Microsoft Security Copilot - Promptbooks and Workflow Automation

Summary

A promptbook turns one good investigation into a routine anyone can run. This article compares prompts, promptbooks, and agents, shows how to build and share your own, explains how Azure Logic Apps connect to Security Copilot in both directions, and closes with the access, capacity, and review guardrails that keep automated results trustworthy.

Three Analysts, Three Methods

Using A.M Finances Group as our business example. On Thursday morning, a high-severity incident lands in the queue, and three analysts handle incidents like it in three different ways. One summarizes the incident and writes a report. Another lists the entities first, then summarizes, then writes the report. The third summarizes, builds a timeline, lists the entities, and never writes a report at all. All three are capable, yet the results are inconsistent, and the security manager cannot tell which steps were skipped.

The fix is not more discipline. It is a way to capture the best routine once and let everyone run it. In Microsoft Security Copilot, that is a promptbook.

Prompt, Promptbook, or Agent

A prompt is a single natural-language request, ideal for an ad-hoc question. A promptbook is a series of prompts grouped to accomplish a specific security task. The prompts run in sequence, each one builds on the previous response, and the final prompt produces the summary or report. An agent is an AI-driven workflow with a defined goal, and it can respond to user requests or run on events and schedules. Microsoft's guidance maps them to the work: use prompts and promptbooks for ad-hoc analysis, and agents for automation. The three also combine well. Start with an agent for structured workflows, use prompts to explore the results, and return to agents for repeatable execution.

Security Copilot also ships with ready-made promptbooks, and each asks for one input. The Microsoft Sentinel and Microsoft Defender XDR incident investigation promptbooks take an incident number and return an executive report for a nontechnical audience. Suspicious script analysis takes a PowerShell or Windows command-line script, Threat actor profile takes an actor's name, and Vulnerability impact assessment takes a CVE number or a common vulnerability name. To run one, select the Prompts button in the prompt bar, search by name, enter the input, and select Submit. A few need a plugin or roles first, so check the requirements before the first run.

Build and Reuse Your Own

The best promptbooks start as real investigations. In a session, run the prompts you want to keep, then check the boxes beside them to turn on the Create promptbook button. Add a name, tags, and a description. You can edit the prompt text and reorder the prompts, though changes can affect the results. Choose who can use it, either Just me or Anyone in my organization, select Create, and then View to test it. Promptbook creation is available to all Copilot roles, including publishing one for the tenant.

Parameters make a promptbook reusable. Replace a value such as a threat actor name with a name in angle brackets and no spaces, like <ThreatActor>, and the promptbook asks for that value each time it runs. You can define more than one. To keep a run going when a single prompt fails, edit the promptbook and turn on Continue on failure. Later, the Promptbook library lets you edit, duplicate, share, or delete it. Reordering prompts can change both the results and the SCU use, because each step builds on the one before it.

Automate in Both Directions

Azure Logic Apps connects to Security Copilot in two directions. Inbound, the Security Copilot connector adds actions to a workflow, on both the Standard and Consumption plans, to submit a prompt, submit a promptbook, or execute an agent. Build the workflow in the Azure portal, set the trigger, add the action, and fill in the inputs. A promptbook action needs every required input, such as an incident ID from the trigger, or the workflow will not run successfully. The person who creates the connection needs Security Copilot access and access to the underlying data. An agent must be set up in the default workspace, and it runs under its own run identity, separate from the connection identity. A.M Finances Group could run its investigation promptbook on every new high-severity incident and deliver the executive report to the on-call analyst.

Outbound, a Logic App becomes a skill that Security Copilot calls. Build the workflow with an HTTP Request trigger, where the skill's inputs arrive as top-level fields of the request body. Describe it in a skillset manifest that names the subscription, resource group, workflow, and trigger, then upload the manifest as a custom plugin. After that, prompts, promptbooks, and agents can invoke it. The Security Copilot tenant must be the same as the Logic App's tenant, and the workspace needs provisioned security compute units. Microsoft's example is a promptbook that analyzes a risky user, reports on it, and after confirmation emails options to add the user to a Microsoft Sentinel watchlist or update Microsoft Entra.

Guardrails Before You Automate

Security Copilot does not go beyond the access you have, so a promptbook returns what the person running it is allowed to see. That is why two colleagues can get different results from the same promptbook. Each plugin also keeps its own role requirements, such as Microsoft Sentinel Reader for the Sentinel plugin. Capacity matters too. Security Copilot is priced on security compute units, and promptbooks and agents use them like any other feature, so start small and watch the usage dashboard.

Treat every output as a draft. AI responses can vary between runs, long-running promptbooks can be canceled, and edits made in a session do not change the saved promptbook unless you save them. Review the results to confirm they are accurate before you act.

Conclusion

A promptbook turns individual expertise into a shared routine, and Logic Apps make that routine run where the work already happens. Start with a ready-made promptbook, build one from a session you trust, add a parameter, and then let a workflow call it. A.M Finances Group went from three analysts and three methods to one routine that every analyst runs the same way, with access, capacity, and review keeping the results worth trusting.