We’ve all heard of human-in-the-loop oversight. It’s become the safe, responsible default for basically every AI deployment, across every industry, in every use case. But it’s the wrong default for most of what a SOC handles.
At SOC alert volume, human-in-the-loop defeats the purpose of automating in the first place. Signing off every action becomes the bottleneck automation was supposed to eliminate.
SOCs need a more nuanced oversight model for agentic defense. They need to choose that model based on the risk and reversibility of the action.
What are the Limitations of Human-in-the-Loop for SOCs?
The NIST AI Risk Management Framework (NIST AI RMF) mandates that organizations establish robust accountability and human oversight AI security processes for high-stakes systems.
Organizations aren't bound to one oversight model, though. SOCs default to human-in-the-loop because it's the most rigorous option available, and in a market full of AI hype and AI skepticism, that rigor buys peace of mind. But rigor and fit aren't the same thing. Applying human-in-the-loop oversight to a modern SOC is like limiting a Formula 1 car to 30 miles per hour. It's much less likely to crash, but the car can't do what it was designed for.
By its nature, human-in-the-loop requires analysts to sign-off every AI action. That would be okay if the AI only took a handful of actions every day. But a modern, AI-enabled SOC handles thousands of alerts every day.
If analysts have to sign off every single one, what was the point of automating in the first place? All it achieves is shifting effort to a different part of the workflow, rather than eliminating it, and turning teams into button-clickers.
Remarkable as it might sound, human-in-the-loop oversight is something of a relic. Many teams adopted it before AI models were accurate enough to trust with more complex tasks. Much of investigation automation today doesn’t demand that level of oversight.
Why is Human-on-the-Loop a Better Fit For Most SOC Decisions?
Human-on-the-loop oversight is a more efficient alternative for most SOCs.
In this model: the AI acts, a human monitors, and a human can intervene, rather than forcing
that human to sign off every action. The human is no longer an approval gate, and becomes an overarching reviewer who watches patterns of outcomes.
Some of you may balk at this suggestion. It might sound like a compromise, a prioritization of efficiency over control in agentic systems. But in reality, it’s the correct operating model for the majority of mid-risk, reversible, high-volume threat AI threat detection, investigation, or response decisions.
Of course, there are caveats. Human-on-the-loop oversight only works when teams have confidence in the AI’s baseline accuracy and explainability. Reviewing outcomes relies on trusting the basic process that produced them.
Where Does Human-in-the-Loop Still Earn Its Place?
However, there are some situations where human-in-the-loop oversight remains essential. If actions are irreversible and the potential consequences are severe, a human needs to sign them off before the AI executes.
For example, if an AI is wrong about needing to isolate or contain a critical system during a suspected ransomware attack, the wider organization could experience unnecessary downtime. And that downtime can be costly: a 2024 Oxford Economics study found that the average Global 2000 company loses $200 million a year to IT shutdowns.
And don’t be fooled into thinking that a more advanced AI could make these kinds of decisions without human approval. However advanced an AI tool is, it can still make mistakes. If the consequences of those mistakes are severe and irreversible, teams cannot allow them to happen. Human-in-the-loop oversight is the best way to do that.
What Determines the Right Model for a Given Decision?
Deciding what model is right for what decision requires a simple equation: Risk X Reversibility.
That means low-risk, reversible, high-volume decisions are the best candidates for on-the-loop oversight. Examples might include disabling a user’s MFA push after repeated failed attempts, blocking a known-malicious IP or domain at the firewall or proxy, or quarantining a phishing email across mailboxes after detonation confirms malicious payload.
High-risk, high consequence, and hard-to-reverse decisions, however, deserve in-the-loop oversight. Disabling a domain controller or core identity provider, wiping or reimaging a device before forensic capture, or terminating a live customer-facing service during a suspected breach are all examples of decisions that demand greater AI oversight in security operations.
However, some cases are borderline, and require other considerations.
Locking an executive’s account for anomalous behavior, for example, is the same action as a
normal user lockout, but has a higher potential impact, so might be better served by in-the-loop.
Similarly, auto-remediating a cloud misconfiguration (like closing an open S3 bucket) would usually be on-the-loop, but if that bucket is actively serving production traffic, reversibility drops and in-the-loop would be better suited.
What Should Security Leaders Look for in a Vendor?
When choosing a SecOps vendor, security leaders need to broaden the questions they ask. It’s no longer enough to determine just whether or not a platform has human oversight for AI security.
To make the most of automation safely, security leaders need to ask:
Which decisions default to on-the-loop review?
Which stay in-the-loop?
Who set the mapping?
Are decisions auditable after the fact?
ReliaQuest GreyMatter is an example of tiered oversight in practice.
Lower-risk, reversible actions run with the analyst reviewing outcomes and patterns rather than approving each one. Higher-consequence actions still route through explicit analyst sign-off before execution. That means SOCs are more efficient, can handle more alerts, and, crucially, lose none of the safety inherent in blanket human-in-the-loop security operations.
A platform that applies the same oversight rule to every action, regardless of risk or reversibility, hasn't solved the problem. It's just moved the bottleneck. The decision, not the system, is what should set the oversight level, and that's the question worth asking any vendor claiming to have "solved" AI oversight.
To see for yourself how ReliaQuest GreyMatter implements tiered oversight decisions, schedule a demo today.
