Blog

qAI agent for access provisioning

AI agent for access provisioning: who approves each grant

An AI agent for access provisioning prepares each grant from the ticket and your policy. The system owner approves it. What is refused, and what to check.

3 min readCredentials, Running it

by Grisha Pertaia

illustration
  1. session
  2. credential
  3. score
  4. intake
  5. sign
  6. sign
  7. file · redacted· client identifiers
  8. notify
  9. write

Access provisioning is the part of the service desk that looks easiest to hand to an agent. The request names a person and a role. The policy says which groups that role gets. The directory has a button for it. An AI agent for access provisioning can do all of that without a person touching the queue.

That ease is the reason to settle one question first: when an agent prepares a grant, who approves it, and what is the agent never allowed to grant? This post walks through the access request workflow on our IT and security page to answer it. The workflow is designed. It has not run in the sandbox yet, and the steps are how we would build it with you in a pilot.

Why one wrong grant is expensive

A wrong grant does not fail loudly. The account works, the person gets on with the job, and nothing looks broken. The cost arrives later. Every action that account takes is in question until someone checks it. If the grant had no end date, the rights stay after the project ends. An access review may find them, or may not.

The other costs are quieter still. A grant made from a comment instead of the request means the approval and the access no longer match. A contractor who keeps a group after leaving is an open door that nobody chose to leave open.

So the workflow is built around one rule: the agent grants only what the request names, the policy allows and the system owner approves.

The request, step by step

  1. Step 1

    A credential for one ticket

    The agent takes the request from the service desk queue. It receives a credential for this ticket only, and it signs in as itself, not as an administrator.

  2. Step 2

    The attachment read at the door

    The ticket carries an attachment. A line in it, in hidden text, asks for administrator rights. It is struck out before the agent reads it, and the strike is on the record.

  3. Step 3

    The requester looked up

    The agent reads the requester's role and current groups in the directory.

  4. Step 4

    The policy checked

    It checks the requested group against the access policy for that role, and notes the policy line it used.

  5. Step 5

    An extra group refused

    A comment on the ticket mentions a second group. It is not in the request, so it is outside the permission for this ticket. The step is refused.

  6. Step 6

    The owner approves

    The grant the request names waits for the system owner, who sees who asked, which role and which policy line.

  7. Step 7

    The ticket closed

    The agent makes the approved change and closes the ticket with what changed and who approved it.

For a contractor the access request workflow adds one more rule. Every grant carries an end date, and the removal is booked at the moment the grant is made. A request for "the same access as my colleague" is refused. Copying another person's groups hands over rights that nobody asked for and nobody approved.

The step a person keeps

The system owner approves the grant. That is the one step the workflow never skips in a pilot. The agent does the assembly: the lookup, the policy check, the draft change. The owner makes the decision with those facts on one screen, and the approval goes on the record with a name.

You choose which other steps wait. You may want a person on every grant to a system you mark sensitive, and on every removal from one. Ares does not decide which systems are sensitive. Your policy does, and the agent follows it.

Why the refusals hold

Step 5 is not the agent being careful. The credential for this ticket names one requester and one requested access. A second group was never in it, so the call fails whatever reason the agent had, and the refusal and its reason are written down. The same holds for step 2: even if the hidden line had reached the agent, the credential does not carry administrator rights. The guarantees page explains how a permission is issued for one job, narrowed and expired.

This is also why the agent never borrows a person's login or a shared service key. It signs in as itself, so the record names the agent on every line it wrote.

Your auditor can export the record and check it offline with the open verifier, without us. If a line was changed after it was written, the check fails.

What it does not do

  • Your identity provider and your directory stay in place; Ares sits on top of them. It connects to one identity provider product today.
  • Who should hold access is not the agent's call. That stays with your system owners and your written policy.
  • Access granted before the agent arrived is not checked by this workflow. Reviewing it is a separate job.
  • Inspection covers stated patterns and languages. An instruction outside them can reach the agent; the permission still limits what it can do.
  • The access request workflow is designed, not yet run in the sandbox. The closest run today is the requisition scenario, where a change outside the job is refused and a large order waits for a person.
Keep reading