# We’ve Been Building Observability for Humans. AI Just Changed the Customer. > Source: > Published: 2026-08-10 19:01:15+00:00 Earlier this year, one of the [world’s largest retailers](https://searchengineland.com/walmart-chatgpt-checkout-converted-worse-472071?utm_source=chatgpt.com) quietly pulled the plug on its ChatGPT-powered online checkout assistant after the system converted customers at roughly one-third the rate of purchases completed through the retailer’s own website. The headlines focused on the AI and how the model wasn’t good enough. The experience wasn’t polished enough. The assistant needed more training. Yeah, maybe. But I suspect those conversations overlooked something much more fundamental. When AI underperforms in production, our instinct is to question the intelligence of the model. But, increasingly, we should be asking a different question: Did we give it enough of the business to reason about? That distinction matters because enterprise AI isn’t running into intelligence limits nearly as often as it’s running into information limits. **AI Doesn’t Need More Data. It Needs Better Access.** Every enterprise has data. Most have more than they know what to do with. The challenge is that information rarely lives in one place. Customer record s, identity data, infrastructure telemetry, and security events are all scattered across different systems and tools. Humans have become surprisingly good at navigating that fragmentation. When something breaks, we know where to look next. We remember which dashboard tends to lag behind, which team owns a particular service, and which log source usually fills in the missing pieces. Experience allows us to stitch together a coherent picture from incomplete information. But AI doesn’t investigate that way. It reasons from whatever information it can retrieve. If that information is fragmented, stale, duplicated, or disconnected, the model has no intuition to compensate. It simply reaches conclusions using the picture we’ve given it. That’s why so many AI failures turn out to be access failures in disguise. **The Missing Layer** Most conversations about enterprise AI eventually arrive at the same recommendation: connect your models to more systems. That sounds straightforward until you consider what “more systems” actually means. Every application has its own interface. Every repository has different permissions. Every API behaves differently. Every dataset has its own structure, owner, and governance model. Adding another connection doesn’t necessarily make AI smarter. Often it simply introduces another place where context can become incomplete or inconsistent. The real challenge is less about connecting AI to everything and more about creating a reliable way for AI to discover the information it actually needs, regardless of where that information lives. Think about the front door of a well-designed building. Visitors don’t need to know where the electrical room is, or how the plumbing is routed through the walls. They simply walk through one entrance and find what they’re looking for. Enterprise AI needs something similar. Not unrestricted access to every system, but a consistent way to discover trusted information without requiring every application to reinvent how that information is exposed. **We’ve Optimized Observability for Eyes** For years, observability has been designed around the way people investigate systems. We built dashboards because humans recognize patterns visually. We created search interfaces because investigations are iterative. We invested in rich visualizations because people absorb information more quickly when they can see relationships unfold. Those investments weren’t wrong. They reflected the needs of the customer. Yet the customer was human. Today, observability has another customer. AI doesn’t open dashboards. It doesn’t wait for someone to click through a series of graphs before deciding what deserves attention. It consumes logs, metrics, traces, events, and identities programmatically. If those signals disagree with one another or arrive without enough business context, the model doesn’t know which source is authoritative. It reasons from whatever it was able to retrieve. That changes the standard. Historically, observability succeeded when it helped an experienced engineer understand a complex system. Increasingly, it also needs to support software making decisions at machine speed. Those are different design problems. **Context is Becoming Infrastructure** One phrase comes up constantly in AI conversations: *AI needs more context.* It does. But what’s often overlooked is where that context comes from. Context isn’t created by giving a model a larger context window or another retrieval plugin. It’s assembled from everything an organization already knows about its systems: telemetry, identities, infrastructure, documentation, APIs, business processes, and operational history. The quality of an AI system increasingly depends on how well those pieces fit together. Organizations that recognize this early start to think differently about observability. They’re asking whether their telemetry is consistent enough to support automated reasoning, whether business context can be discovered instead of hardcoded, and whether information can be governed once and consumed everywhere. Those questions have less to do with AI models than they do with architecture. **A Different Way to Think About Observability** AI hasn’t changed the purpose of observability. It has expanded its audience. People will continue investigating incidents, validating automated actions, and making judgment calls that software shouldn’t. But more investigations will begin with AI. More decisions will be informed by software. More operational context will be assembled automatically before a human ever opens a dashboard. The organizations that adapt won’t necessarily have the most sophisticated models. But they’ll have the clearest picture of their own environments. Because the quality of AI isn’t determined solely by the model making the decision. It’s also shaped by everything the model was able to see before it made one. *Next week, I’ll explore what happens when organizations move beyond a single AI assistant and begin deploying fleets of specialized AI agents—and why that shift raises an entirely new set of challenges for observability and operations.*