Read a sample: 1. Map the Application Before Choosing Controls
A team is preparing to launch an internal AI assistant. The demonstration looks promising: an employee asks a question, the assistant searches company documents, and a concise answer appears with links to relevant sources. The team’s first security meeting quickly turns to controls. Should it enable more logging? Add a content filter? Restrict network access?
Those may all be useful questions, but they are premature. Before choosing a control, the team needs to know what the application can reach, whose authority it uses, where information travels, and what could happen when a response is wrong or inappropriate.
This chapter begins with a fictional application that will recur throughout the book. Its architecture is deliberately ordinary: a web interface, an application backend, an AI service, a document retrieval system, and operational records. The difficulty lies not in any one component but in the relationships among them. An employee may be authorized to use the assistant without being authorized to read every document the assistant can retrieve. A service identity may need access to a search system but not to every source connected to it. A useful diagnostic log may also contain information that deserves protection.
A map makes those relationships visible. It does not need to be a finished architecture diagram, and it does not prove that the application is secure. Its purpose is to give the team a shared, testable account of the system before assumptions harden into design decisions.
Northstar Help is a proposed internal assistant for a fictional organization called Northstar Services. Employees will use it to ask questions about internal procedures, locate guidance, and understand routine workplace processes. A question might be, “How do I request access to a project workspace?” The assistant will search approved internal documents, provide a draft answer, and cite material the employee can review.
The initial release is intended to answer questions, not make changes. It cannot approve access, alter employee records, or submit requests on a user’s behalf. If an answer concerns a consequential decision, such as eligibility for a workplace benefit, the employee is directed to the responsible team and the applicable source material. Northstar’s owners understand that a fluent response is not itself an authoritative decision.
The proposed flow is as follows:
1. An employee signs in to the Northstar Help web application and submits a question.
2. The web application sends the question and relevant session context to a backend.
3. The backend requests relevant material from a search service populated from approved document repositories.
4. The backend sends the question and selected retrieved material to an Azure AI service for answer generation.
5. The backend returns the answer and source references to the employee.
6. Application and platform components produce operational or security records for authorized staff to review.
This description is enough to begin scoping, but not enough to approve the design. “Approved document repositories,” for example, says nothing about whether all employees may read every document in them. Nor does it say how a document’s permissions are represented in the search system, or whether the employee’s identity is considered when results are selected.
The fictional team has not yet resolved those questions. That uncertainty is useful: a map should distinguish what is proposed from what has been verified.
A system map begins with purpose. Northstar Help exists to help employees find and interpret internal guidance. That purpose informs which data is necessary, which actions are appropriate, and what failure would mean.
The team should also describe uses that are outside the initial scope. Employees should not rely on Northstar Help as the final authority for an access approval, personnel decision, or interpretation that requires specialist judgment. This boundary is not enforced merely because it appears in a project document. The interface, response design, review paths, and tests must eventually support it.
Next, identify the people who interact with or affect the application:
- Employees submit questions and receive answers. They may have different permissions to source material.
- Document owners create and maintain the guidance that the assistant may retrieve.
- Application developers and operators build, deploy, configure, and troubleshoot the service.
- Identity and access administrators manage relevant identities, authentication settings, and access assignments.
- Security and governance staff assess risks, review evidence, and help define response and oversight requirements.
- Human reviewers or subject-matter teams handle questions or decisions that should not be resolved by the assistant alone.
These categories are starting points, not a complete organizational chart. One person may hold several responsibilities, and some responsibilities may be shared. The important scoping question is whether each meaningful decision and operational task has an identifiable owner.
The team should also consider people who do not intentionally interact with Northstar Help. A document may contain information about an employee who never submits a prompt. An answer may be copied into a ticket and influence a decision about someone else. Mapping only the direct user would miss those downstream effects.
A component inventory gives names to the places where data is processed, stored, or passed onward. For Northstar Help, the first inventory contains the following elements:
The inventory should name actual services and environments when the team knows them. During early scoping, it is acceptable to record an unresolved choice, such as whether a search component will be shared with another application. It is not acceptable to treat that choice as irrelevant. Shared infrastructure can…