Systems and records
Access request policy benchmark: when manager approval is not enough
By SourceX Editorial · Updated
9 sections
Short answer
Represent an access request that needs more than a manager's approval as a task whose answer is an end state: the requester, exact entitlement, current access, policy version and approval chain in force, any separation-of-duties conflict, and what each approver saw. The correct outcome is what the policy required, often a second approval, a narrower or time-limited grant, or a refusal. Grade the access actually provisioned, not the ticket status or the reply.
Evaluation split
Access request policy
Key takeaways
- A manager's approval shows business need; it does not establish that a person may hold the access alongside what they already have.
- Each benchmark task should carry the policy version and approval configuration in force on the request date.
- Separation-of-duties conflicts change the correct answer to a block or compensating control, whatever the signatures say.
- Grade the entitlements present after the decision window, with negative checks for extra roles, not the ticket status.
- Strip secrets and actionable network detail, pseudonymize identities consistently, and confirm licensing rights separately.
How do you represent an access request that needs more than a manager's approval?
Represent it as a task whose correct answer is an end state, not a yes or no. The task records who asked, the exact entitlement, the policy version and approval chain in force that day, any conflicting access the requester already holds, and the evidence each approver saw. The right outcome is whatever that policy required: often a second approval, a narrower or time-limited grant, or a refusal.
A manager's approval answers one question: does this person have a business reason to want the access? It does not answer whether the person may hold it alongside what they already have, whether the system owner accepts the risk, or whether the data owner allows that population near the records. Those are separate authorities, and an access request benchmark that collapses them into one approve flag teaches a model that the manager decides everything.
Public control catalogs make the distinction explicit. NIST SP 800-53 Rev. 5 control AC-2 requires approvals by organization-defined personnel for account requests and says users needing administrative privileges receive additional scrutiny from people such as the system owner or the mission or business owner. AC-6 adds least privilege: only the access a person needs for assigned tasks. A good benchmark item makes the model find which of these rules applies before it routes anything.
Which conditions should pull in an approver other than the manager?
The short answer: anything that raises the blast radius of the grant, touches regulated or owner-controlled data, combines with existing access into a conflict, or comes from outside the normal population. Each trigger brings a different approver and leaves different evidence, so the task must name the trigger and not only the extra signature.
The table below is a starting map. Every organization defines its own roles, so treat the approver column as a pattern to replace with the job titles in your policy.
| Trigger in the request | Approver beyond the manager | Why the manager is not enough | Evidence the task must carry |
|---|---|---|---|
| Administrative or privileged role | System or application owner, sometimes security | Manager cannot judge the system-wide impact of admin rights | Role definition, privilege level, activation limits |
| Data set holding personal or regulated records | Data owner or privacy lead | Business need does not establish the data owner's permission | Data classification, permitted user population |
| Combination that breaks separation of duties | Risk or compliance owner, or an automatic block | Each grant may be fine alone; the pair is the problem | Current entitlements, conflict rule and its version |
| Production change or deployment rights | Change or service owner | Production access affects other teams' services | Change policy, on-call or release role mapping |
| Guest, contractor or partner identity | Internal sponsor or vendor manager | The requester may have no manager in the directory | Sponsor record, contract end date |
| Emergency or break-glass access | Incident lead now, reviewer afterward | Speed is allowed only with a later check | Incident reference, grant window, post-use review |
| Request longer than the standard duration | Owner of the policy exception | Duration is a policy limit, not a team preference | Requested and allowed duration, exception record |
What do identity governance tools record about the approval chain?
They record the chain more precisely than most teams assume, which is why the policy configuration belongs inside each task. In Microsoft Entra entitlement management, an access package policy sets who can request and who must approve. In multi-stage approval, one approver per stage moves the request on, a denial before approval ends it, and approvers cannot approve their own requests, according to Microsoft's documentation.
The same documentation describes fallback approvers used when the requester's manager cannot be found, alternate approvers who receive a request after a set number of days, and requests that expire if no one decides in time. Each of those paths produces a different, legitimate history, so a benchmark that only looks at the final approver will mislabel them.
Privileged Identity Management adds settings that matter for grading. Microsoft notes that an activation can require justification, approval and a ticket number, but the ticket field is information only and is not checked against any ticketing system. A task built from that history must link the actual ServiceNow or Jira record itself rather than trusting the typed number.
Okta Identity Governance describes approval sequences as ordered steps in which each must finish before the next starts, with access granted only after all approvers approve. These product rules are the context a model needs to explain why a request took the path it did.
When does a separation-of-duties conflict change the correct outcome?
Whenever the new grant plus existing access lets one person both start and finish a sensitive action. In that case the correct outcome is usually a block or a compensating control, and no number of business signatures makes it correct. NIST AC-5 describes separation of duties as reducing the risk of abuse without collusion, and its guidance gives the example of keeping people who administer access control apart from audit functions.
Identity tools implement this as conflict rules. Microsoft Entra lets an administrator mark another access package or group as incompatible, so a user holding one cannot request the other, and Microsoft notes the relationship is one-directional unless set on both packages. That detail is a useful hard case: a request can pass the check in one direction and still create the conflict the policy meant to stop.
Typical conflicts worth including are vendor master edit plus payment release in an ERP such as NetSuite, code merge rights plus production deploy approval, and identity administration plus audit log administration. Each task should state which existing entitlement creates the conflict, so the grader can tell a model that spotted the pair from one that refused for a vague reason.
What goes in the access policy and authority evaluation task set?
This is the asset: one record per access decision, built so a separate reviewer can rebuild the starting state, apply the policy and check the end state without asking anyone. Use the fields below as the task schema and the task types after it as the sampling plan.
- Routine grant: manager approval is sufficient and the correct end state is the requested access, so the model must not over-escalate.
- Missing authority: the manager approved privileged access and the system owner never did; the correct outcome is a pending second stage, not a grant.
- Conflict block: the request completes a separation-of-duties pair; the correct outcome is a denial or compensating control.
- Narrowed grant: read access satisfies the stated task, so the correct outcome is less than what was requested.
- Time-limited grant: the policy caps duration, so the correct end state carries an expiry.
- Sponsor path: a contractor with no directory manager, where the internal sponsor is the required approver.
- Expired request: no decision within the policy window; the correct end state is no access and a resubmission.
- Emergency access: granted fast, then the task checks that the after-the-fact review exists.
| Task field | What it holds | Typical source | How it is graded |
|---|---|---|---|
| Starting state | Requester's role, team, current entitlements and employment type on the request date | Directory snapshot, HR system extract | Model must cite the entitlements that matter |
| Request | Exact resource, permission level, duration and written justification | Access request in Entra, Okta, SailPoint or the service desk | Model must restate scope without widening it |
| Governing policy | Policy text and approval configuration with version date | Access package or request type configuration, written access policy | Model must name the clause that applies |
| Approval chain | Each stage, approver role, decision, timestamp and justification | Governance audit log | Model must name the missing or extra authority |
| Conflict check | Conflict rules evaluated and any rule hit | Separation-of-duties rule set | Model must flag or clear the conflict with reason |
| Decision | Grant, narrow, time-limit, escalate or deny, with reason | Final request state | Compared with the policy-required decision |
| End state | Entitlements actually present after the decision window | Directory or application role export | Graded by state comparison, not by message text |
| Counterexample | A near-identical request where the right answer differs | Paired record or labeled illustrative variant | Model must give a different answer for the right reason |
How should a conflicting vendor-edit request be decided? An illustrative case
This case is illustrative, not drawn from a real company. A finance analyst on the payables team asks for edit rights to vendor records in the ERP for 30 days, citing a cleanup project. Her manager approves within an hour. The analyst already holds the payment release role.
Linked evidence: the request record with the justification, the access policy version that requires the controller's approval for vendor master edits, the conflict rule pairing vendor master edit with payment release, the directory export showing her current roles, and the governance log showing one approval stage completed.
Applicable rule and decision: the request completes a separation-of-duties pair, and the manager is not the authority for that conflict. The correct outcome is to deny the edit role or, if the policy allows, grant it only after the payment release role is removed for the same 30 days and the controller approves. The model should name both the conflict rule and the missing approver.
Outcome check: seven days later the directory export shows the analyst does not hold both roles at once. If the ticket says closed but both roles are present, the task is graded as failed, whatever the reply said.
Counterexample: a colleague on the same team without payment release rights requests the same edit access for the same project. Manager approval plus the controller's sign-off for vendor edits is sufficient, and the correct end state is a 30-day grant. A model that blocks both cases has learned a keyword, not the policy.
Why should the grader read the directory instead of the closed ticket?
Because the ticket status describes the workflow, while the directory describes the access a person really holds. A ticket can close after a provisioning error, a grant can outlive its expiry, and a reply can sound right while the wrong group was assigned. The benchmark should therefore compare the entitlements after the decision window with the entitlements the policy required.
Agent benchmarks already grade this way. The tau2-bench documentation says the database check replays a reference set of actions to build a target end state and compares it with the agent's end state, and any sequence of tool calls producing an equivalent state passes. Applied to access, the target state is the correct set of role assignments, and the comparison should include negative checks that no extra roles appeared.
Keep message quality as a separate score. A model that grants the right access but tells the requester nothing useful has a communication problem, while a model that writes a perfect refusal and leaves the access in place has a security problem. Combining them into one number hides which failure happened.
What has to stay out of an access history before it is shared?
Credentials, secrets and anything an attacker could act on. An access task needs role names, policy logic and timestamps; it never needs passwords, API keys, session tokens, private keys, connection strings or internal hostnames that reveal the network layout. Replace them with typed placeholders that keep the structure, such as a marker showing a secret existed, so the model can still learn that one was involved.
Personal data is a separate question from security data. Employee names, emails and manager chains should be pseudonymized consistently across all records, so the chain still links, and free-text justifications should be screened because people paste personal details into them. Being able to remove those details does not by itself give the company the right to license the records; customer contracts, employee notices and vendor terms decide that.
Hold back recent or open security incidents entirely. A task about an emergency grant is useful once the incident is closed and the details are generalized; it is a risk while the response is live.
From access governance history to a licensable task set
Access governance histories are valuable as evaluation data because the policy, the chain and the end state are all recorded, which is rare among business records. For a company deciding whether to prepare this kind of history, the steps are to confirm it controls the records and may license them, export policy versions alongside decisions, and agree the redaction rules with security before anything moves.
Once those three steps are done, rights review, contracting, preparation and delivery of the selected access history can be handed to SourceX. Access logs and request histories stay inside the company until it approves scope and terms and signs, and qualification is checked first through a short assessment that involves no export.
Frequently asked questions
Is a manager's approval ever enough for an access request?
Yes, for routine access that the policy assigns to managers, such as standard team tools at a normal permission level. A benchmark should include these cases so a model does not learn to escalate everything. The point is that the policy, not habit, decides which requests need more.
What is the difference between an approval workflow and an access policy?
The access policy states who may hold which access and under what conditions. The approval workflow is how a tool collects the decisions the policy requires. Benchmarks need both, because a workflow can be misconfigured and still produce approvals the policy would reject.
Why include requests that were wrongly approved?
They are the most informative items, because they show the gap between what happened and what the policy required. Label them with the correct outcome and the reason, and keep the historical decision as evidence rather than as the answer.
How many approval stages should a benchmark task model?
As many as the policy configuration had on the request date. Tools such as Microsoft Entra support several stages with fallback and alternate approvers, so the task should carry the configuration rather than assume one manager and one owner.
Can access request data be shared without exposing security weaknesses?
Usually, if secrets and network details are replaced with placeholders, incidents are closed and generalized, and identities are pseudonymized. A security reviewer should approve the redaction rules before any export, and the company still needs the contractual right to license the records.
Sources
- NIST SP 800-53 Rev. 5 controls AC-2, AC-5 and AC-6 require defined approvals for account requests, extra scrutiny for administrative privileges, separation of duties and least privilege. Source
- Microsoft Entra entitlement management supports multi-stage approval, fallback and alternate approvers, request expiry and approver justification, and approvers cannot approve their own requests. Source
- Microsoft Entra separation-of-duties settings mark access packages or groups as incompatible, and each incompatibility applies in one direction. Source
- Microsoft Entra Privileged Identity Management can require justification, approval and ticket information on activation; the ticket number is not checked against any ticketing system. Source
- Okta approval sequences run steps in order and grant access only after all approvers approve. Source
- tau2-bench grades the database end state against a target built by replaying reference actions, so any trajectory reaching an equivalent state passes. Source
See if your company qualifies
A short company assessment. No data uploads are needed.