What you will learn, and the big picture: AI as a fast junior, you as the decider
What you will learn in this module. By the end you will be able to run a repeatable alert triage routine in which AI does the slow, boring parts (summarising, formatting, suggesting enrichment) and you do the parts that need judgement (checking the evidence, deciding, owning the outcome). You will learn how to give an AI tool safe and useful input, how to enrich an alert with context, how to verify every claim against raw logs, how to spot and resist automation bias, and how to decide between escalate, close and wait. Everything uses fictional companies, .example domains, documentation IP ranges and outputs marked nk-sim. Think of it like driving with a satnav: it is fast and usually right, but you still watch the road, and you are the one holding the wheel.
- Triage
- Quickly sorting incoming cases by how urgent they are, so the most serious get attention first.
- SIEM
- The central system that collects logs from many devices, runs detection rules on them and raises alerts.
- Enrichment
- Adding facts that the alert itself does not contain, such as who owns the host or whether the address was seen before.
- Escalate
- Hand a case to a more senior person or to incident response because it may be real and serious.
Prerequisites. Finish the earlier module in this track, AI for Security Operations Centre (SOC) and NOC engineers: start here (ano-start). We assume you know what AI does well and badly, the task-risk scoring idea, the data-hygiene rules (no customer secrets in public tools) and the trust-but-verify habit from its first triage example. This module teaches the full method behind that habit. Basic SOC vocabulary (alert, SIEM, severity, true positive, false positive) is helpful; we define each term again when it first matters.
What is in the nine chapters
- This chapter: the big picture and the analogy.
- Anatomy of an alert and what a good triage record contains.
- Summarise: asking AI for a short, structured summary.
- Enrich: adding context about the host, the user and the indicator.
- Verify: checking every AI claim against raw evidence.
- Automation bias: why we believe confident text, and how to resist.
- Escalate, close or wait: the decision and the closing note.
- A ticket-driven triage workflow, including what to do when the AI is wrong.
- Summary and exam checklist.
Start with an analogy: the hospital triage desk
In a hospital emergency department, a triage nurse sees every arriving patient first. The nurse does not treat anyone. She takes a few measurements, asks a few questions and decides who is seen immediately, who can wait and who can go home with advice. Now imagine she gets a very fast assistant who reads the patient forms aloud and writes a neat summary: "chest pain, 54 years old, probably indigestion". The assistant saves minutes. But if the nurse simply believes the summary and never checks the pulse, the one patient who is actually having a heart attack goes to the waiting room.
Alert triage is the same. The SIEM (security information and event management system) is the front door. Each alert is a patient. The AI tool is the quick assistant: excellent at reading and summarising, sometimes wrong, always confident. You are the nurse. The measurements are your raw logs. The decision, and the responsibility, stay with you.
The triage loop. AI speeds up steps 2 and 3. Steps 4 and 5 belong to the human.
In the hospital analogy, what is the "pulse check" for an alert, and why can the AI not do it for you?
Opening the raw log lines and trusted tools yourself. The AI only sees what you pasted, can be confidently wrong, and cannot be accountable, so the check and the decision stay with you.
Three jobs, three owners
- Summarise
- Turn a long alert and its log lines into a short, structured description. AI is good at this.
- Enrich
- Add context: who owns this host, is this user an administrator, has this IP been seen before. Tools and databases do this reliably; AI can suggest what to look up but should not invent the answers.
- Decide
- Escalate, close or wait, with a reason. This is always human.
Worked example. A mid-size SOC (fictional, Kestrel Logistics) receives about 600 alerts a day across 4 analysts. Manual triage averages 7 minutes per alert, which is 4,200 minutes, or 70 hours, against 32 analyst-hours available per day (4 analysts working 8 hours each). With a good routine, AI drafts the summary in 20 seconds and suggests enrichment steps, and the analyst spends about 4 minutes per alert, mostly verifying. That is 2,400 minutes (40 hours), still more than the 32 hours available, so the team also tunes its noisiest rules to cut the alert count (module ano-detect). The saving is real (about 30 hours a day), but only if the 4 minutes include the verification. If the analyst skips verification to save another minute, the saving is paid back in missed incidents.
Common beginner mistake. Pasting the alert into an AI chat and copying the verdict ("benign, close") into the ticket. The ticket then records the AI's opinion as if it were your finding. Always record what you checked, not what the AI said.
Exam trap. A question may ask who is accountable for a triage decision made with AI help. The answer is the analyst or the organisation, never the tool. "The AI recommended it" is not a justification.
The fast close that was not
At Kestrel Logistics (fictional) a new analyst closed 14 alerts in an hour using an AI chat, a record. Next morning a senior analyst reviewing a sample found that one "benign virtual private network (VPN) login" had come from a country the user had never visited, and the AI summary had called it "a normal remote login" because the pasted text did not include the geolocation field. The analyst had not looked at the raw log. The login was a compromised password. The team added a rule: every closed ticket must name at least two raw log lines the analyst personally checked.
Lesson: speed without verification only moves the risk to tomorrow.
"How would you use AI in alert triage?"
Strong answer: AI drafts the summary and suggests enrichment, I verify each claim against the raw logs and trusted tools, then I decide and write the closing note myself. I never paste sensitive data into unapproved tools, I treat the AI output as a draft from a fast junior, and I record the evidence I checked. Mention automation bias by name.
The routine saves about 3 minutes per alert. Why does the worked example insist that the 4 remaining minutes include verification?
Because one missed incident costs far more than the minutes saved. The saving is only real if the quality of decisions stays the same.
Key takeaways
- Triage is a loop: alert, AI draft, enrich, verify, decide.
- AI speeds up reading and writing; the human owns checking and deciding.
- Enrichment answers come from tools and databases, not from the AI's memory.
- Record the evidence you checked, not the AI's opinion.
- Time saved is real only if verification stays in the routine.