Skip to main content

The Real Enterprise AI Risk Isn't Hallucinations

Nils Bunge
Tsuga

Ask most IT leaders about their biggest concern with AI and you'll hear the same answer: hallucinations.

That concern made sense when large language models first entered the enterprise. Early systems often produced confident but incorrect answers, creating understandable skepticism about whether AI could be trusted in business-critical workflows.

Today, however, the conversation has shifted.

Model accuracy continues to improve at a remarkable pace. While hallucinations have not disappeared, they are no longer the primary obstacle preventing organizations from realizing value from AI. In my conversations with enterprise leaders, a different challenge is emerging: understanding what AI systems are actually doing once they are deployed.

As organizations move beyond chatbots and experiments, they are increasingly deploying AI agents that perform multi-step tasks. These systems retrieve documents, query databases, call APIs, generate reports, write code, and make recommendations. The issue is not whether the model can reason. The issue is whether the organization can see, verify, and govern the decisions being made along the way.

That's an operational problem, not a model problem.

A useful way to think about modern AI is that the model is only one component of a much larger system. Around the model sits an operational framework that includes tools, workflows, permissions, retrieval systems, APIs, business rules, and human oversight. This surrounding environment determines how the model interacts with the organization and ultimately whether it delivers reliable outcomes.

Many AI failures today occur not because the model invents information, but because small errors accumulate across extended workflows.

Imagine an AI agent assigned to create a quarterly business report. It retrieves information from several internal sources, summarizes findings, pulls supporting data, drafts conclusions, and formats the final document. If a minor error occurs early in the process—perhaps an outdated document is retrieved or a metric is interpreted incorrectly—that mistake can influence every subsequent step. By the time the report reaches a human reviewer, the underlying issue may be difficult to trace.

I've seen similar patterns emerge across software development, compliance reviews, customer support operations, and data analysis workflows. The challenge is rarely a single catastrophic mistake. It's the gradual drift that occurs when multiple automated decisions build on one another.

Traditional software teams solved this problem years ago through monitoring, logging, testing, and observability. When a distributed application fails, engineers don't simply inspect the final output. They examine the entire chain of events that led to the failure.

Enterprise AI requires the same discipline.

Organizations need visibility into which tools an agent accessed, what information it retrieved, which decisions it made, where confidence levels changed, and where errors began to appear. Without that visibility, AI systems become operational black boxes. Teams may know something went wrong, but they cannot determine why.

This becomes especially important as organizations pursue greater levels of autonomy. Every executive wants AI to increase productivity, but productivity gains depend on trust. And trust comes from transparency.

One of the reasons AI coding assistants have gained traction is that their outputs are highly verifiable. Code can be compiled, tested, scanned, and validated automatically. Feedback arrives quickly, allowing errors to be identified and corrected before they propagate.

The same principle should apply across other enterprise workflows.

Before organizations deploy AI into business-critical processes, they should establish verification checkpoints, define measurable success criteria, and create clear escalation paths when confidence drops or unexpected behavior occurs. Human oversight remains important, but humans should focus on exceptions and anomalies rather than reviewing every action manually.

For IT leaders, this means expanding the conversation beyond model selection. The question is no longer simply, "Which AI model should we use?" The more important question may be, "How will we observe, govern, and validate the decisions that model makes once it becomes part of our operations?"

The organizations seeing the greatest success with AI are not necessarily deploying the largest models or the newest agents. They are building systems that make AI measurable, inspectable, and correctable.

That's why I believe the next major challenge in enterprise AI is not intelligence. It's operational visibility.

Hallucinations may grab headlines, but the real work of scaling AI successfully comes from understanding what happens between the prompt and the outcome.

Nils Bunge is Head of Product at Tsuga

Hot Topics

The Latest

Rapid AI adoption and the unique ways AI workloads operate is redefining the scope and structure of what these teams must deliver. This shift is forcing organizations to rethink how they manage scale, automation, and control, according to The State of SRE and Platform Engineering 2026, a new report from Dynatrace ...

AI is usually talked about as a software tool, but it also depends heavily on the network behind it. Whether a company is using AI for chatbots, automation, monitoring, analytics, or employee support, all of that information has to move across the network in a reliable and secure way. That means AI is not just an application decision. It is also an infrastructure decision. Before organizations rush into AI, they should ask a simple question: Is our network ready to support it? ...

Enterprise AI often lacks governed access to where business processes actually execute. Without that access, AI agents may be able to reason, but they cannot operate reliably across enterprise workflows. For AI agents to effectively carry out workflows, they will require integration-layer context and controls. Organizations can implement these prerequisites by providing AI with managed access to the middleware layer ...

Enterprise networks rarely behave the same way for very long. A routing adjustment in one region may unexpectedly alter application performance in another. A cloud migration may introduce hidden dependencies that go unnoticed until an outage occurs. All the while, the network is managed by several different teams, each of whom use different tool sets — and as a result, have different views of the network ... There’s usually an engineer who remembers why traffic fails over a certain way between sites, or which transparent firewall was added where. The problem is that human memory cannot scale alongside enterprise-scale networks ...

Ask an infrastructure team how confident they are in their ability to govern AI, and most will tell you they've got it handled. A recent survey of 406 IT decision-makers and platform engineering leaders found 86% expressing exactly that confidence. Ask the same group whether they have a formal written AI governance policy, and the number drops to 30%, according to Spacelift's Infrastructure Automation Report ...

In MEAN TIME TO INSIGHT Episode 27, Shamus McGillicuddy, EMA VP of Research, Network Infrastructure and Operations, and Parker Hathcock, EMA Research Director covering IT Service/Operations (ServiceOps), discuss observability unification in modern IT operations ... 

Virtual Private Networks became a cornerstone of enterprise security at a time when corporate infrastructure looked very different from today ... For years, this model worked well. But the architecture behind VPNs assumed a centralized corporate environment—one where the network itself was the hub of activity. In a cloud — first world, that assumption no longer holds ...

Website outages get resolved just as fast in August as they do in November. I went looking for the opposite: the summer slowdown everyone assumes is there once the people who fix things are away. It isn't in the data we collected, covering 1.8 million confirmed outages across tens of thousands of websites ...

This year, many of the cloud infrastructure contracts signed in the early days of the AI boom will come up for renewal. As the year goes on, I anticipate we'll see a significant amount of cloud vendor swapouts and multi-cloud adoption, and the reason isn't just GPU depreciation. It's because they're tired of their current cloud providers ...

There's a moment the many observability teams have experienced days into bringing a new service into production: you realize that the vendor's claims of "intelligent" behavior included a large serving of hype. Their dashboards look nice until they don't, the failure modes are a black box, and no one on the team can confidently explain why the system did what it did at 2 am. Agentic AI is about to force every Ops team to relive that moment at web-scale until they start treating these systems as the dependencies they actually are ...

The Real Enterprise AI Risk Isn't Hallucinations

Nils Bunge
Tsuga

Ask most IT leaders about their biggest concern with AI and you'll hear the same answer: hallucinations.

That concern made sense when large language models first entered the enterprise. Early systems often produced confident but incorrect answers, creating understandable skepticism about whether AI could be trusted in business-critical workflows.

Today, however, the conversation has shifted.

Model accuracy continues to improve at a remarkable pace. While hallucinations have not disappeared, they are no longer the primary obstacle preventing organizations from realizing value from AI. In my conversations with enterprise leaders, a different challenge is emerging: understanding what AI systems are actually doing once they are deployed.

As organizations move beyond chatbots and experiments, they are increasingly deploying AI agents that perform multi-step tasks. These systems retrieve documents, query databases, call APIs, generate reports, write code, and make recommendations. The issue is not whether the model can reason. The issue is whether the organization can see, verify, and govern the decisions being made along the way.

That's an operational problem, not a model problem.

A useful way to think about modern AI is that the model is only one component of a much larger system. Around the model sits an operational framework that includes tools, workflows, permissions, retrieval systems, APIs, business rules, and human oversight. This surrounding environment determines how the model interacts with the organization and ultimately whether it delivers reliable outcomes.

Many AI failures today occur not because the model invents information, but because small errors accumulate across extended workflows.

Imagine an AI agent assigned to create a quarterly business report. It retrieves information from several internal sources, summarizes findings, pulls supporting data, drafts conclusions, and formats the final document. If a minor error occurs early in the process—perhaps an outdated document is retrieved or a metric is interpreted incorrectly—that mistake can influence every subsequent step. By the time the report reaches a human reviewer, the underlying issue may be difficult to trace.

I've seen similar patterns emerge across software development, compliance reviews, customer support operations, and data analysis workflows. The challenge is rarely a single catastrophic mistake. It's the gradual drift that occurs when multiple automated decisions build on one another.

Traditional software teams solved this problem years ago through monitoring, logging, testing, and observability. When a distributed application fails, engineers don't simply inspect the final output. They examine the entire chain of events that led to the failure.

Enterprise AI requires the same discipline.

Organizations need visibility into which tools an agent accessed, what information it retrieved, which decisions it made, where confidence levels changed, and where errors began to appear. Without that visibility, AI systems become operational black boxes. Teams may know something went wrong, but they cannot determine why.

This becomes especially important as organizations pursue greater levels of autonomy. Every executive wants AI to increase productivity, but productivity gains depend on trust. And trust comes from transparency.

One of the reasons AI coding assistants have gained traction is that their outputs are highly verifiable. Code can be compiled, tested, scanned, and validated automatically. Feedback arrives quickly, allowing errors to be identified and corrected before they propagate.

The same principle should apply across other enterprise workflows.

Before organizations deploy AI into business-critical processes, they should establish verification checkpoints, define measurable success criteria, and create clear escalation paths when confidence drops or unexpected behavior occurs. Human oversight remains important, but humans should focus on exceptions and anomalies rather than reviewing every action manually.

For IT leaders, this means expanding the conversation beyond model selection. The question is no longer simply, "Which AI model should we use?" The more important question may be, "How will we observe, govern, and validate the decisions that model makes once it becomes part of our operations?"

The organizations seeing the greatest success with AI are not necessarily deploying the largest models or the newest agents. They are building systems that make AI measurable, inspectable, and correctable.

That's why I believe the next major challenge in enterprise AI is not intelligence. It's operational visibility.

Hallucinations may grab headlines, but the real work of scaling AI successfully comes from understanding what happens between the prompt and the outcome.

Nils Bunge is Head of Product at Tsuga

Hot Topics

The Latest

Rapid AI adoption and the unique ways AI workloads operate is redefining the scope and structure of what these teams must deliver. This shift is forcing organizations to rethink how they manage scale, automation, and control, according to The State of SRE and Platform Engineering 2026, a new report from Dynatrace ...

AI is usually talked about as a software tool, but it also depends heavily on the network behind it. Whether a company is using AI for chatbots, automation, monitoring, analytics, or employee support, all of that information has to move across the network in a reliable and secure way. That means AI is not just an application decision. It is also an infrastructure decision. Before organizations rush into AI, they should ask a simple question: Is our network ready to support it? ...

Enterprise AI often lacks governed access to where business processes actually execute. Without that access, AI agents may be able to reason, but they cannot operate reliably across enterprise workflows. For AI agents to effectively carry out workflows, they will require integration-layer context and controls. Organizations can implement these prerequisites by providing AI with managed access to the middleware layer ...

Enterprise networks rarely behave the same way for very long. A routing adjustment in one region may unexpectedly alter application performance in another. A cloud migration may introduce hidden dependencies that go unnoticed until an outage occurs. All the while, the network is managed by several different teams, each of whom use different tool sets — and as a result, have different views of the network ... There’s usually an engineer who remembers why traffic fails over a certain way between sites, or which transparent firewall was added where. The problem is that human memory cannot scale alongside enterprise-scale networks ...

Ask an infrastructure team how confident they are in their ability to govern AI, and most will tell you they've got it handled. A recent survey of 406 IT decision-makers and platform engineering leaders found 86% expressing exactly that confidence. Ask the same group whether they have a formal written AI governance policy, and the number drops to 30%, according to Spacelift's Infrastructure Automation Report ...

In MEAN TIME TO INSIGHT Episode 27, Shamus McGillicuddy, EMA VP of Research, Network Infrastructure and Operations, and Parker Hathcock, EMA Research Director covering IT Service/Operations (ServiceOps), discuss observability unification in modern IT operations ... 

Virtual Private Networks became a cornerstone of enterprise security at a time when corporate infrastructure looked very different from today ... For years, this model worked well. But the architecture behind VPNs assumed a centralized corporate environment—one where the network itself was the hub of activity. In a cloud — first world, that assumption no longer holds ...

Website outages get resolved just as fast in August as they do in November. I went looking for the opposite: the summer slowdown everyone assumes is there once the people who fix things are away. It isn't in the data we collected, covering 1.8 million confirmed outages across tens of thousands of websites ...

This year, many of the cloud infrastructure contracts signed in the early days of the AI boom will come up for renewal. As the year goes on, I anticipate we'll see a significant amount of cloud vendor swapouts and multi-cloud adoption, and the reason isn't just GPU depreciation. It's because they're tired of their current cloud providers ...

There's a moment the many observability teams have experienced days into bringing a new service into production: you realize that the vendor's claims of "intelligent" behavior included a large serving of hype. Their dashboards look nice until they don't, the failure modes are a black box, and no one on the team can confidently explain why the system did what it did at 2 am. Agentic AI is about to force every Ops team to relive that moment at web-scale until they start treating these systems as the dependencies they actually are ...