AI in the SOC: agents, SOAR and human oversight
A SOC receives more alerts every day than a human team can review calmly. The promise of AI is clear: let a machine do the first pass, gather the evidence and propose what to do. The question risk committees ask me is a different one: who is accountable when the agent gets it wrong?
In this article I explain what AI is changing in security operations centres, where the real risks are and how to govern it so that it fits NIS2, DORA and the EU AI Act. I write from experience in highly regulated banking environments, not from a vendor demo.
AI in the SOC: from processing alerts to owning cases
Until recently, SOC automation mostly meant SOAR: hand-written playbooks that enrich an alert, open a ticket or isolate a host when a rule fires. It works well with the predictable and poorly with the new.
AI agents change the model. They no longer follow a fixed script: they reason over evidence, query several sources, form a hypothesis and propose (or execute) an action. Gartner describes this as a shift in the SOC’s primary function, from processing alerts to end-to-end case ownership. It also expects security operations teams to have 50% more engineers than analysts by 2028.
My reading: AI does not eliminate the SOC, it turns it into an engineering team that designs, monitors and corrects its own agents. If you do not plan for that change in profile, the agent ends up deciding alone simply because nobody is left to supervise it.
What AI already does in a SOC: machine learning and SOAR
Before talking about agents, it helps to separate what AI already does reliably in a SOC. Most of it is machine learning applied to four tasks:
- Alert triage. Models trained on closed alerts separate false positives from cases worth investigating and group duplicate alerts from the same incident.
- Anomaly detection and behaviour analytics (UEBA). The system learns what is normal for each user, host or service and flags what deviates: logins at odd hours, unusual data volumes or lateral movement.
- Incident prioritisation. A score combines alert severity, asset criticality, exposure and threat intelligence to decide what is looked at first.
- Automated response. SOAR runs playbooks (isolating a host, blocking an indicator, forcing a password reset) when confidence passes a defined threshold.
Used well, it multiplies the team’s capacity. But everything depends on the data it learns from: if the history is incomplete or badly labelled, the model repeats those mistakes faster. That is why its quality has to be measured (precision and false negatives found in reviews) and its drift watched over time, like any other control.
The four autonomy levels of an AI SOC
The most common mistake I see is talking about “AI in the SOC” as if it were one thing. It is not. Before approving anything, define with the team the level at which each use case operates:
- Assistance. The agent summarises, enriches and prioritises. The analyst decides everything.
- Recommendation. The agent proposes a specific action with its justification. The analyst approves or rejects it.
- Bounded autonomy. The agent executes reversible, low-impact actions (for example, quarantining an email) within defined limits, and logs them.
- Conditional autonomy. The agent acts on critical systems (isolating a server, blocking a privileged account) only with preconditions, confidence thresholds and guaranteed rollback.
In banking, almost anything that touches a critical system should stay at level 2 at first. You move up a level with data, not enthusiasm: measured accuracy, reviewed false positives and tests of what happens when the agent fails.
The real risks of AI in the SOC
The risks that concern me are not science fiction. They are operational, and they can be managed once they are named:
- Explainability debt. If the agent dismisses an alert and nobody can reconstruct why, you have lost the chain of evidence. In a reportable incident that is a regulatory problem, not just a technical one.
- Manipulation of the agent itself. An attacker who understands how the agent reasons can plant data so that their activity is classified as benign. The agent’s input is attack surface.
- Excessive permissions. An agent that can isolate hosts or block accounts is one more privileged identity, and it must be governed as such (I cover this in non-human identities and AI agents).
- Loss of craft. Gartner warns that over-reliance on automation can erode teams’ foundational analysis skills. If analysts only validate what the machine says, they stop knowing how to investigate.
- Automation bias. When the agent is right many times in a row, the human stops really reviewing. Oversight becomes a click.
Automated prioritisation: bias, unequal treatment and proportionality
There is a risk that gets less attention and has an ethical dimension, not just a technical one: when a machine decides what is handled first and how to respond, it is treating assets, teams and people differently.
- Bias inherited from the data. The model learns from what the team investigated in the past. Areas that were barely monitored will keep looking low-risk, even when they are not.
- Unequal treatment. Prioritising by the value of the asset or the user can leave incidents unattended in less visible sites, subsidiaries or groups or, the other way round, subject some employees to more monitoring than others without a risk-based reason.
- Proportionality of the response. Locking a person’s account or isolating their laptop because of an anomaly has real consequences for their work. The response must be proportionate to the evidence and graduated: contain what is reversible first and let a person decide anything that affects someone.
This also has a legal side. If the SOC analyses employee behaviour, the GDPR applies: a legal basis, prior information, a data protection impact assessment and limits on automated decisions that affect people, plus employees’ digital rights at work. And if the system is used to monitor or evaluate the behaviour of workers, it may be considered high-risk under Annex III of the EU AI Act, which requires effective human oversight.
The controls are simple to state: document the prioritisation criteria, sample what the system left behind, compare outcomes across areas or groups to spot differences that risk does not explain, and require human approval for any action that falls on a person.
AI in the SOC and regulation: NIS2, DORA and the EU AI Act
Regulation does not forbid using AI in the SOC, but it does require you to prove what happened and who decided.
NIS2 and DORA. Both require incident management with short reporting deadlines. An agent that speeds up detection helps; one that silently dismisses a major incident makes you miss the deadline without knowing it. I detail the deadlines and how to meet them in 72-hour incident notification under GDPR, NIS2 and DORA. In Spain, the NIS2 transposition law is still pending; what already applies is summarised in NIS2 in Spain.
EU AI Act. A SOC agent is not automatically a high-risk system. It can be, if it acts as a safety component in the management of critical digital infrastructure (Annex III). Narrow preparatory tasks, on the other hand, may fall outside under the Article 6(3) exception. Classification is done case by case and must be documented. Following Regulation (EU) 2026/1744, Annex III high-risk obligations apply from 2 December 2027 (see the EU AI Act postponement). Postponed is not removed: what the regulation asks for —effective human oversight, logs that allow events to be reconstructed, risk management— is exactly what good SOC practice requires.
How to govern an AI SOC: seven controls
These are the controls I recommend having in place before letting an agent touch operations:
- An inventory of agents and use cases, with each one’s autonomy level and a named human owner.
- A full record of every decision: what data the agent saw, what it concluded, what it did and who approved it. Without it, no audit is possible.
- Minimal, separate permissions for each agent, reviewed like any privileged account.
- Input and output guardrails against manipulation and data leakage (along the lines of guardrails for generative AI).
- Human sampling of what is dismissed: a percentage of the alerts the agent closes is always reviewed. It is the only way to measure what you do not see.
- Quality metrics, not just speed: false negatives found in sampling, reopened cases and time to human decision, as well as mean time to respond.
- Failure drills: test what happens when the agent is wrong, goes down or receives manipulated data, and keep a manual mode ready.
The decision that adds most value: what not to automate
In regulated environments, the biggest step forward rarely comes from automating more. It comes from deciding, with judgement, what not to automate. Let the machine enrich, correlate and prioritise; keep actions on critical systems going through a person until the data shows the agent is reliable. That reduces noise for the team without losing a chain of evidence a supervisor can review.
For the broader picture of what AI can and cannot do in a bank’s cybersecurity, see AI in banking cybersecurity.
Frequently asked questions about AI in the SOC
Will AI replace SOC analysts?
No. It changes their work: less manual alert review, more detection engineering, agent supervision and investigation of complex cases.
Is an AI agent in the SOC a high-risk system under the EU AI Act?
It depends on the use. It can be if it acts as a safety component of critical digital infrastructure; narrow preparatory tasks may fall outside. It has to be classified and documented case by case.
Where do I start?
With an assistance use case, clear metrics, a human owner and a full decision log. Move up the autonomy ladder only with data.
Are you considering bringing AI agents into your SOC and need it to fit NIS2, DORA and the EU AI Act from day one? Let’s talk.
Do you have to apply this under DORA, NIS2 or ENS? Tell me about it.
Book 20 minutes