Back to Blog Home

How to Roll Back AI Onboarding When a New Hire Does Not Start

Empty sofa with rumpled blankets in a quiet home with no one present

A new-hire no-show should cost you a short, checkable rollback, not a week of teams closing their own tickets. We recommend three design rules. Do not create anything before the start date whose undoing is expensive or impossible. Write what each action created, and how to undo it, before the action runs. And suspend a confirmed no-show instead of deleting it for a grace window, so that a late arrival is a re-enable and not a rebuild. Otherwise accounts stay active, a laptop sits in transit, and the company brain keeps assigning reading to someone who will not start.

Why the original event cannot undo itself

Onboarding automation runs one way, and a late change arrives after several systems have accepted the original event. Replacing start_date in the HR system may move a welcome email, but it does not show that a temporary password expired, a shipping label was intercepted, a payroll record was paused or a confidential onboarding page became inaccessible. Each system reports success for its own operation while the overall outcome is wrong: the identity provider knows an account exists, not whether the employment decision still applies, and the company brain knows reading was assigned, not whether the reader still has a relationship with the company.

A changed employment state has to trigger a new workflow and not merely edit the data that triggered the old one. Microsoft's lifecycle workflows model three phases, joiner, mover and leaver, and have none for someone who was provisioned and never arrived, so decide in advance which path a no-show takes. The same page lists leave of absence next to termination among the reasons to revoke access in a timely way, which is a fair precedent for suspending a delayed hire.

Make the rollback small: price every action before it runs

The cheapest rollback is the side effect you never created. List what your pre-start state allows and rank each action by what undoing it costs: a calendar invite and a disabled account are cheap, a laptop in transit may not be interceptable, a badge, once produced, has to be recovered or destroyed, and a payroll or tax record enters a file you may have to keep. We recommend scheduling the expensive actions as late as their lead time allows, gated on something the hire has done, such as confirming a delivery address or returning signed documents. Where you can, create accounts disabled and enable them on the start date, so that a no-show's account never worked at all.

Microsoft's compensating transaction pattern gives the principle: define clear points of no return, and design the workflow so that irreversible steps happen only after all critical validations succeed. Where pre-start tasks need sign-in, give the narrow access they need and nothing more.

Define the states

A single Boolean such as onboarding_active cannot tell a delay from a withdrawal, and those need different actions. Use a small authoritative state model:

  • planned: the offer is accepted, but no side effects are allowed yet.
  • approved_for_preboarding: named pre-start actions may run.
  • delayed: the person may still join, so reversible resources are suspended or re-dated.
  • withdrawn: the employment will not start, so resources are revoked and records handled under policy.
  • no_show_pending: the expected start passed, and an accountable HR owner has not yet confirmed the outcome.
  • no_show_confirmed: HR confirmed that the person did not start, and rollback may complete.
  • started: an authoritative employment event confirms the worker began.

Only the HR or recruiting system that owns the employment decision may move a person into withdrawn, no_show_confirmed or started. A manager's chat message can open an urgent case but must not become the authoritative event, so an automation mistake cannot remove access from a legitimate employee. Give every state change a monotonically increasing version and have consumers reject an older event that arrives after a newer one, so a delayed approved_for_preboarding message cannot recreate an account after withdrawn has been processed.

Write the ledger before the action

A rollback is only as complete as its list of what onboarding created. Write the ledger entry before each forward action runs, not after. If the process dies between creating an account and recording it, the account exists and no rollback can find it. Each entry holds the onboarding identifier, the lifecycle event version that authorized the action, the target system and resource, the accountable owner, and the compensating action with a flag for automatic, manual or impossible. The result and an evidence reference are filled in afterwards.

Tag every resource with the onboarding identifier wherever the target system has a tag or description field, and once a week compare the tagged resources in each system with the ledger. Anything tagged and missing from the ledger is an orphan. Microsoft's pattern page describes the same discipline: as the original operation runs, the system records information about each step and how to undo it.

The ledger records what to inspect and where to find proof, while the identity provider stays authoritative for account state and the carrier for shipment state. SCIM can carry part of the compensation, with a catch: under RFC 7644, DELETE is final from the client's side. The provider may keep the data but must answer 404 for the deleted resource and leave it out of query results. The reversible tool is the user's active attribute, which typically means suspended when false (RFC 7643), although the provider defines the exact meaning. Shipments, payroll cases, calendar events, knowledge assignments and building access sit outside the protocol altogether.

def on_lifecycle_change(person, event):
    if event.version <= person.last_processed_version:
        return  # stale or duplicate event
    freeze_new_onboarding_actions(person)

    if event.state == "delayed":
        suspend_access(person)
        redate_reversible_tasks(person, event.new_start_date)
        route_nonreversible_items_to_owners(person)
    elif event.state in ("withdrawn", "no_show_confirmed"):
        suspend_access(person)
        compensate_all_recorded_side_effects(person)
        verify_no_active_footprint(person)
    elif event.state == "started":
        require_hr_start_confirmation(person)
        resume_approved_actions(person)

    record_results_and_exceptions(person, event.version)

Make every operation idempotent: receiving the same withdrawn event twice must not create duplicate tickets, delete unrelated records or send another sensitive notification.

Freeze, suspend, then compensate

First stop new side effects: pause the person's onboarding orchestration, outbound messages, role assignments and company-brain task generation, because otherwise one integration may recreate a resource while another team removes it. Next suspend privileged access before any destructive cleanup, since suspension is fast, reversible and easier to verify across uncertain states. NIST SP 800-207 describes continual monitoring with possible reauthentication and reauthorization throughout user transactions, triggered by things such as time or anomalous activity (tenet 6). Do not rely on that to end an open session. Revoke it explicitly. Then compensate system by system:

  • Identity and applications: disable sign-in, revoke sessions, remove pre-start groups, invalidate one-time credentials and cancel pending invitations. Check applications that allow direct accounts as well as the identity provider, because a directory suspension does not prove that every software account follows it.
  • Company brain and onboarding knowledge: remove employee-specific assignments, revoke preboarding links, stop reminders and clear cached permission projections, but keep stable policy pages. The Kipwise employee onboarding workflow provides product context for assigned reading, searchable knowledge, and progress tracking. Rollback should reconcile those employee-visible tasks without deleting shared company knowledge.
  • Calendar, communications and people: cancel meetings, buddy introductions, check-ins and welcome messages, remove the person from private channels, and tell internal attendees only that the start is delayed or onboarding has stopped. Keep the reason out of calendar descriptions and chat, and keep sensitive recruiting context in its source system.
  • Equipment and physical access: ask the logistics owner to intercept an unshipped or in-transit package, and if interception fails, open a custody exception with a named owner and deadline. Cancel badge production where possible, revoke building authorization and check that temporary visitor credentials are inactive. An email to the carrier does not make equipment rollback complete.
  • Payroll, benefits and HR tasks: pause pending records and route them to the owners of those systems. Employment, tax, recruiting and security records may have separate retention requirements, so the rollback engine changes operational state and records the exception, and the record owner decides on deletion.

The GitLab onboarding handbook shows two backstops for the issue nobody closes: a Senior People Operations Specialist audits all open onboarding issues monthly, and an employment bot closes any onboarding issue still open after 60 days. Closing a ticket does not remove anything that was provisioned, so the ledger is still what tells you what is switched on.

Delayed, suspected and confirmed are different cases

A delayed start needs a controlled pause. Suspend access that is not needed during the delay, re-date tasks from the new authoritative start date, and reconfirm the manager, role, location, equipment destination and policy scope before resuming. Put an expiry on the delay: if HR has not confirmed a new date by then, the case goes to manual review, so nothing stays suspended while licenses, group membership or shipment custody go unchecked. A suspected no-show gets an immediate freeze and a wait for authoritative confirmation before anything destructive, which balances security against acting on a scheduling or communication error. A withdrawn hire gets the full compensating sequence.

Keep the way back open

People who miss a start date sometimes still arrive, a day late or after a mix-up about the date. If no_show_confirmed deleted the account and sent the laptop back, the company has to rebuild the account and re-ship the laptop for someone who is already there. We recommend a grace window, set by HR, in which a confirmed no-show stays suspended: sessions revoked, groups removed, shipments intercepted where possible, accounts disabled and not deleted. Deletion and record disposal start when the window closes, under the retention rules. Test the late arrival as a scenario in its own right, rollback followed by reinstatement from the same ledger, and not as an edge case.

Plan for partial rollback

Rollback fails in pieces: a carrier rejects an interception, an application is down, a direct account has no automated suspension. Record each exception with the resource, owner, next action, deadline and a safe interim control instead of a single rollback_complete: false, and keep the case open until every item is verified complete or accepted as a documented exception by an accountable owner. If access cannot be revoked, block it upstream and open a critical owner task. If an external system is unavailable, bound the retries and require manual confirmation after the deadline. If states conflict, keep the safer, suspended state until HR resolves the employment decision. If compensation could destroy a required record, restrict access and leave deletion to the record owner.

The company brain should never tell the former prospective employee that rollback is complete. It should stop answering through that identity. Internal owners can query a restricted status view that lists the safe state, unresolved exceptions and deadlines without exposing recruiting notes.

Prove the footprint is gone

Negative verification looks for state that should no longer exist, because completed rollback tasks alone prove nothing. For a confirmed withdrawal or no-show, check from the authoritative systems and not from the ledger alone that:

  • sign-in is blocked, sessions are revoked, and pre-start groups and direct application accounts are inactive;
  • company-brain assignments and employee-specific links no longer resolve;
  • welcome messages, reminders and future employee meetings are canceled;
  • badges and building permissions are inactive;
  • shipments are intercepted, returned, recovered or tracked as owned exceptions;
  • payroll, benefits and training work is paused or closed by the authoritative owner;
  • internal owners received a minimal status update;
  • every exception has an owner and a deadline.

Then replay the original hire event with its old version and confirm that consumers reject it, which catches the case where rollback succeeds once and a late message rebuilds the footprint. Test at least five scenarios before launch: a delayed start that later resumes, an offer withdrawal before access creation, a no-show after accounts and equipment exist, a partial rollback while one system is down, and a late arrival after no_show_confirmed. Each expected result names the final state and the remaining evidence.

Trace one recent hire

Take one recent onboarding record, ideally a remote employee who received pre-start knowledge and a laptop, and trace its identity, knowledge, calendar, equipment, physical-access and HR side effects. Any action without a compensating step or a verification query is the first control to fix. Do not launch until a simulated no-show shows that the person cannot use any pre-start access and that every item you cannot reverse automatically appears as an owned exception.

Topics

Related posts

Your team’s knowledge, answered in Slack

Add the Kipwise Agent. It answers from your pages with the source, and asks before it changes anything. 7-day trial, no card needed.