AI observability: How CIOs can see past their org blind spots - InformationWeek

No corner of the modern enterprise remains untouched by artificial intelligence. But as use cases expand and adoption spikes, cracks appear in the technology's deployment. Increasingly, CIOs struggle to keep track of what AI systems are doing, who uses them, and how they perform.

In many cases, CIOs are discovering they have no way to monitor or measure critical factors such as model drift, latency, hallucination rates, performance degradation, shadow AI and output decay. Not surprisingly, as AI systems make increasingly consequential decisions — and handle critical activities — the risks escalate.

"CIOs feel confident that they know how AI is being deployed within their organization, but they typically can't tell you how it's actually performing," said Arnab Chakraborty, chief responsible AI officer at Accenture.

According to the Stanford HAI 2026 AI Index (using McKinsey data), organizations that rated their AI incident response as "excellent" dropped from 28% in 2024 to 18% in 2025. Meanwhile, 88% of organizations report using AI in at least one business function, but fewer than 10% have fully scaled AI in any single domain.

The takeaway? As enterprises navigate a rapidly changing AI space, observability is critical. Yet AI requires a fundamentally different way of thinking than conventional IT. "In order to understand day-to-day performance and manage risk, it's critical to think beyond traditional IT measures," Chakraborty said.

Visibility into AI performance matters

What sets AI oversight apart from conventional IT monitoring is unpredictability. Uptime, throughput, utilization rates and errors — metrics that anchor IT — do not capture the factors and risks germane to AI. That's because AI is probabilistic by design. The same input can produce drastically different outputs.

These issues can take many shapes and forms. CIOs generally know the intended purpose of AI systems but lack insight into accuracy, latency, user interfaces, costs and risks. There are also model drift, agent behavior and shadow AI issues to grapple with. Unfortunately, no vendor has created a tool that delivers observability across all the AI layers.

The problem is rooted in the way AI works. It isn't a single model with a single output. AI is typically a stack of components: data pipelines, foundation models, retrieval systems, agents and other components — all interacting with humans and workflows. Agentic AI introduces additional risks. These include: "Cascading errors, integration failures, unclear accountability and difficult-to-anticipate emergent behavior when multiple agents interact across workflows," said Ilana Golbin Blumenfeld, responsible AI partner at PwC US.

Consider: A miscalibrated retrieval policy can corrupt outputs across a dozen downstream applications. Drift in a vector database can pop up as hallucinations in a chatbot. As enterprises chain agents together to handle longer-running tasks, the number of things that can go wrong expands faster than the tools designed to watch the environment. "It isn't just a linear effect, it's a compounding effect," Chakraborty points out.

Often, these problems go unnoticed for weeks or months — until something suddenly breaks. That's because the level of performance degradation isn't noticeable — until it is. "If you don't intervene early enough, within days you can suddenly find yourself in an undesirable place," said Grace Trinidad, research director of AI security and trust at IDC.

Existing dashboards and security tools cannot solve the problem, Trinidad said. Most rely on risk scores and confidence ratings that are insufficient and entirely opaque for AI. In fact, two organizations can run identical models and arrive at very different views of the same risk factor. "There's no standardization of what goes into a risk score," she said.

How AI monitoring is evolving

You can't govern what you can't see. Microsoft found that 73% of organizations have detected unauthorized AI tools in their networks, yet only 28% have comprehensive monitoring or blocking capabilities in place. McKinsey's "2026AI Trust Maturity Survey" found that the average maturity score for organizations is 2.3 out of 4, with only about one-third reaching maturity level 3 or higher in strategy, governance and agentic AI oversight.

"One of the biggest blind spots for organizations is that they still monitor AI like traditional software. They can see that AI infrastructure is running, but they don't understand why it is producing poor or unreliable results," Blumenfeld said. Often, organizations design front-loaded intake and risk assessment processes that do not address how an AI system is actually used and how risk within an application can drift. "The key is choosing tools that can integrate across multi-cloud, multimodel and agentic AI environments," he said.

In fact, AI observability is rapidly evolving to full-stack visibility along with more nuanced insight into AI behavior. In this world, telemetry data takes a back seat to things like semantic mapping and intent interpretation, continuous monitoring and audits, role-appropriate views and controls, and tooling that oversees security and regulatory requirements in a more comprehensive way. Blumenfeld said that these tools must span governance, infrastructure monitoring and model-level visibility.

A robust discovery process is foundational, Trinidad said. It's important to catalog models, agents, owners, versions, deployment contexts and logs — preferably in an AI registry. With a clear idea of what systems are supposed to do and an understanding of what needs to change, an enterprise can begin to build observability into the entire stack. With this information, CIOs can spot data and model drift, performance degradation, hallucinations, shadow AI and security risks before they cause problems or reputational damage.

Layered monitoring also requires automated guardrails, Chakraborty said. This means establishing the right thresholds for key factors, including hallucination rates, latency, bias, privacy, costs, data and model drift, regulatory compliance, and the quality of output. It also requires the right mix of tools from hyperscalers and third-party vendors to manage and measure tasks.

With an integrated control plane — a single architectural layer that collects and displays all the signals — managers and leaders from different departments can see what really matters for them. For instance, a chief risk officer sees risk thresholds and breaches, a CFO views consumption and runaway cloud costs, a chief human resources officer sees workforce impact, and engineers have their fingers on the pulse of auditability and explainability. "It creates your DNA, almost like a nervous system for your AI," Chakraborty said.

Where AI observability is headed

"CIOs should treat AI observability as a core design principle rather than something added after deployment," Blumenfeld said. It's also essential to treat observability as a cross-functional effort involving IT, business, risk compliance and internal audit teams, he said. "The industry is moving beyond monitoring individual AI models and toward monitoring entire ecosystems of agents, orchestration layers, data pipelines, and autonomous workflows."

When organizations get the equation right, they can scale AI faster and more safely, control costs even as workloads grow, generate an airtight audit trail and boost customer trust. Gartner forecasts that large language model observability investment will cover 50% of GenAI deployments by 2028, up from 15% today.

To be sure, observability isn't a bolt-on item, and it doesn't follow an IT-as-usual formula. It's a fundamental element that has to be built into an AI framework. "Organizations that get this right from the get-go and invest in building the muscle around it are the ones who will emerge as leaders in the age of AI," Chakraborty said.