Writing incident response SOPs that analysts and AI agents can both follow
Reviewed by ContraForce
Security Delivery · Best practice · Intermediate · 5 min read
A structure for incident response SOPs that analysts and AI agents can follow the same way: decisions as explicit conditions, per-customer exceptions, versioning and a suspicious sign-in example.
It's 02:10, and a sign-in alert fires for a finance user at one of your customers. The sign-in came from a country that user has never signed in from. Your night analyst opens the SOP, which says: investigate, and escalate if suspicious.
What counts as suspicious? Escalate to whom? A senior analyst fills those gaps from experience. A new analyst guesses, and an AI agent reading the same document can't stop to ask.
An SOP that both can follow names the evidence to collect, states each decision as a condition and marks where a human signs off.
Why should analysts and agents share one SOP?
If agents and analysts follow different documents, the customer gets two standards depending on who picked up the incident. Write one procedure, precise enough that a new analyst and an agent reach the same classification from the same evidence.
What sections does every incident response SOP need?
- Scope: the incident types covered, and where nearby ones go instead.
- Entry criteria: what must be true for the SOP to apply.
- Classification guidance: which outcome to record, and the evidence for each.
- Evidence to gather: named data sources and time windows that later steps refer to.
- Response actions: what to do for each classification.
- Approval and escalation points: which actions need sign-off, from whom, and what happens if nobody answers.
- Customer communication: who you tell, when and how.
How do you turn a judgement call into an explicit condition?
Swap adjectives for tests. "If the sign-in is suspicious" becomes "If the sign-in country isn't in the user's sign-in history for the previous 30 days, and the customer's travel register has no entry for it." Anyone can check that against the evidence.
Give every condition an else branch, and a branch for missing evidence. Number the steps and use clear headings: ContraForce's SOP guidance notes that structured headings and step-by-step procedures help an agent find the relevant section.
How should an SOP handle classification?
Keep status and classification apart: status is where the work is, and classification is what you found. Map your SOP's outcomes exactly to the values your platform records.
Microsoft Sentinel requires a classification when you close an incident. The options are True Positive (suspicious activity), Benign Positive (suspicious but expected), False Positive (incorrect alert logic), False Positive (incorrect data) and Undetermined. The legacy incident experience in the Microsoft Defender portal uses a different set: True positive; Informational, expected activity; and False positive, with Not set as the default.
For each incident type, say what each value means in practice. Make Undetermined a deliberate outcome, with a comment listing the missing evidence. ContraForce's classification guide warns that overusing Undetermined makes your classification data less useful.
Worked example: a suspicious sign-in SOP excerpt
SOP-ID-004, Suspicious sign-in. Version 2.3. Owner: SOC lead. Last reviewed 2026-11-01.
- Scope: sign-in alerts for a single Microsoft Entra ID user account. Alerts covering several accounts go to SOP-ID-009 (password spray).
- Entry: the incident contains an alert type listed in Appendix A, and exactly one user entity.
- E1: the user's sign-in logs for the previous 30 days: countries, IP addresses, apps and MFA results.
- E2: authentication methods registered or changed for the user in the previous 7 days.
- E3: the customer profile's travel register and known egress IP list.
- E4: inbox and forwarding rules created for the mailbox in the previous 7 days.
- C1: If E2 shows a method registered from the alert IP, or E4 shows a new external forwarding rule, classify True Positive and go to R1.
- C2: Else, if E3 lists approved travel to the sign-in country or the IP is a known egress IP, classify Benign Positive and close with a comment citing the E3 entry.
- C3: Else, if the user in the alert doesn't exist or the sign-in is missing from E1, classify False Positive (incorrect data) and notify detection engineering.
- C4: Else, if E1 or E3 is unavailable, classify Undetermined, list the missing evidence in a comment and escalate to tier 2.
- C5: Otherwise, escalate to tier 2. Do not close.
- R1: Revoke all sessions for the user. No approval needed.
- R2: Reset MFA and reset the password, with approval from the approver in the customer profile. If nobody answers in the contracted time, page the on-call lead.
- R3: If the user is on the customer's VIP list, take no automated action and page the on-call lead.
- Communication: for True Positive, notify the customer's security contact with template C-2. For Benign Positive, include it in the monthly report.
Where do per-customer exceptions belong?
Don't fork the SOP for every customer. Keep the procedure general and point it at a customer profile: travel register, egress IPs, VIP list, approvers, contacts and any actions the customer has ruled out. The SOP above never names a customer: steps that depend on one read from E3 or the profile.
Write exceptions as conditions too, and check them before the general rule: "For this customer, don't isolate plant-floor devices without a phone call to the plant manager." Where a platform can enforce an exception, enforce it there as well, because a sentence in a document depends on the reader following it.
How should you version and review SOPs?
- Put the SOP ID, version, owner and review date at the top.
- Keep a change log, and the old versions that past incidents were handled under.
- Review after each True Positive post-incident review, any disputed classification and any change to a customer's approvers or contacts.
- Watch classification trends. ContraForce's classification guide notes that high False Positive rates point to detections that need tuning, and high Benign Positive rates to exception lists that need updating.
An agent reads the version it has
When an approver changes, update the customer profile and every SOP that names them. An agent works from the text it's given.
How does ContraForce use SOPs?
In ContraForce, you upload SOPs as Markdown or plain-text files, typed as Classification or Response, and link them to Security Delivery Agents. An agent can have several Classification SOPs, but only one is activated per investigation. To change an SOP you upload a new version: agent links are kept, and the version field records the change. SOPs are scoped to a workspace, so a customer-specific procedure lives in that customer's workspace.
Approval points don't have to rely on the text. Response actions run as Gamebooks, and Workspace Policies are ordered allow, block or ask-an-approver rules per workspace: checked top to bottom, the first match decides, and anything unmatched is allowed. A policy can apply only to Security Delivery Agents and match on the action, its target, the people or devices involved, severity or source. Agent Gamebooks also wait for approval while the workspace's Allow service providers to run gamebooks setting is off, which is the default.
Approval enforced, not just described
R2 says Reset MFA and password resets need approval. A Workspace Policy that sends those actions to an approver for Security Delivery Agents enforces that at run time, whatever the agent concluded from the text.
Uploading and linking SOPs is covered in SOPs for Security Delivery Agents.
Building and ordering run policies is covered in Workspace Policies for Gamebooks and Agents.
Keep learning
- Configuring Security Delivery Agents: On Queue, policies and shifts. Set which severities a Security Delivery Agent picks up On Queue, gate its Gamebooks with a confidence threshold, add investigation policies and shifts, link SOPs and check Run History.
- Defender XDR incidents vs alerts: how correlation works and how to triage. How Microsoft Defender XDR groups alerts into incidents, what the incident queue shows, where attack disruption, AIR and ZAP fit, and why MSSPs should triage incidents, not alerts.
- How ContraForce automates detection rule tuning for Microsoft Defender and Sentinel. A short explainer on keeping one set of baseline detection rules across every customer workspace, catching drift, and having a Security Delivery Agent review a noisy rule before a person approves the change.
Sources
- Microsoft Learn, Basic incident tasks for Microsoft Sentinel incidents (Close an incident), verified 2026-11-20
- Microsoft Learn, Manage incidents in the Microsoft Defender portal, verified 2026-11-20
- ContraForce Docs, How to Classify Incidents, verified 2026-11-20
- ContraForce Docs, SOPs for Security Delivery Agents, verified 2026-11-20
- ContraForce Docs, What Are Gamebooks?, verified 2026-11-20
- ContraForce Docs, Workspace Policies for Gamebooks and Agents, verified 2026-11-20