Home / Blog / SIEM and Tools

Intermediate SIEM and Tools

What an Agentic SOC Is and Where Human Approval Fits

AI triage agents are reaching SOC queues. Learn what an agentic SOC is, how NCSC oversight models apply, and where analysts approve, review and own the outcome.

An agentic SOC is a Security Operations Centre where AI agents carry out parts of the investigation, such as reading an alert, gathering context and drafting a verdict, while people keep control of the decisions that carry risk. The idea is simple, but the details decide whether it helps or hurts. This guide is for SOC analysts and team leads in financial services who want a clear picture of how agentic triage works and where human approval belongs.

Core ideaAgents investigate, people approve risky actions
Best first jobUser reported phishing, high volume and low risk
Key sourceNCSC interim advice on agentic AI, August 2026
>

Scenario: At 06:40 on a Monday, 312 staff at a bank have reported the same fake invoice email. The first analyst on shift would normally open each report, check the sender, detonate the link and write a note. Today a triage agent has already grouped the reports, checked the sender domain and the link, attached its reasoning to each case and marked 296 as the same known campaign. It has also proposed deleting the message from every mailbox. The analyst's job has changed: instead of opening 312 cases, they review the evidence and decide whether to approve one bulk action.

What an agentic SOC actually is

A chatbot answers a question. An agent is given a goal, chooses a tool, uses it, reads the result and decides what to do next, repeating the loop until it finishes or gets stuck. In a SOC the tools are things you already use: the SIEM (Security Information and Event Management) search, a threat intelligence lookup, the ticketing system, the email security console.

Vendors use the phrase in different ways, so check what a product really does. As far as I could verify, Microsoft's Phishing Triage Agent in Microsoft Defender classifies user reported phishing alerts, explains its decisions in plain language and learns from analyst feedback. Microsoft has also described a Security Alert Triage Agent that extends the same approach to identity and cloud alerts, but its availability has been changing, so check the current Microsoft documentation before you plan around it.

What it readsAlerts, emails, sign in logs, endpoint events and threat intelligence.
What it decidesA proposed verdict, such as malicious, benign or needs a human, with the evidence behind it.
What it may doAdd notes, enrich cases and, only if allowed, take a response action.

The third pillar is where the risk sits. An agent that only reads and summarises is a better search box. An agent that can disable accounts or delete email is a new kind of actor in your environment, and it needs the same controls you would apply to a privileged person.

Where analysts fit: three oversight models

The NCSC (the UK National Cyber Security Centre) published interim advice on managing the cyber risk of agentic AI in August 2026. It says controls should be proportionate to how much autonomy an agent has, and it describes three ways to oversee one. The advice is interim, and the NCSC says formal guidance will build on it and eventually replace it, so treat it as a starting point and check for updates.

Most control

Human in the loop

A person approves each action before it happens. Slower, but the right default for anything that changes a mailbox, an account or a firewall.

Analyst approves the mailbox purge
Balanced

Human on the loop

The agent acts and a person monitors and can step in. Suits reversible, low impact work where speed matters.

Agent closes a duplicate report, analyst samples the closures
Least control

Human out of the loop

The agent acts with no human review. The NCSC ties controls to potential impact, so this needs the strongest safeguards and a clear case for it.

Rarely right for containment

For consequential activity, the NCSC also recommends that named individuals or groups are responsible for the agent's actions, that its behaviour is monitored in real time, and that it can be stopped quickly. In a regulated firm that last point matters: when an auditor asks who approved an action, "the agent" is not an answer.

Start with one narrow job

Several vendor and analyst write ups suggest starting with a low risk, high volume task such as phishing triage, beginning in an observe mode where the agent proposes and people decide. That matches the NCSC's suggestion to run early experiments in office hours, when more human oversight is available, and to widen the hours only once you trust the controls. A sensible ladder looks like this.

  1. S1
    ObserveThe agent triages in parallel with your analysts and writes proposed verdicts. Nobody acts on them. Compare results for a few weeks.
    Read only
  2. S2
    ProposeAnalysts see the agent's verdict and evidence first, then confirm or correct it. Every correction is a lesson for the agent and a data point for you.
    Analyst decides
  3. S3
    Act on the reversibleThe agent may close clear duplicates or add notes by itself. A sample of its closures is reviewed every day.
    Sampled review
  4. S4
    Gate the riskyContainment, such as deleting email or disabling accounts, always waits for a named human approver, with the evidence on screen.
    Approval gate
+

Tip: an agent only improves if people correct it. If your analysts treat the verdicts as background noise, the feedback is lost and quality stays where it started. Make confirming or correcting a verdict part of the normal workflow, not an extra step.

Write the permissions down

Before an agent touches a live queue, write down what it may do alone, what needs approval and what is never delegated. The NCSC advice says each agent should have its own unique identity, with only the permissions the task needs and credentials that last as short a time as possible. The sketch below is an illustration of the idea, not the format of any product. The identity and team names are placeholders.

Example agent policyYAML
agent: phishing-triage-01
identity: svc-agent-phish-01        # its own account, not an analyst's
owner: soc-email-team-lead          # a named, accountable owner
mode: propose                       # observe, propose, then act
may_do_alone:
  - read_reported_email
  - lookup_sender_reputation
  - add_case_note
needs_approval:
  - delete_email_from_mailboxes
  - block_sender_domain
  - disable_account
never_delegate:
  - contact_regulator
  - any_action_on_executive_mailboxes
stop_switch: disable svc-agent-phish-01 and revoke its tokens

Read it as a contract. The first list is small on purpose. The second list is where your analysts add value. The third is not a technical limit but a decision about accountability. The last line matters as much as the rest: the NCSC says you should always be able to halt agent activity immediately, which can mean cutting its access as well as stopping the process.

!

Watch out: a prompt is not a control. The NCSC advises combining clear instructions with technical and operational controls, because an agent can pursue a goal in ways you did not expect. If a rule matters, enforce it with permissions, not just wording.

Treat the agent as an identity you monitor

The NCSC advises treating agent activity as a form of user activity, included in your normal 24/7 security monitoring and incident response, with logs that are protected from change. That turns the agent into something your own SOC watches. A few checks to build in from the start:

  • The agent has its own account, separate from any analyst or shared service account.
  • Every action it takes is logged with the case it relates to and the evidence it used.
  • Its logs are stored where the agent cannot edit or delete them.
  • An alert fires when it does something outside its policy, such as calling a tool it was never given.
  • Someone owns a monthly review of its verdicts, corrections and false positive rate.

Two risks deserve special attention in a bank, because the agent reads content that outsiders control.

Prompt injection through reported content

High
What it is

A phishing email or web page contains text that tries to give the agent new instructions, such as telling it to mark the message as safe.

Defence

Treat all reported content as data, never as commands. Keep the agent's own instructions separate and gate any action that follows from untrusted text.

Quiet growth in permissions

Medium
What it is

The agent starts with read access, then gains write access for a pilot and keeps it. Its scope grows without a fresh decision.

Defence

Review its permissions on a schedule, record each change with an approver and remove access that is no longer needed.

What to study next

  • Read the NCSC blog "Managing the cyber risk of agentic AI" on ncsc.gov.uk, especially the sections on oversight, observability and emergency shutdown.
  • Read the OWASP Top 10 for LLM applications, starting with prompt injection and excessive agency.
  • Practise on paper: pick one alert type in your SOC and fill in the policy above for an imaginary agent.
  • Next in this series: how to test a triage agent with a set of expected answers before you trust it.

Key takeaways

  1. An agentic SOC uses AI agents to investigate and propose, while people keep the decisions that carry risk.
  2. Match oversight to impact: in the loop for containment, on the loop for reversible work.
  3. Start with one narrow job such as phishing triage, and move from observe to propose to act in stages.
  4. Give each agent its own identity, minimal permissions, a named owner and a stop switch.
  5. Monitor the agent like any other user, and enforce rules with permissions, not only prompts.

You do not need to trust an agent to benefit from one. You need to know what it may do, who answers for it and how to stop it. Get those three things written down first, and the rest of the rollout becomes an ordinary change you can review.

Written by Riz, Senior Cybersecurity Consultant, Finance

Keep going

More in SIEM and Tools