What’s Inside
Why invisible automation can create new risks for enterprises
What founders must know about training and overseeing AI agents
How NeuBird’s Hawkeye blends into IT teams and why that’s a strength and a concern
When the enterprise IT team at one of NeuBird’s clients started receiving detailed incident reports in their system, they assumed a new hire had joined.
The IT team began asking, “Who is this new employee? These reports are really good,” said Gou Rao, co-founder and CEO at NeuBird.
The client’s employees didn’t know the reports were being generated by Hawkeye, NeuBird’s AI agent.
For Rao, the confusion was a compliment, but it also points out a blind spot. What happens when software is so convincing, employees forget it isn’t human?
That’s the promise and the concern of agentic systems. AI tools are increasingly being deployed deep inside enterprise workflows, responding to alerts, solving infrastructure issues and doing it so stealthily that teams barely notice.
But that invisibility raises new questions: who’s in charge when the software acts independently? And what happens when something goes wrong?
This story is part of my ongoing series exploring how AI agents are reshaping startups. We previously covered the legal risks (Kelly Lawton-Abbott) and the limits of efficiency (Tarun Raisoni). In this installment, I spoke with Rao about the operational blind spots that can emerge when agents are too good at blending in and why he believes founders and investors need to rethink how they deploy, train and oversee these systems.
Founded in 2023 and backed by Mayfield and Microsoft, NeuBird develops an agentic AI platform focused on IT operations. Their flagship tool, Hawkeye, connects with systems like PagerDuty or ServiceNow and monitors infrastructure for anomalies.
It pulls from logs, traces, alerts and telemetry to generate real-time diagnostics—what Rao calls “root cause analysis reports”—before a human looks at the issue.
The product aims to reduce outages, speed up recovery and cut downtime-related costs. Rao cited one study estimating $400 billion in global losses from unplanned IT outages in a single year. In NeuBird’s early deployments, he said, customers saw as much as an 80% reduction in time spent resolving incidents.
But unlike some AI tools that require teams to change how they work, NeuBird designed Hawkeye to operate in a system’s background.
“We don’t ask teams to retrain or adopt new workflows,” Rao told me. “The agent adapts to them.”
In practical terms, that means Hawkeye receives the same alerts as a human would, but instead of forwarding them, it begins investigating and offering answers.
That seamless integration, Rao said, is the only way to make agents viable in a high-stakes enterprise environment. But it also means the automation can go unnoticed and unexamined.
Rao is quick to point out that AI agents don’t magically “know” what matters.
“There’s too much data, too many logs, metrics, alerts. You can’t expect an LLM to reason effectively if you just dump raw telemetry into it,” he said.
His team’s solution is to treat agents like new employees. Provide them with good inputs, train them on company context and allow for a learning curve.
“We tell customers Hawkeye will work well out of the box, but like any new engineer, it improves when you teach it your processes,” he said.
That comparison may be helpful, but it underscores a leadership reality: if agents require onboarding and oversight, then someone must be responsible for reviewing their actions and ensuring they’re aligned with company goals.
Hawkeye provides transparency by citing its sources and outlining its “chain of thought,” but Rao acknowledges that the need for some human validation never fully disappears.
In this way, agents don’t just reduce human labor, they’re reshaping it. Alert fatigue may drop, but someone still needs to supervise how the agent filters, clusters and responds to incidents.
Rao noted how much work is required for humans to manually sift through hundreds of daily alerts. Hawkeye may alleviate that burden, but it’s not a hands-free or fully autonomous system.
At the enterprise level, there’s growing interest in deploying agents across the org and not just in IT, but in sales, marketing and operations, Rao said. Unlike conventional software, agents can’t be judged by rigid outputs. They are, in his words, “more cognitive,” and their behavior may shift depending on the problem.
That makes managing them a new kind of leadership challenge.
Agents operate asynchronously. They reason. And while Rao believes they “make quality of life better” for teams, he also warns that they can’t be treated like plug-and-play tools.
In some ways, the success of Hawkeye may be its biggest risk: when automation fades into the background, it’s easy to stop paying attention.
“You have to ease it in,” he said. “You can’t ask people to change overnight, but you can’t just ignore the agent either.”
In the coming years, he expects enterprise AI agents to become commonplace, as they’re integrated into workflows, monitored like teammates and, increasingly, personalized.
Rao pointed to one customer in Latin America whose engineers, after realizing Hawkeye wasn’t human, asked if it could deliver responses in Spanish and even suggested tweaking its tone.
“That kind of personalization,” he said, “is probably the next step for agentic systems.”
It’s a small but telling signal that the more AI agents resemble coworkers, the more teams expect them to behave—and sound—like one of their own.
The moment we start asking AI agents to sound like teammates, we risk treating them like teammates. But ask why an AI agent failed, you may get a confident answer, not a clear one.
That’s why supervision, even if it’s ever so subtle and ongoing, matters more than ever.
