IT & Systems Administration · Advanced
Translate Conditional Access policies into a plain-language matrix
An executive asks a simple question. "Who can sign in without MFA?" Your Microsoft Entra tenant has dozens of Conditional Access policies, several exclusions, guest rules, device conditions, and one report-only policy that looks active in a screenshot. A policy called "Require MFA for everyone" isn't proof that everyone is covered by it.
Understand the evaluation boundary
This guide uses AI to turn policy configuration into a reviewable description of the intended rules. It doesn't ask AI to act as the Conditional Access evaluation engine. Whether a sign-in succeeds depends on the actual conditions, identities, target resources, sessions, external dependencies and enforcement state. The matrix you'll build is an audit aid, not a verdict on any live sign-in.
Conditional Access decisions are built from assignments and conditions (who, which resource, under what circumstances), grant controls (block, or what must be satisfied), session controls, and policy state. Several policies can apply to one sign-in. One policy's display name never overrides another policy's block. Users or groups excluded from a policy are excluded from that policy only, not from Conditional Access in general.
The states enabled, disabled and report-only (enabledForReportingButNotEnforced) matter. Report-only is for observation; it doesn't enforce anything. Resource targeting matters too: a user whose Teams sign-in fails may actually be blocked when the client requests Exchange Online or another dependent service. See Microsoft Learn: planning Conditional Access and viewing applied policies.
Define the audit question precisely
"Review our Conditional Access" is too broad. Pick a concrete objective:
- Coverage: which defined user populations appear to require strong authentication for an identified resource?
- Exclusions: which policies exclude emergency access accounts, service identities, guests or named groups?
- Overlap: which policies target the same population and resource with different controls?
- Migration: how do today's enabled rules differ from a proposed report-only policy?
- Troubleshooting: why did a specific user's sign-in to a specific resource succeed, fail or prompt?
Name the population, resource, client type, device and location assumptions, policy state, and the time of the audit. A matrix without those dimensions can create false assurance.
Export the policy definitions, read-only
Use the Microsoft Graph v1.0 endpoint GET /identity/conditionalAccess/policies through an approved tool, with a role and permission suited to viewing policies. See Microsoft Graph: list policies.
If your organization allows Microsoft Graph PowerShell, this requests read-only access and saves a local JSON snapshot. Module availability and authorization vary by tenant. The script doesn't change any policy.
# Example from AI Field Manual. Provided as-is, without warranty.
# Read it, then test it in a non-production environment before any production use.
# Run from an approved administrative workstation, not a public shell.
Import-Module Microsoft.Graph.Identity.SignIns
Connect-MgGraph -Scopes 'Policy.Read.All' -NoWelcome
$policies = Get-MgIdentityConditionalAccessPolicy -All
$policies | ConvertTo-Json -Depth 40 |
Set-Content -Path '.\ca-policies-raw.json' -Encoding UTF8
# Store in a restricted working folder; do not upload the raw export.
Policy.Read.All is a Graph permission, not a guarantee that your account has every role it needs. Group IDs, named locations, authentication strengths, app IDs and roles may need separately authorized read-only lookups. Don't broaden permissions casually. If a property is missing or unresolved, keep its ID and label it unknown rather than inventing a display name.
What to preserve in the snapshot: each policy's ID, name and state; included and excluded users, groups and roles; included and excluded resources; condition selectors such as platform, location, client app and device filters; grant controls, including the operator (require all versus require one), built-in controls and authentication strength; session controls; and the date and time of the export. Avoid resolving privileged exclusions to personal names unless the audit needs it.
Normalize the policies into a matrix
Use a row for each meaningful policy and scope combination, not one sentence for the whole tenant. For example:
| ID / state | Who | Resource | Conditions | Control | Exclusions | Reviewer note |
|---|---|---|---|---|---|---|
| CA-01 / On | All users | Office 365 | All device platforms; interactive clients | Require authentication strength X | Emergency access group | Check the named strength resolves; check how guests are treated. |
| CA-02 / On | All users | Admin portals | Selected client types | Block under the stated condition | Emergency access group | Check the resources actually requested match the intended portal scope. |
| CA-03 / Report-only | Guests | SharePoint Online | Selected locations | Require approved device or MFA (operator matters) | None shown | Not enforcing; inspect the report-only results. |
These rows are fictional and simplified; they describe no real tenant. Don't flatten nested conditions into an "MFA: yes/no" column. "Require one of the selected controls" isn't the same as "require all selected controls." "All users" can include guests, while an excluded group or an app-specific scope can change the outcome substantially.
Use AI to explain the rules without claiming effective access
Act as a read-only Microsoft Entra Conditional Access policy reviewer.
Analyze ONLY the sanitized JSON and ID-resolution table supplied.
You cannot inspect live users, tokens, sign-in results or hidden policy
configuration. Do not assert that a user can or cannot sign in based
on the policy export alone.
AUDIT QUESTION: [specific user/resource/device/location/client scenario]
SNAPSHOT DATE: [time zone and timestamp]
POLICY JSON: [sanitized export]
ID MAP: [known groups, apps, locations, strengths; unknown IDs retained]
OUTPUT:
1. One matrix row per policy with state, scope, conditions,
include/exclude assignments, grant operator and session controls.
2. A human-readable interpretation for each row; cite policy ID and
exact JSON property path supporting it.
3. Apparent overlaps, exclusions and unresolved identifiers.
4. Differences between On, Off and Report-only policies.
5. Five targeted scenarios to validate using What If and sign-in logs.
6. A separate OPEN QUESTIONS list; never guess at missing values.
7. Findings classified as CONFIGURATION OBSERVATION, INFERRED RISK,
or VERIFIED SIGN-IN OUTCOME (only with separately supplied logs).
RULES: Never propose a production mutation command. Do not reduce
AND/OR grant controls to an ambiguous summary. A policy display name
is not evidence of effective scope. Do not call report-only enforced.
Ask for a matrix you can compare directly with the source JSON. Any "risk" label must explain its scenario and limits. An exclusion from a broad MFA policy deserves review, but it might be a documented emergency account with its own safeguards.
Validate with the actual platform
The What If tool simulates selected scenarios and lists the policies that would apply. It isn't a complete substitute for real sign-in testing, and Microsoft documents its limits around service dependencies. Compare its output with the Conditional Access details in Entra sign-in logs (subject to your permissions), including report-only results where relevant. See Microsoft Learn: the What If tool and report-only mode.
At minimum, exercise:
- A standard employee on a compliant, managed device in a normal location.
- The same employee on an unmanaged device or from an untrusted location.
- A guest accessing a specific SharePoint resource.
- A privileged administrator accessing an admin resource.
- An approved emergency access account, through its documented process.
- A known service or workload identity, evaluated as the correct identity type with the correct controls.
Some of these may be impossible or inappropriate to run in production. Record what was simulated, what was observed in logs, and what was tested in a controlled environment. If user-risk or sign-in-risk conditions are configured, make sure the reviewer understands their licensing and signal dependencies. If a resource makes downstream service requests, examine those too.
Emergency access. Design and test emergency accounts according to your documented recovery strategy. Microsoft advises keeping them usable when a Conditional Access or identity configuration fails, while still protecting and monitoring them. See Microsoft Learn: emergency access accounts.
Worked example: "MFA for everyone" isn't the whole story
Fictional scenario. A policy named "CA-01 MFA Required for All Staff" is enabled for all users and Office 365, with an emergency group excluded. "CA-05 External Locations" blocks access to a sensitive admin resource from non-approved locations. "CA-09 Stronger Auth" is report-only for privileged accounts.
The first AI draft says: "All employees must use MFA, except emergencies; admin users will need phishing-resistant MFA because CA-09 applies." That's misleading twice over. CA-09 is report-only and doesn't enforce anything. And "MFA required" describes a policy configuration for a defined scenario, not proof that every real sign-in was challenged and satisfied the control.
Corrected conclusion:
CA-01 is enabled and configured to require the specified control for included users requesting the named resources, subject to the recorded conditions and exclusions. CA-05 may independently block certain admin-resource sign-ins. CA-09 would require stronger authentication in simulated report-only scenarios but isn't enforcing. The configuration suggests coverage; actual outcomes need targeted What If and sign-in log validation. The emergency group exclusions must be reviewed and tested against the complete set of applicable policies.
Turn the findings into a controlled review
Tenant / environment: [sanitized]
Snapshot date / Graph endpoint / reviewer:
Question and scenario boundaries:
Policies inspected / unavailable data:
Matrix: ID | state | population | resource | conditions | grant |
session | exclusions | evidence path
Confirmed observations:
Potential gaps (not yet verified):
Exception register and owner:
What If scenarios and outputs:
Sign-in log samples and observed results:
Unknowns / licensing or dependency limitations:
Proposed changes (SEPARATE change request, no auto-apply):
Emergency access verification:
Approval / next review date / evidence archive:
AI tool and data-handling classification:
What stays with you: deciding which exclusions are intentional, whether a risk is acceptable, and when a change is safe to enforce.
Further reading on Microsoft Learn: Graph: list policies, grant controls, the What If tool, planning Conditional Access.