AI Field Manual
FM-IT-06 · Guide

IT & Systems Administration  ·  Intermediate

Build a change request with a real rollback plan

A network gateway is being replaced. The new hardware is configured, the VLANs are staged, and a maintenance window is approved. The plan says "Cut over the connections. Verify networking. Roll back if necessary." That isn't a complete change plan. It doesn't say how you'll know each function works, when to stop, what can be restored, or what happens if you lose remote access halfway through.

Status
Published
Reading time
About 16 minutes
You'll need
Current-state diagram and configuration, target design, dependencies, affected users, a maintenance window, validated backups, an authorized implementer, and an approver or observer where policy requires
Last reviewed
November 2026
Before you start

A backup isn't a rollback

AI can challenge a change request before the maintenance window, the way a skeptical reviewer would. It can expose hidden dependencies and missing tests. It can't certify that a rollback will work; someone has to prove the steps in the real environment.

A backup is a copy. A rollback is a tested sequence that restores acceptable service within an agreed recovery time. A real rollback plan answers:

  • Exactly what state will be restored: device, configuration version, connections, policy, software, data or routing?
  • What triggers the decision, and who's authorized to make it?
  • How much time is reserved to carry it out before the window ends?
  • Which paths stay available if the original management interface fails?
  • What evidence proves the previous state actually works again?
  • Which changes can't be undone, and need a compensating action instead?

If you can't answer those, the change may not be ready to schedule.

Step 1

Define the change as a controlled experiment

Write the desired result separately from the implementation:

  • Desired result: "The new gateway provides the same permitted user and service connectivity as the existing gateway, with validated segmentation, remote access and monitoring."
  • Implementation: "Move the designated uplinks from the old gateway to the new device in the approved sequence."

They aren't interchangeable. The gateway can boot and show green while client DHCP, DNS, inter-VLAN rules, VPN access, printing or security controls fail.

Record the starting conditions:

Element Current Target Evidence to compare
WAN Old gateway, working circuit New gateway, same approved routing objective External reachability and provider status
VLAN routing Existing routes and ACLs Matching allowed and denied paths Segment-by-segment tests
DHCP / DNS Existing leases and resolvers Intended equivalent Client lease and name-resolution checks
VPN Existing approved remote access Intended remote connectivity Authenticated remote test with an approved account
Management Known console or local path Documented access to the new device Out-of-band or on-site recovery path
Monitoring Current alerts and logs New device integrated into monitoring Test alerts, timestamps and dashboards

Keep any intended behavior change separate from feature parity. A migration isn't the time to quietly loosen a firewall rule or change guest access because it's simpler.

Step 2

List risk, scope and dependencies

Use a real risk matrix instead of "low/medium/high" with no reasoning.

Risk Condition Consequence Mitigation / owner
Losing management access Admin access depends on a path being migrated Can't finish or revert remotely Named on-site operator, console access, tested alternate management path
VLAN mismatch A port profile or trunk differs from what's expected Department devices isolated Preflight switch-port mapping and a staged test client
DHCP/DNS failure Client lease or resolver path changes Users have a link but no useful connectivity Lease capture and internal/external DNS checks
Security regression The new device's default rules are permissive Unapproved cross-segment access Test denied paths as well as allowed ones
Provider issue The new WAN handoff doesn't match the circuit Site-wide outage Confirm the handoff and provider escalation path before the window
Slow rollback Old cables or configuration misplaced Outage runs past the approved window Labeled cables, the old gateway staged, rollback time reserved

Testing a denied path is as important as testing an allowed one. A guest VLAN mustn't reach the management subnet just because internet access works.

Step 3

Ask AI to interrogate the plan

The change-request prompt
You are an independent reviewer of an IT infrastructure change. Your
job is to find missing assumptions, unsafe sequencing and weak tests.
You may draft documents but must not execute or approve the change.

OBJECTIVE: [measurable end state]
CURRENT STATE: [sanitized architecture and configuration summary]
TARGET STATE: [sanitized design]
SCOPE: [systems, users, sites, services]
DEPENDENCIES: [network, identity, DNS, vendors, third parties]
MAINTENANCE WINDOW: [start/end/time zone]
AVAILABLE RECOVERY: [backups, prior devices, physical/remote access]
IMPLEMENTERS / APPROVERS: [roles]

DELIVER:
1. Clear change summary, reason and expected user impact.
2. Assumptions requiring explicit validation before the window.
3. Stepwise preflight, implementation and verification checklists.
4. Positive and negative path validation matrix with expected result.
5. Stop conditions, go/no-go checkpoints and decision authority.
6. Rollback actions with time budget and proof-of-restoration checks.
7. Communication templates for start, delay, rollback and closure.
8. Post-change monitoring and evidence to archive.

RULES:
- Do not invent network values, device commands or provider support.
- Mark missing details as BLOCKERS rather than filling them with guesses.
- Prefer minimal change; do not bundle unrelated improvements.
- Separate read-only verification from state-changing steps.
- Identify changes that cannot be undone by restoring a config file.
- Never say "rollback verified" without test evidence I supplied.

A useful answer may be uncomfortable: "BLOCKER: you haven't shown a working out-of-band path. A failed remote migration could strand the team." That isn't obstruction. It's the point of a review.

Step 4

Worked example: replacing a gateway

Fictional site. A 90-person office is moving from Gateway A to Gateway B. It has employee, guest, voice and security-system networks. The current gateway stays physically available during the window. One on-site implementer and one remote verifier are assigned. The window is 19:00–21:00 ET. A configuration export exists but hasn't been restored in a test environment.

Change request, decision-ready version

  • Title: Replace the perimeter gateway while preserving approved routing and segmentation.
  • Business reason: The existing gateway is approaching end of support; the replacement improves maintainability and lifecycle position.
  • Scope: One site's WAN handoff, internal VLAN routing and firewall, VPN, monitoring and supported remote administration. No intentional change to user access policy. Endpoint configuration is out of scope unless listed.
  • User impact: Intermittent site connectivity during the scheduled period. Remote users who rely on this site may be affected. The exact duration depends on testing; don't promise an interruption of under five minutes.

Prerequisites, all complete before 19:00:

  1. Approved inventory and verified cable and port mapping, including the WAN and upstream core connections.
  2. An exported, versioned backup of the old configuration, stored somewhere reachable without the network being migrated.
  3. The new device staged with current approved firmware, and its configuration reviewed against the target design.
  4. The owner of each DHCP scope, DNS forwarder, static route, WAN assignment, VPN path and firewall segment recorded.
  5. Read-only preflight results from representative clients on each VLAN and from an approved remote VPN user.
  6. Tested physical or console access, and explicit agreement on who has rollback authority.
  7. Confirmed contacts for the provider, the switch engineer and the decision-maker.
  8. Monitoring and communication owners assigned, with draft announcements ready.
  9. Time reserved for rollback and post-rollback verification. No implementation steps scheduled into that reserve.

Go/no-go gate at 18:55. The change lead checks every prerequisite, especially access to the old device, physical cable labels, test clients, and a recovery route that doesn't depend on the new gateway. Any missing critical prerequisite means NO GO.

Implementation sequence (outline, not vendor-specific commands)

Phase Operator action Required evidence Stop / continue condition
0 · Baseline Capture reference state: routes, test results, DHCP leases and firewall policy version. A dated baseline attached to the change. No baseline means no go.
1 · Transition Move only the approved handoffs in the diagram, with on-site verification. Physical port and cable mapping, and the device's link state. Unexpected links or missing management access: stop.
2 · Control plane Verify WAN, routing, DNS, DHCP and device management. Screenshots or exports of the checks. Any critical failure: start the rollback clock immediately.
3 · Services Test representative employee, guest, voice and security-system paths. A pass/fail matrix with device, source and time. All required allow and deny checks must pass.
4 · Remote Validate approved VPN and monitoring from outside the site. The remote tester's records. A failed recovery path may require rollback.
5 · Soak Watch logs, errors and service health for an agreed period. A monitoring snapshot and ticket notes. Repeated faults or unresolved impact: make the rollback decision.

Validation matrix (illustrative; tailor it to the site's actual rules)

Source / test Expected result Evidence
Employee client → approved public site Allowed DNS resolution and a successful application connection
Employee client → internal business application Allowed An authenticated application transaction
Guest client → public internet Allowed within policy Lease, DNS and HTTPS test
Guest client → management subnet Denied A negative connection attempt, with source and destination documented
Security-device VLAN → approved controller Allowed Device heartbeat or a validated session
Security-device VLAN → employee file share Denied if policy requires A negative test and the firewall log
Approved remote client → internal target Allowed with the intended authentication VPN connection plus an application test
Monitoring system → gateway telemetry Allowed A fresh log, event or polling result

Don't invent rule allowances. Validate against the approved existing policy.

The rollback plan, the part that makes this a change request

  • Triggers: losing administrative control, unexpected broad production impact, failed critical segmentation tests, unrecoverable DHCP or DNS problems, or being unable to finish essential validation before the rollback decision point.
  • Decision authority: the named change lead, escalating to the incident commander if user impact is material.
  • Decision time: by 20:15 ET, stop expanding the work and decide whether to accept the validated target or revert. That keeps a nominal 45-minute recovery and verification reserve before the window closes at 21:00. The time budget is illustrative and must be proven against the real cabling and configuration.

Procedure:

  1. Announce the rollback and freeze any unrelated changes.
  2. The on-site operator returns the WAN and internal uplinks to their labeled original ports, following the archived diagram.
  3. Restore and power the known-good old gateway with its confirmed previous configuration. If its configuration was changed, restore the specific verified version through the documented management or console path.
  4. Confirm the expected link state, then rerun the baseline tests: WAN reachability, internal DNS and DHCP, an employee application, guest isolation, the security-device controller, VPN and monitoring.
  5. Check for leftover states that need manual cleanup, such as duplicate active DHCP servers, changed upstream routes, stale ARP caches or conflicting gateway addresses. Act only through approved steps.
  6. Record the actual restoration times and user confirmation. Tell staff the previous environment is restored and the migration has been deferred.
  7. Preserve logs and observations, and open a follow-up engineering task instead of retrying straight away without a new plan.

If the rollback itself fails: invoke the established incident escalation path, keep any working console session open, engage the network lead and the provider, and prioritize restoring essential access. "Reboot both gateways until it works" isn't a contingency plan.

Key limitation: a prior configuration export is only useful if it can be applied to a working device. If physical paths, identity dependencies, certificates, upstream routing or the management plane have changed, more than a configuration restore may be needed.

Step 5

Close the loop after implementation

The change isn't complete when the implementer says "looks good." It closes when the agreed acceptance criteria are met, monitoring has been reviewed, the help desk has current information, and the evidence is attached.

Change request template
Change ID / title / requester / implementation owner:
Business reason and success definition:
Scope and explicit out-of-scope items:
Current-state reference / target-state reference:
Affected users, services and vendors:
Risk assessment (likelihood, impact, basis, mitigation):
Maintenance window / time zone / communication owner:
Preflight checklist with evidence and sign-off:
Backup artifact and independent recovery access:
Implementation steps (owner, expected result, stop condition):
Validation matrix (source, target, expected allow/deny, result):
Rollback triggers / authority / decision deadline:
Rollback steps / expected duration / restoration checks:
Incident escalation if rollback fails:
User and leadership communications:
Monitoring / support observation period:
Actual outcome / deviations / unresolved issues:
Approval, implementer and independent reviewer sign-off:
AI assistance disclosure, if used:

What stays with you: the risk rating, the timing, the approval, the go/no-go call, and the decision to keep the new state or restore the old one.

Esc

Saved in this browser

Reading list

Contents

Reference

Keyboard shortcuts

/ or Ctrl K
Search the manual
↑ ↓
Move through search results and index entries
Enter
Open the selected result
← →
Switch discipline tabs; previous / next entry in a workflow card
Esc
Close any dialog
?
Show this list