Xiao, and Katie Keller - for writing such an insightful and informative piece. The full report is linked in the comments.