Avoid using AI where a wrong answer is costly and hard to undo, where you cannot check its output, or where the data is too sensitive to share. Keep people in charge of legal judgments, irreversible actions and anything you have not measured. Use it elsewhere with checks.
Cases to avoid or keep manual
Irreversible or high-impact actions. Disabling an executive account, wiping a device or blocking a business-critical service should not run on model output alone. OWASP's excessive agency guidance lists excessive autonomy, meaning no independent verification or approval of high-impact actions, as a root cause of damaging outcomes, and recommends human approval for those actions.
Legal and regulatory decisions. Whether an event is notifiable, what to tell a regulator, and who is at fault belong to the people accountable for them, with counsel as needed.
Alert types you have not measured. If you have no reference set for an alert category, you have no idea how often the model errs on it. See metrics for AI alert triage.
Sensitive data without controls. Credentials, health or financial data and customer content should not reach a model unless you have redaction, an approved provider and reviewed data terms. See what log data to redact.
Pure lookups. If a plain query or script gives the exact answer, use that. A model adds cost and error for nothing. Checking a hash against a list is a job for code.
Attacker controlled text with powerful tools. A model that reads untrusted strings and also holds strong permissions is the risky combination. OWASP says there is no clearly foolproof prevention of prompt injection. Separate reading from acting. See prompt injection in SOC AI assistants.
Evidence you must defend. Anything that may end up in court, an audit or a dispute needs a verified chain of facts. A model summary can help you draft, but cite the original records.
Cases where AI is a good fit
These share a feature: a person checks the output cheaply.
- Summaries next to raw evidence.
- First drafts of queries, rules and runbooks that go through review.
- Explaining unfamiliar log fields or commands to a junior analyst.
- Drafting hunt write-ups and status updates for editing.
A simple test
Ask four questions about a task.
- If the model is wrong, what happens?
- Can a person catch the error quickly?
- Can the action be reversed?
- Is the data safe to share?
If you answer badly on any of them, keep a person in the loop or leave AI out. For help choosing a level, read how much autonomy an AI SOC agent should have.
Quick check
Does this mean AI has no place in the SOC?
No. It is useful for drafting, summarizing and explaining, where reviewers catch errors. The point is to match the task to the risk.
What is the first task to take away from AI?
Any action that changes a production system without human approval.
Sources
Last reviewed: 19 October 2026.
