schedule 14-min read

DMARC Change-Management Policy for MSP Portfolios

DMARC AI editorial team · Last updated

A DMARC change-management policy for MSPs: change categories, approval and evidence gates, freeze windows, TTL-aware rollback and audit trail.

01

Introduction

A DMARC change-management policy defines how an MSP plans, approves, executes, verifies and rolls back changes to client SPF, DKIM and DMARC records. It sorts changes into standard, normal and emergency categories, sets the evidence each needs, and keeps an audit trail, so dozens of small DNS edits across a portfolio never turn into a delivery incident.

Most of the work in an MSP DMARC practice is small: a new SaaS sender added to SPF, a DKIM key rotated, a policy stepped from p=quarantine to p=reject, a reporting address moved. Each change is low risk on its own. The risk lives in the volume, across many clients and many engineers. This guide gives you a policy you can adopt as it stands or adapt. For the protocol background, start with the DMARC guide; for the commercial side of running this as a service, see DMARC for MSPs.

02

Key takeaways

  • Change control covers anything a receiving mail server will observe: _dmarc, SPF, DKIM, BIMI, MTA-STS and TLS-RPT records.
  • Classify every change as standard (pre-approved), normal (peer review, sometimes client approval) or emergency (act now, review within a working day).
  • Policy steps follow p=none, then p=quarantine, then p=reject. RFC 9989 removed pct, so there is no percentage ramp to manage.
  • Rollback means re-publishing the previous TXT record, but DNS caching delays both the change and the revert, so plan TTLs before the change window.
  • A change record with before and after values, evidence, approvals and verification is the artefact that survives a client audit.
03

What counts as a DMARC change?

A change needs change control when it alters something a receiving mail server will see. That means any edit to a _dmarc TXT record, SPF record, DKIM selector, BIMI record, or MTA-STS and TLS-RPT records on a client domain. Reading reports, running diagnostics and editing internal runbooks are routine work, not changes.

In scope:

  • _dmarc records: policy (p, sp, np), alignment (adkim, aspf), the test flag t, and reporting addresses (rua, ruf).
  • SPF: adding or removing include: mechanisms or IP ranges, or changing a flattening setup.
  • DKIM: publishing a new selector, switching signing to it, retiring or revoking an old one.
  • BIMI, MTA-STS and TLS-RPT records.
  • Platform settings that change what gets published in client DNS.

Out of scope: reading aggregate reports, running a lookup with the DMARC checker or SPF analyzer, client conversations with no technical change attached, and documentation edits.

04

How should MSPs categorise DMARC changes?

Use three categories borrowed from standard IT change practice: standard changes are pre-approved and repeatable, normal changes need review before execution, and emergency changes are made immediately to stop harm and reviewed afterwards. The category decides who approves, what evidence is required and when the change may run.

Category Approval Typical examples
Standard Pre-approved template, engineer executes and logs Adding a sender from your approved list to SPF when the lookup count stays within the limit; scheduled DKIM rotation following the documented two-selector procedure; removing a leftover pct tag where the intended p= is already confirmed
Normal (peer review) Second engineer reviews before execution p=none to p=quarantine; adding or changing sp or np; changing rua to a new destination (including an RFC 9990 external authorisation record); DKIM key length change; SPF restructuring or flattening
Normal (peer review plus client approval) Peer review and written client approval p=quarantine to p=reject; removing a sender from SPF; any change during a client freeze window; strict alignment (adkim=s or aspf=s)
Emergency Act now, file the record immediately, review within one working day Reverting a policy step that is blocking legitimate mail; fixing an SPF permerror; revoking a compromised DKIM key by publishing an empty p=

A few of these deserve a note:

  • SPF lookups. RFC 7208, Section 4.6.4 limits evaluation to 10 DNS-querying terms (include, a, mx, ptr, exists and redirect). Exceeding it returns permerror, so an SPF addition is only "standard" when the analyser confirms headroom. See how to fix too many SPF lookups.
  • np and sp. Under RFC 9989, Section 4.7, sp applies to existing subdomains and np to non-existent ones. Tightening np is usually low impact; tightening sp can affect real subdomain senders, so both are normal changes. New DMARC tags: np, psd and t explains the fallback order.
  • Reporting destinations. If a client's rua points at a domain other than the client's own, the receiving domain must publish an authorisation record as described in RFC 9990, Section 4. Without it, receivers must ignore that address and the reports silently stop. The record lives at client.example._report._dmarc.msp.example with the value v=DMARC1, or a wildcard *._report._dmarc.msp.example covering every client.

When in doubt, classify higher. A misclassified change is a typical way a "routine" edit becomes an incident.

05

What evidence should each change require?

Every normal change needs evidence that the new state will not block legitimate mail, plus a named reviewer. Standard changes rely on the template having been proven once; their evidence is a lookup showing the record parses correctly afterwards. Policy steps need the most evidence, all drawn from aggregate reports.

Useful evidence by change type:

  • SPF or DKIM change: the sender's documented SPF include or DKIM selector from its own setup guide, a lookup count before and after, and a test message whose Authentication-Results header shows a pass (check it with the email header analyzer).
  • rua change: a DNS lookup proving the RFC 9990 authorisation record exists before the _dmarc edit, and confirmation of the first report received at the new destination.
  • Policy step: a defined observation period in aggregate reports at the current level, every known sender aligned on SPF or DKIM, and every remaining failing source identified as either spoofing or a known exception such as forwarding. How to move from monitoring to enforcement safely covers the evidence gates in detail.

Set your thresholds in the policy itself, not per engineer. A rationale such as "all twelve known senders aligned for the last 28 days of reports, no unexplained sources" is reviewable. "Looked fine" is not.

06

How do policy steps work without pct?

Policy steps now move between whole levels: p=none, then p=quarantine, then p=reject. RFC 9989 dropped the pct tag, so older runbooks that ramped through 25, 50 and 100 percent sampling values describe a mechanism modern receivers ignore. Each step is one change record, not four.

RFC 9989 adds an optional test flag instead. Its Section 4.7 defines t=y as a request that receivers apply the policy one level below the one published: p=reject; t=y is treated as quarantine, and p=quarantine; t=y as none. The default is t=n, t=y has no effect at p=none, and it does not change report generation. Aggregate reports under RFC 9990, Section 3.1.1.5, include a testing element showing the t value receivers saw, which is useful verification.

Treat t=y as a signal, not a safety net. It is a request, and a receiver that has not implemented RFC 9989 ignores unknown tags and applies p= as written. Many MSPs therefore keep the evidence gate at the lower level and skip the rehearsal, or use t=y only to announce intent while monitoring. If your runbooks still mention percentages, pct has been removed: staging enforcement after RFC 9989 explains how to rewrite them.

07

What does a worked change request look like?

Below is a complete normal change with client approval: stepping a client from quarantine to reject. The thresholds in it are examples an MSP might choose, not figures from any standard. Adapt them to your own policy, and keep the same structure for every change type.

CHANGE RECORD CR-2026-0412
Client:           Contoso Ltd
Domain:           contoso.example (organizational domain)
Record:           _dmarc.contoso.example TXT
Category:         Normal, peer review plus client approval
Change type:      Policy step, quarantine to reject

Before:
  v=DMARC1; p=quarantine; rua=mailto:contoso@rua.msp.example
After:
  v=DMARC1; p=reject; rua=mailto:contoso@rua.msp.example

Evidence (example thresholds set in our policy):
  - 30 days at p=quarantine with no t flag
  - All 9 known senders aligned (M365, CRM, invoicing, newsletter...)
  - Remaining failures: unknown sources only, reviewed as spoofing
  - No client reports of missing or quarantined legitimate mail
    in the last 14 days
  - SPF lookup count 7 of 10; DKIM 2048-bit on all senders

TTL plan:
  Current TTL 3600. Lowered to 300 on 2026-11-09 09:00 UTC,
  at least one full hour before the window. Restore to 3600
  after the rollback window closes.

Rollback plan:
  Re-publish the "Before" record exactly. Expect up to 300 s
  of cached p=reject after revert. Rejected mail is not
  recoverable; senders receive a bounce and must resend.

Approvals:
  Proposed:  A. Patel, 2026-11-06 14:10 UTC
  Peer:      R. Okafor, approved 2026-11-07 10:02 UTC
  Client:    IT manager, email approval 2026-11-08, ref CL-8841

Change window:   2026-11-10 10:00-12:00 UTC (Tuesday)
Rollback window: until 2026-11-12 17:00 UTC
Executed:        2026-11-10 10:14 UTC, A. Patel
Verification:    DNS lookup from two resolvers shows p=reject at
                 10:21 UTC; first aggregate reports with
                 policy_published p=reject received 2026-11-11
Outcome:         No legitimate-mail failures in rollback window.
                 TTL restored to 3600. Closed.

Three details make this record useful. The evidence is specific enough for a reviewer to check against the reports. The rollback plan is honest about what reverting cannot undo. And verification uses two independent sources: a live DNS lookup and the policy_published section of the next aggregate reports, which shows the policy receivers actually applied.

08

How does DNS caching affect rollback?

Rollback for DMARC is simple in principle: re-publish the previous TXT record. In practice, resolvers that cached the new record keep using it until its TTL expires, so a revert is not instant. RFC 2181, Section 8 defines the TTL as a maximum caching time, so plan TTLs before the change, not during an incident.

A practical TTL routine for policy steps and SPF edits:

  1. Check the current TTL with the DNS lookup tool.
  2. Lower it (for example to 300 seconds) at least one full current TTL before the change window, so the long-lived cached copies have expired.
  3. Make the change, verify it, and hold the short TTL through the rollback window.
  4. Restore the normal TTL when the change is accepted.

Also be clear about what rollback cannot fix. Mail rejected under p=reject has already bounced. Mail quarantined under p=quarantine sits in the recipient's junk folder or quarantine, which only the recipient's mail administrator can release. That is why rollback windows are a monitoring commitment, not just a plan on paper.

09

When should changes run, and when should they be frozen?

Run planned changes in a fixed weekday window when the engineer and the client contact are both reachable, typically Tuesday to Thursday mornings. Avoid Fridays, weekends and the day before public holidays. Freeze windows are periods when no normal changes run for a client unless it is an emergency.

Each client runbook should record:

  • Standing change windows, with client-specific exceptions (a weekly bulk newsletter, a month-end invoicing run).
  • Freeze windows: financial year-end, peak trading such as Black Friday, the client's own migrations, and the weeks around a mail platform move.
  • Scheduled changes: DKIM rotations, planned policy steps, certificate renewals for MTA-STS.
  • Recent changes: a rolling log, useful for incident triage ("when did we last touch this domain?") and for quarterly reviews.

Rollback windows scale with risk. A standard SPF addition might need a few working hours of attention; a step to p=reject deserves at least two full business days, so you see a complete daily reporting cycle. RFC 9990, Section 3.1.1.4, notes the reporting period typically covers a single UTC day, so the first reports confirming a change arrive the following day.

10

How should MSPs communicate DMARC changes to clients?

Tell clients before, during and after any change they might notice, in plain language. Before: what changes, why, when, and what they should watch for. After: confirmation it is done and the rollback window. Policy steps and sender removals need written approval; standard changes can be summarised in the monthly report.

A short pre-change notice works better than a technical one:

On Tuesday 10 November between 10:00 and 12:00 UTC we will move contoso.example to the strictest DMARC setting. Mail from all your known systems is passing authentication. If anyone reports missing outbound email until Thursday evening, contact us and we can reverse the change within minutes.

Never treat silence as approval. If a client does not answer a request for a p=reject step, the change waits. How MSPs can avoid breaking client email with DMARC covers the client-side checks that catch forgotten senders before they become an incident.

11

What should the audit trail contain?

The audit trail should let anyone reconstruct who changed what, when, why and with whose approval, for every client domain. Each record holds before and after values, category, evidence, approvals, timestamps in UTC, verification and outcome. Records are never edited after closure; follow-ups get a new record that references the original.

Properties that hold up under a client's ISO 27001 or SOC 2 review:

  • Immutable after closure, with corrections as linked records.
  • Searchable by client, domain, date, engineer and change type.
  • Linked to client approval, the email or ticket where the client agreed.
  • Linked to evidence, an export or screenshot of the aggregate report view used.
  • Emergency records reviewed within one working day, with the post-review notes attached.

If you use Autotask, DMARC AI notification emails can be routed into it, which gives detected problems a natural starting point for a change record. Wherever you keep records, keep them in one place per portfolio, not in individual inboxes.

12

How do you roll the policy out across your team?

Start small: one page of policy, a change-record template and three or four ticket templates for the changes you make most. Then review it every quarter against what actually happened. A short policy that people follow beats a long one nobody reads.

  1. List the change types you made in the last three months and assign each a category using the table above.
  2. Write the evidence thresholds for policy steps into the policy, labelled as your own.
  3. Build templates for new sender, DKIM rotation, policy step and emergency revert.
  4. Agree standard windows, and capture each client's freeze windows during onboarding (the DMARC onboarding checklist is a good place to add them).
  5. Have new engineers shadow normal changes before executing their own.
  6. Each quarter, sample ten closed records at random and check they are complete.
13

Frequently asked questions

Is t=y a replacement for pct?

No. pct applied the policy to a share of failing mail; t=y asks receivers to apply the policy one level lower to all failing mail. It is a request that receivers without RFC 9989 support will ignore. Use it to signal a transitional state, and rely on aggregate-report evidence at the lower level before each step.

Can an engineer peer-review their own change?

No. Peer review exists to catch what the proposer missed, and self-review keeps the same blind spots. In a small team, the reviewer can be anyone who did not draft the change, including a senior engineer from another client team. Record the reviewer's name and timestamp in the change record so the approval is auditable.

How long should the rollback window be for a step to p=reject?

Long enough to see at least one complete daily cycle of aggregate reports and normal business mail, which in practice means two full business days for most clients. Clients with weekly or monthly mail runs, such as invoicing, may need the window extended until that run has passed without failures.

Do we need change control for an rua address change?

Yes, as a normal change. If the new address is on another domain, the receiving domain must publish the RFC 9990 authorisation record first, or receivers ignore the address and reports stop. Verify the authorisation record before editing _dmarc, then confirm the first report arrives at the new destination.

What if a client wants changes faster than the policy allows?

Explain that the evidence gates and rollback windows protect their mail, not your process. Standard changes can already move quickly. If the client still wants to skip steps, record that the change ran at their direction, with their written instruction attached, so the audit trail reflects who accepted the risk.

14

Related articles

Related tools

Ready to Implement?

Get authenticated mail moving in minutes: start free, book a guided demo, or talk to the team about your stack.