Coordinating a departing employee's access review across HR and IT with AI agents
Agent Swarm can prepare a source-linked offboarding checklist for HR and IT, while people authorize and perform every access removal and ownership transfer.
Agent Swarm can prepare an offboarding access review for human owners. A configured shared-agent workflow collects the systems assigned to one departing employee, the named owner of each system and the transfer tasks still open, then puts a source-linked checklist in front of the people who act on it. People authorize the departure, perform every access removal and ownership transfer, and handle any message to the departing person or to an outside party.
The scope is one authorized departure: the entitlements that person already holds, the ownership of what they leave behind, and the evidence a person checks after the change. It is a different job from granting a current employee access to a system they asked for, described in employee access requests; from a customer asking what data the company holds about them; and from reviewing a prospective supplier's security evidence, covered by the supplier security evidence review guide.
When another tool fits better
An identity-management or identity-governance platform. If most systems sit behind one identity provider, if the company already deprovisions through SSO and SCIM provisioning, or if it needs continuous entitlement review, a platform built for identity work fits better. It reaches applications the way it was designed to.
A manual checklist. If the departing employee holds access to a handful of systems and one person knows each owner, a checklist in a document is the cheaper control.
Agent Swarm fits a different shape: the systems, the owners and the open transfer tasks sit in several places, and several people each hold a piece. The HR record lives in one system and the identity provider in another, while per-system access sits in separate admin consoles. Coordinating that means gathering what each source holds and putting one checklist in front of the people who can act. Choosing an AI operating model sets out where that work belongs.
An illustrative workflow
This is an illustrative adaptation, not a built-in offboarding feature, not a customer story and not a measured run. It needs connectors and checks your team builds and tests.
1. An approved departure notice triggers a scoped review. The review opens from a departure HR has already authorized and covers one person. Nothing is collected before that approval exists.
2. Configured sources collect the inventory. Scripts read the sources the team has connected: the HR departure record, the identity provider's accounts, the admin consoles that hold per-system access, and the ticket queue where transfer requests are tracked. Each script fetches what it is allowed to fetch and emits JSON, so no model chooses which records to open. Reports from multiple sources documents that pattern.
3. Deterministic checks flag what does not line up. Configured rules compare the sources: a system where the person is assigned but no current owner is named, a system in one source and missing from another, two sources that disagree about who holds access. These are code your team writes and reviews, not a judgment the agent makes.
4. An agent prepares a source-linked checklist. One row per system: the source and a link to that record, the named owner, the proposed action (revoke or transfer) and the status. It is a file, so every reviewer opens the version the gate message links to.
5. HR verifies the departure event. Someone in HR confirms the notice, the last working day and what the person held. The reason for the departure is not read and not needed.
6. IT and each system owner confirm the change. A human-in-the-loop node pauses the run and posts to Slack, then resumes only when a person responds. Each owner revokes access or transfers ownership in the system and records what they verified. A row closes with a note naming the check that was run, not only with a click.
Exceptions worth writing into the checks
The cases the first pass gets wrong need a named path before the workflow runs.
- Unknown ownership. A system has no current owner, or the listed owner has also left. Route it to the IT lead to assign one. The row stays open and the owner field is never filled by guessing.
- An unavailable system. An admin console the connector cannot reach, an expired credential, or a source returning nothing where records were expected. Mark the row unverified and keep it open, with a name for who can restore access. A missing answer is not a clean answer.
- Conflicting identity records. Two accounts for one person, or a name mismatch between the HR record and the identity provider. Stop that row and ask HR and IT to resolve it before any action.
Privacy and data boundaries
Collect only what the review needs. The system, the role, the owner and the open task. The reason for the departure and performance history stay out of both the collection and the checklist, because the review does not need them.
Restrict the checklist to the people who act on it. HR, IT and the named system owners. The governance guide covers roles and one named owner per workflow, plus its limit: RBAC gates requests made with user tokens; operator-key calls and agent calls are not affected.
Keep source links rather than copying sensitive records. A row points at the record in the system that owns it, so the checklist does not become a second, stale copy of the identity inventory.
Set a retention rule before the first run. Decide how long the completed checklist is kept, who can read it later and when it is deleted.
Keep connectors narrow and secrets server-side. Give each connector read access to what its part of the review needs. Scripts reach external systems through script connections, and credentials come from the credential broker, so the raw secret stays out of script source and arguments.
What stays human-owned
Authorizing the departure, every revocation, every ownership transfer and any message to the departing person or an outside party stay with people. The workflow assembles and tracks; it does not act on an account.
An approval gate records a person's decision. It does not establish that the checklist is correct, and it does not establish that access was removed. The evidence that a system no longer grants access is separate, and a named owner produces it in that system: a screenshot or the console's own confirmation. Write down which form counts per system before the first review, or rows close on a click and nothing else.
Escalation follows the same shape. When an owner does not answer, the gate needs a reminder then a named backup, and an explicit hold so the row stays open. Never proceed on silence.
What is documented and what needs custom setup
Documented and shipping today. Deterministic swarm-script nodes that fetch from sources and emit JSON, one custom skill per source, and agent synthesis into a report: Reports from multiple sources. A workflow node that blocks and resumes only when a person responds, plus the request-human-input tool for a decision outside a workflow: HITL gates. Server-side credentials for outbound calls: script connections and the scripts credential broker.
Custom setup the team builds. Every connector to an HRIS, an identity provider, a SaaS admin console or a ticketing system; the access-source scripts that read them; and the missing-information checks. Agent Swarm ships no native HRIS or identity-governance connectors, so none of this arrives switched on, and each connector needs its own permissions and its own owner. The platform does not discover accounts on its own: the inventory is as complete as the sources you connect.
Not claimed here. The workflow does not revoke access by itself, does not find every account and does not certify compliance with an access-control framework.
Building blocks
Each piece below ships in agent-swarm and is documented; pointing them at an offboarding review is the illustrative part.
- HITL gates: the blocking
human-in-the-loopnode, its branches, timeout and fallback. - Reports from multiple sources: deterministic collection that emits JSON, then agent synthesis.
- Script connections: register an API once, call it from a script with a typed client.
- Scripts credential broker: outbound calls without the raw secret in source or arguments.
- AI operations governance: a named owner per workflow, roles and approval thresholds.
- Choosing an AI operating model: which model fits the work.
Preparing contractor credential expiry checks for human review with AI agents
Agent Swarm assembles a contractor's credential records against one assignment's requirements, flags the gaps in a sourced packet, and a person decides.
Preparing workforce training completion evidence for human review with AI agents
Agent Swarm assembles one training requirement, the roster in scope and dated completion records, flags gaps in a source-linked review, and a named person decides.