Skip to main content
Systems Engineering

Surviving the Final Domain Controller Cutover

Retiring the last on-prem DC exposes hardcoded LDAP binds, NTLM handshakes, and RADIUS dependencies that Entra ID cloud-only cutover plans routinely miss.

KBY Technologies11 min read
Surviving the Final Domain Controller Cutover

The last domain controller in any estate is never actually “the last”. It is a single point of failure masking twenty years of undocumented dependencies: printers with hardcoded LDAP binds, HVAC controllers doing NTLM handshakes, and a finance application that has never heard of OAuth. When a CTO mandates a cloud-only cutover to Entra ID, the marketing narrative is “flip a switch and decommission the server”. The engineering reality is a distributed systems decoupling exercise with four distinct blast radii. Get any one of them wrong and you don’t get a support ticket — you get a 3am outage bridge with three hundred locked-out employees and a badge reader that no longer opens the front door.

This article ignores the Microsoft slide deck and focuses strictly on the protocol-level failure points that surface when you attempt an Entra ID cloud-only migration and pull the plug on the final on-premise Active Directory server. We are treating Active Directory not as a directory service, but as a legacy monolith with hundreds of tightly coupled consumers, and the cloud-only cutover as the strangler-fig pattern applied to identity infrastructure.

#Architectural Breakdown: The Four Blast Zones

Before any GPO is touched or a single server is powered down, you need a dependency map of everything that still talks LDAP, Kerberos, or NTLM on port 389, 88, or 445. Active Directory’s design assumes a chatty, always-on domain controller sitting on the same broadcast segment as its clients. Entra ID assumes the opposite: a stateless, internet-facing OAuth2/OIDC token issuer with no concept of a local network. The gap between these two models is where outages live.

#
1. Ghost LDAP Binds on Port 389

Every environment has hardware and software nobody remembers deploying — building management systems, legacy ERP connectors, print servers, and Java applications using raw LdapContext calls instead of a modern identity provider. These clients will happily bind to a domain controller right up until the object no longer resolves, at which point they fail silently or, worse, lock out service accounts through repeated retry storms.

#
2. The 802.1X/RADIUS Blackhole

Corporate Wi-Fi and VPN concentrators frequently depend on an on-premise Network Policy Server (NPS) performing PEAP-MSCHAPv2 authentication against AD. The moment the domain controller backing that NPS instance disappears, every laptop in the building drops off the corporate SSID simultaneously. This is arguably the single most catastrophic failure mode in a cloud-only cutover because it is instantaneous and affects the entire physical estate at once, not a gradual application-by-application degradation.

#
3. The NTLM/Kerberos Gap

Legacy thick-client applications that were never rearchitected for modern auth still expect Kerberos ticket exchange or NTLM challenge-response against a domain controller reachable on the local network. You cannot simply federate these applications to Entra ID because they were never written with a SAML or OIDC library. They need network-layer proximity to an authentication authority that no longer exists once the on-prem AD is decommissioned.

#
4. GPO to MDM Translation Drift

Group Policy Objects accumulate architectural debt over a decade of “quick fixes” — registry hacks, logon scripts, and drive mappings nobody dares remove. Attempting to “lift and shift” these GPOs into Intune configuration profiles verbatim just relocates the debt into a new platform with a different enforcement engine, and most of those registry paths don’t even exist under the MDM CSP model.

Surviving the Final Domain Controller Cutover architecture diagram 1

#Implementation Logic: Sequencing the Cloud-Only Cutover

A cloud-only cutover is not a single migration event; it is a sequenced decommissioning pipeline with rollback gates at every stage. The logic we enforce on every engagement follows this order, and skipping steps is the most common cause of a failed rollback:

  • Audit phase (2–4 weeks): Enable LDAP diagnostic logging on every remaining domain controller and correlate Event ID 2889 hits against your CMDB to identify every unauthenticated or clear-text bind.
  • Isolation phase: Reroute identified legacy binds to a dedicated, ring-fenced read-only domain controller (RODC) rather than the production DC, so you can shrink the writable tier without breaking the stragglers. An RODC still needs a writable DC to replicate from, so it cannot outlive the last writable DC; the stragglers must be fixed or retired before the final demotion.
  • Protocol substitution phase: Migrate 802.1X onto certificate-based authentication issued via Intune SCEP, eliminating the password-based PEAP-MSCHAPv2 dependency on AD. A RADIUS service (cloud-hosted or on-premises) is still required to validate the certificates.
  • Application wrapping phase: Place any unpatchable NTLM/Kerberos-dependent thick client behind Entra Private Access so access is gated by Entra ID and Conditional Access. Private Access does not replace the application’s own Kerberos or NTLM, and its Kerberos SSO still requires domain controllers, so these apps must be modernised or retired before the final DC goes.
  • Policy re-authoring phase: Rebuild GPOs as CIS-benchmarked Intune configuration profiles from first principles rather than importing ADMX templates wholesale.
  • Decommission gate: Only after 30 consecutive days with zero Event ID 2889 hits and zero NPS authentication attempts do you demote the final domain controller.

#Code & Configurations

#
Hunting Ghost LDAP Binds with Event ID 2889

Active Directory Domain Services logs Event ID 2889 whenever a client performs an unsigned or clear-text LDAP simple bind. This is your primary forensic tool for identifying legacy hardware before you decommission on-prem AD entirely. Run this against every remaining DC:

1# Enumerate unsigned/clear-text LDAP binds in the last 14 days
2$Servers = Get-ADDomainController -Filter *
3$Since = (Get-Date).AddDays(-14)
4
5foreach ($DC in $Servers) {
6 Get-WinEvent -ComputerName $DC.HostName -FilterHashtable @{
7 LogName = 'Directory Service'
8 Id = 2889
9 StartTime = $Since
10 } | Select-Object TimeCreated, MachineName,
11 @{N='ClientIP';E={$_.Properties[0].Value}},
12 @{N='BindUser';E={$_.Properties[1].Value}} |
13 Export-Csv -Path "C:\Audit\GhostBinds_$($DC.HostName).csv" -Append -NoTypeInformation
14}

Feed the resulting CSV into your CMDB join to identify which physical asset owns each IP before it goes dark permanently. Anything still appearing on day 13 of a 14-day window is a hard blocker on your decommission gate.

#
Killing the RADIUS Blackhole: SCEP Certificate-Based Auth via Intune

Rather than moving PEAP-MSCHAPv2 to a cloud RADIUS proxy that still checks passwords against a directory, the correct architectural pattern is eliminating password-based 802.1X entirely in favour of Certificate-Based Authentication (CBA) with EAP-TLS. Deploy a SCEP certificate profile through Microsoft Graph (beta) so devices present Intune-issued certificates. A RADIUS service that can validate them without a domain controller, cloud-hosted or on-premises, is still required, and NDES with an AD-integrated CA also depends on AD, so plan the CA (or Microsoft Cloud PKI) before the final DC goes:

1{
2 "@odata.type": "#microsoft.graph.windows81SCEPCertificateProfile",
3 "displayName": "CorpWiFi-CBA-Root",
4 "subjectNameFormat": "commonName",
5 "subjectAlternativeNameType": "emailAddress",
6 "keyUsage": "keyEncipherment,digitalSignature",
7 "keySize": "size2048",
8 "hashAlgorithm": "sha2",
9 "certificateValidityPeriodScale": "years",
10 "certificateValidityPeriodValue": 1,
11 "renewalThresholdPercentage": 20,
12 "scepServerUrls": [
13 "https://ndes.corp.example.com/certsrv/mscep/mscep.dll"
14 ],
15 "extendedKeyUsages": [
16 {"name": "Client Authentication", "objectIdentifier": "1.3.6.1.5.5.7.3.2"}
17 ],
18 "rootCertificate@odata.bind": "https://graph.microsoft.com/beta/deviceManagement/deviceConfigurations/{root-ca-id}"
19}

Pair this profile with a Wi-Fi configuration profile referencing EAP-TLS instead of PEAP-MSCHAPv2, and the RADIUS server validates the device certificate against the issuing CA’s chain instead of a password hash on a domain controller.

#
Translating GPO to CIS-Benchmarked Intune Profiles

Do not export ADMX registry values and re-inject them as OMA-URI custom profiles — that is a lift-and-shift, not a redesign. Instead map each legacy setting to its equivalent CIS Level 1 control and deploy through the settings catalog (the summary below is illustrative, not an import format):

Surviving the Final Domain Controller Cutover architecture diagram 2
1profileName: Windows11-CIS-L1-Baseline
2platform: windows10AndLater
3technologies: mdm
4settings:
5 - settingDefinitionId: device_vendor_msft_policy_config_deviceLock_devicepasswordenabled
6 value: true
7 - settingDefinitionId: device_vendor_msft_policy_config_smartscreen_enableappinstallcontrol
8 value: true
9 - settingDefinitionId: device_vendor_msft_policy_config_lanmanworkstation_enableinsecureguestlogons
10 value: false
11 - settingDefinitionId: device_vendor_msft_bitlocker_requiredeviceencryption
12 value: true
13assignments:
14 - target: allLicensedUsersAssignmentTarget

This settings-catalog approach also gives you granular reporting per device rather than the binary “applied/not applied” state GPO gives you, which is essential when you’re proving compliance ahead of the final decommission.

#Failure Modes and Edge Cases in a Cloud-Only Cutover

Every cloud-only cutover we’ve run into trouble has failed for one of the following reasons, none of which appear in a standard migration checklist:

  • Kerberos-constrained delegation chains: Applications using KCD to impersonate users against SQL Server backends break silently once the delegating service account’s DC is gone — there is no cloud equivalent of KCD, so this requires app-level rearchitecture, not a policy change.
  • SYSVOL-dependent logon scripts: Scripts referencing UNC paths to \\domain.local\NETLOGON fail with an opaque “network path not found” error long after GPOs have been migrated, because the share itself, not the policy, disappeared.
  • Time synchronisation drift: Kerberos has a 5-minute clock skew tolerance. Devices that lose their w32time hierarchy root (the PDC emulator) during decommission drift out of tolerance and start failing authentication with cryptic KRB_AP_ERR_SKEW errors.
  • Certificate Authority orphaning: If your on-prem CA was AD-integrated, decommissioning the last DC without first migrating the CA role leaves every SCEP and CBA certificate you issued during the migration unable to renew.
  • Hybrid Exchange hard dependency: Organisations still running Exchange hybrid for mail flow retain a hidden dependency on on-prem AD for the Exchange schema, regardless of how “cloud-only” the rest of the identity stack has become.

From an architectural standpoint, the safest mitigation is running the final domain controller in a network-isolated, change-frozen state for a minimum burn-in period rather than an abrupt shutdown. This gives you a rollback path measured in minutes rather than a full forest restore measured in days. Teams building resilient architectural patterns for identity infrastructure treat this burn-in phase as a mandatory circuit breaker, not an optional nicety.

#Scaling and Security Trade-offs

Once the cloud-only cutover is complete, you inherit a different risk profile, not a risk-free one. Weigh these trade-offs explicitly before you sign off the decommission:

  • Attack surface: Removing on-prem AD eliminates NTLM relay and Kerberoasting as viable attack vectors, but shifts risk to token theft and consent phishing against Entra ID — Conditional Access with token protection becomes non-negotiable.
  • Latency: Certificate-based 802.1X (EAP-TLS with SCEP-issued certificates) removes the password check against a domain controller but still needs a RADIUS round trip, so measure authentication times in your own estate rather than relying on a general figure.
  • Blast radius on outage: A misconfigured Conditional Access policy in Entra ID can lock out the entire tenant globally in seconds; a misconfigured GPO historically only affected a single OU. Cloud-only architectures trade slow, contained failures for fast, global ones.
  • Cost model: Server licensing and DC hardware refresh cycles disappear, but Intune and Entra P2 per-user licensing scales linearly with headcount rather than with server count — model this before promising CFO-level savings.
  • Recovery time objective: Rebuilding a compromised on-prem forest after ransomware can take days or weeks; rolling back a bad Conditional Access change in a cloud-only tenant takes minutes, provided break-glass accounts are correctly excluded from every policy.

The engineering discipline that separates a successful Entra ID cloud-only migration from an outage post-mortem is refusing to treat any of the four blast zones as optional. Every ghost LDAP bind you fail to reroute, every RADIUS dependency you fail to replace with CBA, and every GPO you lift-and-shift instead of redesigning becomes the reason the final domain controller can never actually be switched off. Microsoft’s Entra Private Access documentation covers how access to the legacy applications that remain can be brokered through Entra ID in the meantime.

Evidence trail

Sources and verification

Primary documentation and external technical references used in this article.

  1. 01Entra Private Access documentationlearn.microsoft.com

About this page

Credited to

Dates

Published

Last updated

Corrections

No corrections are listed for this page. Some review sweeps corrected pages without naming each one; those are in the corrections log.

Reader Interaction

Comments

Add a thoughtful note on Surviving the Final Domain Controller Cutover. Comments are checked for spam and held for moderation before appearing.

Loading comments...

Discover more

Was this useful?

Reports go to the site owner and are used to correct pages. Corrections are listed at /corrections.

Engineering insights, direct to you.

Receive the latest Systems Engineering tutorials, production guides, Engineering Labs and operational best practices.