Introduction
Yes, a general AI assistant such as ChatGPT can analyse a DMARC aggregate report usefully: it can explain the fields, spot alignment problems and suggest what to check next. It cannot safely run DMARC for you. Reliable intake, history, tenant isolation and enforcement decisions need a deterministic platform, with a person approving every DNS and policy change.
This article is published by DMARC AI, a DMARC platform provider for MSPs. The evaluation criteria are designed to be useful when comparing our platform, another vendor, open-source tooling, or an internal build.
This article is for MSP engineers and IT administrators who already paste the odd report into an AI chat and want to know where the line is. For the structure of the reports themselves, start with DMARC aggregate reports explained; for a field-by-field walkthrough, see how to read DMARC XML reports. The protocol is covered in the DMARC guide.
Key takeaways
- One-off use (explaining a report, drafting a client summary) is a good fit for AI. Continuous operations are not.
- Aggregate reports contain no message content, but they do reveal which services send for a client, which is confidential.
- Failure reports can contain message headers or content and should not go to an external AI model.
- Parse reports deterministically first, and give AI structured data, not raw attachments.
- AI may recommend; a person approves every DNS change and every move to
p=quarantineorp=reject. - DMARC tools with a built-in AI assistant exist: DMARC AI's assistant works on your parsed report data and live DNS, inside the platform's controlled pipeline.
Is there a DMARC tool with an AI assistant that helps fix issues?
Yes. DMARC AI includes an AI assistant that works on the platform's parsed aggregate report data for your domains and can look up live DNS records (SPF, DKIM, DMARC, MX). That lets it answer the questions an MSP engineer would otherwise take to a DMARC specialist: which senders fail and why, which record change fixes an alignment problem, and whether a domain is ready for the next policy step.
The difference from pasting a report into a general chat is the context and the controls around it:
| General AI chat | AI assistant inside a DMARC platform | |
|---|---|---|
| Data it sees | Whatever you paste in | Parsed, validated aggregate data for your tenant, with history |
| DNS | Cannot check records | Looks up the domain's live records |
| Tenant isolation | None; clients mix in one history | Per-client access control |
| Failure reports | Risky if pasted | Never used for AI processing at DMARC AI |
| Changes | Up to you | Recommended by the assistant, approved and published by a person |
The rest of this article explains why that structure matters, and how to use any AI tool, built-in or general, safely with DMARC data.
What are the two kinds of AI use?
The difference that matters is between asking AI a question and asking AI to run a process.
One-off, copy and paste use:
- explaining a single XML report;
- summarising unfamiliar fields;
- discussing likely causes of a failure;
- drafting a technical explanation or a client-friendly summary.
Operational, continuous use:
- receiving reports in mailboxes and processing every one reliably;
- normalising and deduplicating thousands of reports from dozens of receivers;
- keeping history and detecting changes over weeks and months;
- isolating one client's data from another's;
- raising alerts and running a workflow;
- making controlled enforcement decisions.
A general AI chat is good at the first list. It does not provide any of the infrastructure in the second list: no mailbox, no database, no schedule, no tenant model and no record of what was decided. That is the gap between an AI assistant and a reporting platform, and it is the same gap described in build vs buy for MSP DMARC platforms.
What can AI do well with DMARC data?
Given reliable, structured input, AI is a capable analyst. It can:
- Explain fields and patterns, such as why
dkimpassed inauth_resultsbut failed inpolicy_evaluated(the signing domain did not align). - Interpret SPF, DKIM and alignment problems, for example an envelope domain belonging to an email service rather than to the client.
- Suggest hypotheses and next checks: "this IP range looks like a CRM platform; confirm whether the client uses one and whether DKIM is set up for it".
- Summarise data across many rows into a readable paragraph.
- Draft reports in two registers: one for the engineer, one for the client's managing director.
- Prioritise changes, such as a new sender appearing or a pass rate dropping, when the input already contains trustworthy history.
What must AI never decide on its own?
Some decisions carry too much risk to hand to a model without human approval. AI should not:
- classify an unknown sender as malicious with finality;
- remove a legitimate sender from SPF or stop it signing;
- change DNS records without validation;
- move a domain to
p=quarantineorp=rejectautomatically; - treat results as complete when reports from some receivers are missing;
- send personal data or failure-report content to an external model without proper safeguards.
Each of these can stop a client's legitimate mail. RFC 9989 itself recommends a cautious, staged path: Section 7.4 says domains whose users post to mailing lists should spend time at p=none and then p=quarantine before considering p=reject. A model that cannot see the client's mailing lists, forwarding rules or gateway cannot judge that on its own.
What are the technical limits and risks?
The risks fall into four groups.
Accuracy limits. Language models can produce confident wrong answers (hallucinations), and the same prompt can produce different answers on different days. Neither is acceptable for a decision you need to repeat and defend.
Limits of the data itself. Aggregate reports are summaries: a source IP, a message count and authentication results. They do not show the message, the path it took or why it was sent. Forwarding, mailing lists, security gateways and ARC all produce failures that look like abuse but are not. Receivers also report differently, and some do not report at all, so a quiet report is not proof of a clean domain. Why DMARC reports are not always perfect covers these gaps.
Untrusted input. The rua address is published in DNS, so anyone can send a file to it. A fabricated "report" can carry text written to steer an AI model (prompt injection), and free-text fields are written by whoever generated the report. Treat report content as data, never as instructions.
Privacy and governance. Check the AI provider's data processing agreement, retention period and whether inputs are used for model training. A general chat has no tenant isolation, so client data can mix in one conversation history, and it leaves no reproducible audit trail of the input, the model version and the answer.
Do DMARC reports contain personal data?
Aggregate reports contain no message bodies, subjects or recipient addresses. RFC 9990, Section 7.2 states that no personal information such as individual mail addresses or message content is included. They do contain source IP addresses, counts, domains and DKIM selectors, which together show which services send mail for a client. Treat that as confidential client information.
Failure reports are different. RFC 9991 notes they can contain the header fields or sometimes the entire content of a failed message, and says their generation raises significant privacy concerns. They should not go to a general AI model. DMARC AI never uses failure reports for AI processing.
What does a safe AI workflow look like?
A safe workflow uses AI in one bounded step, with deterministic processing before it and human approval after it:
- Parse and validate reports deterministically. Code, not a model, decodes the attachment and checks the XML against the schema.
- Normalise the data into a fixed schema: domain, date, source IP, count, results.
- Separate tenants and apply access control before any analysis happens.
- Remove or minimise sensitive data. Send aggregates, not raw files; never send failure-report content.
- Let AI classify or summarise within limited categories, for example "known service", "probably forwarding", "unknown, investigate".
- Show evidence and confidence with every recommendation, so the engineer can check it.
- Require human approval for any DNS or policy change.
- Log the input version, the output, the model version and the human decision.
What does this look like on a real report?
Here is a small, fictional aggregate report fragment using documentation addresses:
<record>
<row>
<source_ip>198.51.100.24</source_ip>
<count>412</count>
<policy_evaluated>
<disposition>none</disposition>
<dkim>fail</dkim>
<spf>fail</spf>
</policy_evaluated>
</row>
<identifiers>
<header_from>client.example</header_from>
<envelope_from>bounce.mailer.example</envelope_from>
</identifiers>
<auth_results>
<dkim>
<domain>mailer.example</domain>
<selector>s1</selector>
<result>pass</result>
</dkim>
<spf>
<domain>bounce.mailer.example</domain>
<result>pass</result>
</spf>
</auth_results>
</record>
What comes deterministically from the XML: 412 messages from 198.51.100.24 claimed to be from client.example. SPF passed for bounce.mailer.example and DKIM passed for mailer.example, but neither domain aligns with client.example, so DMARC failed. The policy was none, so nothing was blocked.
What AI can reasonably suggest: this looks like a legitimate third-party sending service that authenticates with its own domains. The likely fix is to set up DKIM signing with the client's domain in that service and, if supported, a custom return-path.
What the engineer must still check: whether the client actually uses that service, who owns the account, whether the IP belongs to it, and whether the volume matches what the client expects. Only the client can confirm the first point.
Why the report alone cannot justify enforcement: it covers one receiver and one day. A decision to move to p=quarantine needs weeks of data from all reporting receivers, every legitimate sender aligned, and the client's agreement. How to move from monitoring to enforcement safely sets out that evidence gate.
FAQ
Can I paste a DMARC XML report into ChatGPT?
You can, for a one-off explanation, if your provider's terms and your client agreement allow it. Remove or replace client domains and IPs first if you are unsure. Do not make it a routine for client data, and do not paste failure reports, which can contain message content.
Do aggregate reports contain email bodies or personal data?
No message bodies, subjects or recipient addresses: RFC 9990 excludes them. They do contain source IPs, domains, selectors and counts that reveal a client's mail services, so treat them as confidential business data rather than as harmless.
Can AI tell whether a sender is legitimate?
It can suggest a likely identity from an IP range, a DKIM domain or a naming pattern, and those suggestions are often helpful. It cannot know whether the client has a contract with that service. Legitimacy is a business fact that only the client can confirm.
Can AI safely move a domain to p=reject?
Not on its own. Moving to p=reject needs complete report coverage over time, every legitimate sender aligned, knowledge of forwarding and mailing-list use, and the client's sign-off. AI can prepare the evidence and draft the change request; a person must approve and publish it.
What should be removed before sharing report data with an AI provider?
Send structured aggregates rather than raw attachments. Consider replacing client domain names and IP addresses with placeholders, drop free-text fields such as contact details and comments, never include failure-report content, and keep each client's data in a separate request.
What is the difference between an AI assistant and an AI-powered DMARC platform?
An assistant answers questions about the data you give it. A platform receives every report, validates and stores it, isolates tenants, keeps history, raises alerts and logs decisions, and may use AI inside that controlled pipeline. The assistant helps an analyst; the platform is the operation.
Final thoughts
AI can be a valuable DMARC analyst, but it is not by itself a DMARC reporting and operations platform. Use it to explain, summarise and draft. Keep parsing deterministic, keep tenants apart, keep failure reports away from external models, and keep a person in charge of every change that can stop a client's mail. To see the platform side with your own data, the free trial processes one real domain end to end.