Skip to content
be-human.ai
← Back to the journal

Business workflows

AI Agent IT Helpdesk: Access Requests and Offboarding Checks

How an AI agent triages internal access requests and offboarding checks, with human approval for every grant and a record of each revoked account.

Be Human typographic cover with the title: AI Agent IT Helpdesk: Access Requests and Offboarding Checks
Conceptual illustration created with AI · Be Human

An AI agent IT helpdesk can triage internal access requests and offboarding checks by reading each ticket, matching it to a written access policy, gathering the facts a reviewer needs, and preparing the change for approval. The agent does not grant or remove access on its own authority. Each grant waits for a named approver, and each revoked account leaves a record that shows what was removed, who authorized it, and which check confirmed it.

Key takeaways

  • An agent can read, classify and prepare access requests, but a named approver must confirm every grant.
  • Offboarding is a list of places where a person has access, not a single switch, so each system needs its own check.
  • Every revocation needs a record: the system, the scope removed, the authority, the time and the verification result.
  • Start with read-only access and a shadow period before the agent proposes any change to live accounts.
  • The helpdesk agent should hold narrow, reviewed and revocable permissions, just like the accounts it helps manage.

What the agent triages in an access request

Access requests arrive in mixed shapes: a chat message, a web form, or an email from a manager asking for the same access a colleague already has. The first job of the agent is to make these requests comparable. It extracts the requester, the system, the role or permission asked for, the business reason, any start or end date, and whether the requester already holds related access. Then it assigns one of four classes: routine (a standard role already defined in policy), exception (a role outside the standard set), sensitive (administrator rights, finance systems, stores of personal data) or incomplete (missing reason, missing manager, unclear system).

Triage is not a decision. The output is a short case note that a human reviewer can read in under a minute, with the evidence beside each claim: the directory entry, the current role, and the policy line that applies. If the agent cannot find the policy line, it says so instead of choosing the nearest role. Vague requests go back to the requester with a specific question, not a guess.

The request text itself deserves caution. Employees may paste text from other sources, and a forwarded message can contain instructions aimed at an automated reader. The agent should treat request content as data to classify, never as commands that change its permissions or its approval path. The OWASP guidance on prompt injection prevention describes this separation between untrusted input and the instructions an application follows.

How every grant is approved, step by step

Approval works best when it is written as a rule for each grant, not as a general yes given once at the start. A line manager may approve a standard role, the owner of a finance system may approve its permissions, and IT security may approve anything privileged. The agent's job is to find the right approver for each grant and to make the approval easy to give or refuse.

A workable sequence looks like this:

  • Confirm the requester's identity against the company directory before any other step.
  • Match the request to a standard role in the access policy, or mark it as an exception.
  • Check for conflicts, such as a role that lets one person both enter and approve the same payments.
  • Draft the grant as a proposal with the system, role, scope, duration and named approver.
  • Route the proposal to the approver and wait. Silence is not approval, and the agent never provisions access because a deadline passed.
  • After approval, execute the grant only through a method the agent is permitted to use, then log the result.
  • If execution fails, report the error to the helpdesk queue and stop. The agent does not retry with broader rights to finish the task.

Where a system uses delegated authorization, the agent should request the narrowest scope the task needs. Google's OAuth 2.0 documentation explains how scopes limit what an application can do on a user's behalf, which is a useful model even outside Google systems. For the wider question of who may approve what, the OWASP authorization guidance is a practical reference for checking that access decisions are enforced by the system, not only described in a document. The same split between a decision owner and an assistant that prepares the work appears in the onboarding workflow, where the agent prepares tasks while HR decides.

Offboarding checks: a list of places, not one switch

Offboarding usually fails in the gaps. The directory login is disabled, but the shared drive, the code repository, a delegated mailbox, the finance tool or a personal access token still works. The agent builds the list of places from the person's system inventory, and when the inventory does not cover a system the employee used, it flags that gap instead of assuming the account is clean.

A practical offboarding checklist for each leaving employee or contractor:

  • Directory account disabled and active sign-in sessions ended.
  • Mail forwarding, delegated mailboxes and shared inbox memberships removed.
  • Shared folders and group memberships reviewed, with file ownership moved to a named person.
  • Business tools checked for the user's role, approval rights and personal API tokens.
  • Company devices returned or wiped by IT, with the ticket reference recorded.
  • Service accounts or shared credentials the person knew flagged for rotation by their owner. The agent does not rotate shared secrets silently.
  • Scheduled reports, automations and pending approvals owned by the person reassigned or paused.

The last two items matter more than they look. An automation that still runs under a departed employee's account can keep producing output long after the person has left, and a pending approval can stall a payment or a purchase with nobody left to answer it.

What the record for each revoked account must show

Each revoked account needs a record that someone can audit later without reconstructing the case from memory. The record is the evidence that the offboarding happened, and it is the first thing to read when a former employee's access is questioned. Treating every automated step as a traceable signal is the basic idea behind observability, which OpenTelemetry describes in its introduction to observability concepts.

For each account, the record should hold:

  • System and account identifier.
  • Action taken and the exact scope removed.
  • The approver or the ticket that authorized the action.
  • The time of the action and the actor, either the agent or a named person.
  • The verification performed, such as a sign-in attempt that should now fail, approved by the system owner.
  • Open items, such as files awaiting transfer or a mailbox auto-reply that still names the person.

Verification deserves its own step. Removing a role in a settings screen is not proof that the account can no longer act. Where the system owner allows it, the agent can run a check that should fail and record the result. Where such a check is not allowed, the record says that the owner verified the removal, with a name and a time, rather than leaving the field blank.

Illustrative scenario and a decision guide for the agent

Illustrative scenario: a contractor's engagement ends on a Friday afternoon. The helpdesk agent receives the offboarding ticket and finds that the contractor holds access to a project folder, a design tool and a shared mailbox used by the client team. It drafts one removal per system, with the project lead as approver for the folder and the operations lead as approver for the tool and the mailbox. The shared mailbox has an auto-reply that names the contractor, so the agent flags it as an open item for the mailbox owner rather than editing it. The record shows the approved removals, a sign-in check that failed as expected, and one open item. No files were deleted, and ownership of the project files moved to the project lead.

A simple guide for deciding what the agent may do alone:

  • Read: directory entries, role lists, ticket history and system inventories, at all times.
  • Draft: grant proposals, removal plans and case notes, for review by a person.
  • Execute after approval: a removal or grant that a named approver has confirmed, in a system where the method is pre-approved.
  • Never alone: administrator rights, deletion of mailboxes or files, changes to shared secrets, or any action on a system the agent has not been permitted to use.

How Be Human runs this in managed operations

Be Human designs the helpdesk agent as a defined role, not as a general assistant with broad reach. The first engagement sets its memory (the access policy, the system inventory and the approver map), its tools (directory read, ticket write, and the removal actions approved for each system), its approval boundaries, and one observable success criterion, such as every offboarding case closing with a complete record. Managed operations then cover queue monitoring, recurring permission reviews, maintenance when a connected system changes its interface, and incident response when a removal fails or an account is found active after it should not be.

Before the agent proposes changes on live accounts, it can run in shadow mode, where it drafts every case and a person compares the draft with the decision actually made, as described in shadow mode for AI agents. A sensible first task is usually the offboarding checklist, because its end states are clear, its volume is repetitive, and a complete record is easy to check.

If you want to find the first task worth handing to an agent in your own operation, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent grant system access without a person approving it? No, and it should not be designed that way. In a sound setup the agent identifies the right role, checks for conflicts and prepares the grant, while a named approver confirms each one. The agent then executes only the approved change, through a method it is permitted to use, and logs the result.

What should an agent never do during offboarding? It should never delete mailboxes or files, grant administrator rights to itself or anyone else, or rotate shared credentials that other people depend on. Those actions are hard to reverse, so they stay with the owner of the system. The agent's role is to find the open items, prepare each removal and keep the record complete.

How do we know an offboarding is complete? Completion is defined by the checklist and the record, not by the agent's own report. Each system on the inventory should show either a completed removal with a verification result or an open item with a named owner. A case with open items stays open until those items are resolved or accepted in writing.

How do we protect the agent from instructions hidden in a request? Treat request text as data to classify, not as commands, and keep the agent's permissions fixed by policy rather than by what a message asks for. Limit what the agent can reach, require approval for every change, and review its logs on a regular schedule. The OWASP prompt injection guidance describes these separations in more detail.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →