Short · Identity and zero trust · 17 min read
When Is Unused Access Safe to Remove?
Unused access is a candidate for removal when the observation actually covers the permission and period in question, no current business need justifies keeping it, and the proposed change has a defined impact and acceptance test. A blank last-used field doesn't establish all three conditions. Recent activity doesn't establish a continuing need either: an unnecessary or inappropriate permission can still be used.
I would start an access review by asking what decision the reviewer can defend, not how quickly they can clear the queue. “Approve everyone who signed in recently” confuses activity with authorization. “Remove everyone who didn't” confuses an observation window with the business calendar. Both rules can be useful triage mechanisms. Neither should silently become the entire rationale for a consequential access decision.
This paper develops a method for reviewing existing grants when inactivity is the main signal. It addresses ordinary users and workload identities, with scoped examples from Microsoft Entra, AWS and Google Cloud. The framework and fictional case are analytical proposals; they have not been empirically validated. It does not report an assessment, promise disruption-free removal, or provide a command to delete production access.
Review the need for a capability, not the appearance of activity
NIST's least-privilege review control asks organizations to validate the need for assigned privileges and reassign or remove them as needed to reflect mission and business requirements (Joint Task Force, 2020, AC-6(7)). That places business justification at the center. A timestamp can inform the review, but it cannot establish what work an individual or service should be allowed to do next.
Define the unit being reviewed before gathering evidence. Is it an account, group membership, application assignment, role, individual permission or relationship to a resource? These units can overlap without being interchangeable. A person may actively use an application while having no need for its export privilege. A workload may use one permission in a broad role while leaving the rest unnecessary. Reviewing the whole role as a single “active” object conceals that difference.
I would describe each review item as a principal, capability, resource scope and grant path. Then add the business purpose that originally justified it and the person accountable for whether that purpose continues. That structure supports a concrete question: does this principal still need this capability over these resources through this arrangement? A display name and green recommendation are not enough context to answer it.
The objective is not maximum deletion. It is justified access with a visible treatment for uncertainty. Removing a critical recovery capability without a replacement can be a bad decision. Retaining broad standing power because someone writes “may be needed” can also be a bad decision. Both failures arise when the reviewer never has to identify the task, operating conditions and evidence behind the choice.
Ask what the timestamp measures
A last-sign-in field can record an attempt rather than successful use. Microsoft's Graph documentation says lastSignInDateTime includes unsuccessful interactive attempts. Its lastSuccessfulSignInDateTime property instead records the most recent successful interactive or non-interactive sign-in, and its historical data was not backfilled when introduced (Microsoft, 2025, “Properties”). Those semantics matter before anyone labels a user active or inactive.
Even successful authentication is not proof of using the specific capability under review. A person may sign in to read a dashboard but never use its administrative controls. Conversely, a background client may operate without the interactive sign-in the reviewer expected to see. I would keep authentication, resource access and privileged action as separate evidence categories, rather than mapping them all to a generic last-used column.
Microsoft's access-review creation guidance illustrates the scope distinction. Its no-sign-in recommendation uses a thirty-day interval including interactive and non-interactive sign-ins; for application reviews, the guidance identifies application sign-in rather than tenant sign-in as the relevant scope (Microsoft, 2026b, “Next: Settings”). That describes a configured recommendation, not a universal definition of business inactivity or proof that every application action has been observed.
Suppose a contractor signs in to the tenant for an unrelated collaboration tool. That observation can be real while contributing nothing to whether they still need access to a completed project's data. Conversely, the absence of a new sign-in may need explanation if access is exercised through an already-established application session. These are hypotheses to investigate for the actual application, not assertions that every system behaves identically.
The review record should therefore preserve the field name, source, principal identifier, target scope, success semantics and observation period. If a data pipeline renames an attempted-sign-in timestamp to “last successful use,” the error has occurred before the reviewer opens the campaign. A polished interface cannot repair a meaning that was lost during extraction.
Distinguish no observed use from no usable observation
There are at least three different states behind an empty activity field: no covered event occurred, the event type was not collected, or the evidence cannot be linked to this principal and permission. I would represent those states separately. Combining them into “unused” turns a data-quality gap into a false authorization conclusion.
AWS provides a useful concrete example. IAM last-accessed information can include attempted API access, including denied attempts. Its documentation also identifies limits: action-level information excludes data-plane events and iam:PassRole, and the report does not represent every policy type (Amazon Web Services [AWS], n.d.-a, “Things to know about last accessed information”). These are material boundaries on inference, not reasons to discard the tool.
A recent denied attempt should not establish that a business process requires the permission. An absent action entry for an unsupported action should not establish disuse. The appropriate next step differs: investigate the attempted operation in one case; identify another evidence source or retain an explicit coverage gap in the other. Automated recommendations become more useful when they preserve this distinction instead of erasing it.
The same issue appears in historical searches. CloudTrail Event history covers the last ninety days of management events in a selected Region and account, with limitations including the absence of data events and organization-wide aggregation (AWS, n.d.-c, “Event history limitations”). An empty result there cannot be represented as a comprehensive search of all resource activity across an enterprise.
I would ask the evidence owner to state which event classes, accounts, Regions, applications and identity mappings the query covers. Record missing intervals and retention boundaries. If the permission was granted only recently, the relevant exposure period may be shorter than the nominal report window. If logging began halfway through the period, extending the query's start date does not create the missing records.
Treat recommendations as models with declared boundaries
A recommendation engine can reduce a large review population to manageable questions, but its output has a scope. AWS distinguishes unused-access analysis from external- and internal-access analysis and documents that service-linked roles are excluded from unused-access analyzers (AWS, n.d.-b, “How IAM Access Analyzer generates findings for unused access”). A clean findings list is therefore not evidence that every category of identity has been assessed.
For each automated recommendation, retain the analyzer type, configured population, observation basis and known exclusions. An analyst reviewing the result later should be able to explain why this principal appeared and another did not. Otherwise, changes in configuration can look like changes in risk. A smaller queue may mean fewer excessive grants, or merely fewer grants included in the analysis.
Google Cloud's IAM recommender considers a bounded period of permission use and can use machine learning to identify permissions likely to be needed despite not being recently used. Its documented maximum observation period is ninety days, with configurable minimum periods for project-level recommendations (Google Cloud, 2026a, “Observation period” and “Machine learning for policy insights”). A recommendation is thus neither a raw event count nor a guarantee about future work.
I would ask the reviewer to distinguish observed use, model-inferred need and owner-stated need. All three may be relevant, but they answer different questions. If the model predicts that a permission will be needed, seek the business task that could justify it. If the owner disputes a removal recommendation, ask for the specific future or infrequent operation rather than treating the dispute itself as sufficient evidence.
Organizational proximity is another proxy. Microsoft's user-to-group affiliation recommendation uses the reporting structure and requires appropriate directory information, including a manager attribute (Microsoft, 2026c, “User-to-Group Affiliation”). That can identify an unusual membership. It doesn't establish whether a cross-functional auditor, incident responder or project specialist has a legitimate reason to sit outside the group's usual reporting line.
Build a compact evidence record that supports an actual decision
The review should not become an invitation to collect every log about a person. I would assemble only the information needed to assess the grant, with references to restricted evidence rather than unnecessary copies. Reviewers need scope and meaning, not an unfiltered export of authentication history or sensitive application content.
My proposed record has five parts: the capability being considered, the observed activity and coverage, the current business justification, the proposed change, and the acceptance conditions. Together they make it possible to challenge both a blanket approval and an unsupported removal. The following table is a design aid, not a vendor schema or a mandatory compliance form.
| Decision component | What I would record | What it prevents |
|---|---|---|
| Capability | Principal, permission or role, resource scope, grant path | Mistaking account activity for use of every privilege |
| Observation | Exact signal, success/attempt semantics, period, coverage gaps | Treating missing or irrelevant telemetry as disuse |
| Business need | Current task, accountable owner, operating cycle, conditions | Accepting “still employed” or “might need it” as full justification |
| Proposed treatment | Retain, narrow, time-bound, remove or investigate | Forcing every uncertainty into approve/deny |
| Acceptance | Authorized change, functional checks, residual gaps, next review trigger | Closing a review before the decision is implemented |
A permission can be observed and unjustified, unobserved and justified, or unobserved with uncertain coverage. Those combinations should lead to different work. For instance, use after a project ended can warrant investigation rather than retention. An annual reporting function can warrant a narrower, scheduled grant. A missing data source can warrant an evidence repair without claiming that either removal or retention has been proven correct.
I would also preserve the reviewed policy version and extraction time. A reviewer can make a reasonable decision about yesterday's grant while an administrator widens it today. That change should reopen the affected decision or be reconciled before application. The record needs a stable identifier connecting the decision to the actual entitlement; a human-readable label alone may hide a renamed or replaced object.
Choose an observation window from the work, not from a convenient default
A thirty- or ninety-day interval may be technically available and operationally useful. It is not automatically long enough to observe quarter-end settlement, annual certification, disaster recovery or a seasonal workload. I would ask the service owner to describe the operating cycle, then compare that cycle with the evidence window and the last occasion on which the capability should have been exercised.
This doesn't justify keeping every historical privilege for a year. It changes the choice of treatment. A role used only for a scheduled task might be replaced with time-bound activation or a narrower service capability. A recovery permission might belong in a separately governed emergency path with an exercise requirement. The argument is for matching access to the task, not stretching inactivity thresholds until every old grant survives.
Newly granted permissions require similar care. A permission created for work that starts next month will naturally have no past use. The right question is whether the future assignment is approved, when it begins and whether standing access is needed beforehand. A reviewer should not manufacture historical activity to validate a prospective need. Nor should a future start date excuse granting broader access than the task requires.
For dormant accounts, identify whether the review concerns continued eligibility, current authentication capability or a particular grant. Someone returning from a legitimate absence may need some relationships restored later, but preserving all access indefinitely is not the only design. The people responsible for the business process should define the acceptable lifecycle without exposing unnecessary personal details to the technical reviewer.
Make uncertainty visible without letting it become permanent approval
A binary campaign can pressure reviewers to choose “approve” when they lack enough information to defend either answer. I would allow a distinct investigation state with an accountable owner, a missing-evidence description and a resolution condition. That state should not be reported as recertified access. If the platform only supports binary decisions, retain the unresolved assessment separately and make the operational treatment explicit.
Uncertainty also needs a risk response. If a grant permits a consequential action and no owner can justify it, leaving it indefinitely unchanged is itself a decision. The appropriate response may be restriction, time-bound retention, an alternative workflow or removal under the organization's authority. The paper does not supply a universal default; it supplies the evidence questions that should inform the chosen policy.
I would separate evidence completeness from operational urgency. A critical service with weak telemetry may need temporary continuity while its owner repairs the evidence. An abandoned test account with clear ownership history may be removable without a lengthy observation exercise. The same blank timestamp should not force both cases into the same action, because the consequence and available corroboration differ.
A defensible exception names the capability retained, the reason, the accountable person and the event that ends or reopens it. “Business critical” without a task or service reference is too broad. So is “keep until the next review” when the next review has no better evidence source planned. An exception should create a path toward resolution rather than a recurring copy of the same unexplained approval.
Test proposed narrowing against both history and expected work
Historical simulation can identify a change that would have broken previously observed access. Google's Policy Simulator reports changed outcomes for historical attempts and separately reports replay errors and unsupported resource types (Google Cloud, 2026b, “Understand Policy Simulator results”). A result with no reported access changes must be read alongside those gaps; errors are not passing evidence.
Simulation is bounded by the events and conditions it can evaluate. An annual operation outside the available history, a new business workflow or an unsupported resource can remain untested. I would use simulation to discover specific counterexamples to a proposed change, then ask the owner about important operations that the replay could not have represented. That avoids treating historical compatibility as a forecast of every future need.
NIST's assessment guidance distinguishes examining artifacts, interviewing knowledgeable people and testing assessment objects (Joint Task Force, 2022, §2.4.2). For an access decision, those methods can complement one another: inspect the grant and activity evidence, discuss the actual task with its owner, and conduct an authorized exercise of the proposed state. A screenshot of a recommendation should not stand in for all three methods.
An acceptance exercise should include a permitted task that must continue and a capability intended to disappear. If the test environment lacks the resource relationships that matter in production, record that limitation. A positive test through an alternative broad grant can also hide a failure to narrow the intended route. Define which identity and authorization path the exercise is meant to evaluate before interpreting its outcome.
Restoration planning should be specific. Record how to recover the necessary capability if the change interrupts approved work, who can authorize restoration and whether restoration would recreate the entire broad role. A precise rollback can restore one needed permission while leaving unrelated excess removed. “Put everything back” may be necessary in an emergency, but it should remain a visible exception rather than the assumed definition of recovery.
Apply the decision through the real source of access
The reviewer can correctly decide to remove an entitlement while the integration fails to enact that decision. Microsoft's access-review completion guidance documents examples: nested-group membership can retain access, dynamic groups remain governed by their rules, and some synchronized or group-assigned relationships require changes elsewhere (Microsoft, 2026a, “Apply the changes”). Review completion is therefore not interchangeable with effective removal.
For this paper, the immediate requirement is to identify the object the campaign can change and the authority that owns the remaining grant path. A denial against an application should produce a follow-up when the actual access arrives through a group the campaign cannot modify. The reviewer should not be expected to infer that mismatch from a generic completion status.
Detailed revocation behavior deserves separate treatment. Temporary privileged-access expiry examines what must stop at resource enforcement points while ordinary access continues. Here, the prior question is whether the business and observation evidence justify the entitlement decision at all. Keeping those questions separate prevents a successful removal test from being mistaken for proof that the removal was appropriate.
I would close an item only when the treatment is recorded, the responsible system has applied it or an explicit follow-up remains, and the acceptance evidence matches the intended scope. Keep recommendations, reviewer decisions, applied changes and verified outcomes as separate counts. Otherwise, a campaign can report complete even though the unresolved work has merely moved out of its interface.
A fictional review with three different meanings of inactivity
Consider fictional Westmere, which reviews a reporting application and associated cloud roles. Three items are labeled unused. One belongs to an analyst whose project ended. Another supports an annual reconciliation task. The third is a workload permission whose action type is not represented in the selected report. A blanket inactivity rule would treat all three as equivalent.
For the analyst, the application owner confirms that the project is closed and no successor assignment requires the grant. The review identifies the direct assignment and defines a check that ordinary collaboration remains available while the project resource is denied. Removal is justified by the ended task plus the scoped evidence, not solely by the timestamp.
For annual reconciliation, the owner identifies the actual task, its next scheduled occurrence and the permissions it needs. Westmere proposes replacing a broad standing role with a bounded activation path and tests the required operation in an authorized environment. The exception record includes the new access mechanism and its owner. The absence of recent activity led to redesign rather than either permanent approval or unexamined deletion.
For the workload, the evidence owner discovers that the selected report cannot establish use of the relevant action. Westmere records a coverage gap, consults the service owner and identifies an appropriate source and exercise. It does not count that item as confirmed unused or safely removed. The service-identity ownership questions are developed further in workload identity lifecycle design.
Now add a complication: the analyst's access also arrives through a nested group. The business decision remains valid, but the initial removal does not complete it. Westmere routes the remaining relationship to its owner and verifies the intended result. This constructed case demonstrates separate failures of inference and execution. No real organization, measured savings or observed interruption rate is represented.
Measure decision quality and unresolved work, not just clicks
Campaign completion is an administrative measure. It tells you whether participants supplied responses under the campaign's rules. I would pair it with measures that expose the quality and implementation of those responses: items with a current purpose, items with usable observation coverage, decisions awaiting application, exceptions past their resolution condition and changes whose acceptance evidence is complete.
Define the denominator for each measure. An activity-coverage rate across imported review items is not coverage of every entitlement in the enterprise. A removal count can rise because duplicate entries were imported, because scope expanded, or because genuine excess was reduced. Preserve the campaign population and mapping rules so changes in the measure can be interpreted rather than celebrated automatically.
Track restoration and interruption after changes, but investigate the reason before scoring the reviewer. A restoration may expose a missed dependency, an inadequate test, a new business requirement or a decision that was intentionally temporary. Those are different lessons. Equally, no reported interruption does not prove every removed grant was unnecessary; the relevant operation may not have occurred yet.
Evidence handling matters to these measures. NIST's log-management guidance discusses inconsistent content, formats and timestamps across sources (Kent & Souppaya, 2006, §2.3.1). I would retain source identifiers and time semantics when joining review decisions to operational events, and flag unmatched records rather than treating them as successful outcomes. The older guidance supports this stable evidence principle, not contemporary platform-specific retention or logging defaults.
Counterarguments and limits
The strongest objection is scale: a large organization cannot investigate every grant as if it were a unique incident. I agree. Use automated signals to prioritize and standardize evidence collection. Apply pre-authorized rules to well-understood, bounded populations where their assumptions hold. What should remain visible is the difference between a routine case meeting those assumptions and an exception that does not.
A second objection is that business owners can rationalize almost any access. Requiring a concrete task, resource scope and operating cycle makes that claim testable, but it doesn't guarantee honesty or good judgment. Independent challenge may be appropriate for consequential permissions. The framework cannot eliminate collusion, hidden workflows, compromised telemetry or poorly understood application behavior.
There is also a cost to excessive caution. An indefinitely retained privilege remains available while the review searches for perfect evidence. The method does not require certainty. It requires a bounded statement of what is known, a proportionate decision under the organization's authority and an accountable treatment for what remains unresolved. Restrictions and time-bound access can be useful when neither blanket retention nor immediate deletion is satisfactory.
The twelve cited works establish control objectives, observation semantics and platform limits. They do not validate this combined decision framework or demonstrate a universal safe inactivity threshold. Product behavior, licensing and supported objects can change; the cited platform passages were checked for this manuscript. A real assessment must confirm the actual tenant, analyzer configuration, evidence sources and application paths before using the model operationally.
If your access review is mostly a spreadsheet of names and last-sign-in dates, I would start with one consequential permission population and trace what each recommendation actually means. Discuss an access-review evidence assessment. We can define the decisions reviewers need to make, identify the missing evidence and design a bounded acceptance exercise that distinguishes justified removal from a tidy-looking queue.
References
- Amazon Web Services. (n.d.-a). Refine permissions in AWS using last accessed information. Retrieved October 29, 2026, from https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_last-accessed.html
- Amazon Web Services. (n.d.-b). Understand how IAM Access Analyzer findings work. Retrieved October 29, 2026, from https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-concepts.html
- Amazon Web Services. (n.d.-c). Working with CloudTrail event history. Retrieved October 29, 2026, from https://docs.aws.amazon.com/awscloudtrail/latest/userguide/view-cloudtrail-events.html
- Google Cloud. (2026a, October 26). Overview of role recommendations. https://docs.cloud.google.com/policy-intelligence/docs/role-recommendations-overview
- Google Cloud. (2026b, October 26). Test role changes with Policy Simulator. https://docs.cloud.google.com/policy-intelligence/docs/simulate-iam-policies
- Joint Task Force. (2020). Security and privacy controls for information systems and organizations (NIST SP 800-53 Rev. 5). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.800-53r5
- Joint Task Force. (2022). Assessing security and privacy controls in information systems and organizations (NIST SP 800-53A Rev. 5). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.800-53Ar5
- Kent, K., & Souppaya, M. (2006). Guide to computer security log management (NIST SP 800-92). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.800-92
- Microsoft. (2025, July 5). signInActivity resource type. Microsoft Learn. https://learn.microsoft.com/en-us/graph/api/resources/signinactivity?view=graph-rest-1.0
- Microsoft. (2026a, March 12). Complete an access review of groups and applications. Microsoft Learn. https://learn.microsoft.com/en-us/entra/id-governance/complete-access-review
- Microsoft. (2026b, March 12). Create an access review of groups and applications. Microsoft Learn. https://learn.microsoft.com/en-us/entra/id-governance/create-access-review
- Microsoft. (2026c, March 12). Review recommendations for access reviews. Microsoft Learn. https://learn.microsoft.com/en-us/entra/id-governance/review-recommendations-access-reviews