AI assistant
It tells you what happened to your email. And how it knows.
Delivered, deferred, bounced, quarantined, or DNS to fix: the assistant explains it to users and administrators from system data, with the sources in view and the declaration that it is an AI.
- Metadata only, unless you ask
- Common questions, no AI model
- Always says it is an AI
In the webmail I only see metadata of your own mailbox. I read the content of an email only if you ask me for that email.
You pressed “Why didn't it arrive?”Why didn't the message “Invoice 118/2026” arrive?
Assistant checks
- Reading the message status (in progress)(done)
- Interpreting the error code (in progress)(done)
AI assistant 0.3 s
The recipient's server replied with a temporary error at 09:41: code 452, “mailbox full”. The message is queued and the platform retries on its own. The rejection is not final.
Sources: Tracking · 3 events
Answer computed from system data, without an AI model.
AI assistant: answers are generated automatically and may contain errors.
YouWhat should I tell Jane Rowe?
AI assistant
You can tell her the message is waiting because her mailbox is full: as soon as she frees up space, delivery resumes on its own within the retry window: no need to resend it straight away.
Sources: Tracking · 3 events
Pseudonymized data: addresses, names, IPs and domains were replaced with placeholders before reaching the model.
AI assistant: answers are generated automatically and may contain errors.
01 — Where it lives
In the webmail and in the panel. Never beyond your permissions.
In the webmail it answers only about your mailbox. In the panel it answers with the permissions of the administrator asking: it sees what they can manage, and nothing else. Either way, permissions are decided by the system, not by the model.
- The “Why didn't it arrive?” button sits in every sent message, next to the delivery status.
- In the panel it also explains customers' deliveries, quarantine and domain DNS: delivered, deferred, bounced, quarantined, records to fix.
- It reads the body of an email only if you ask it to, for that email. The reading stays in the audit log.
What it works on
Delivery log
Every stage of every message, inbound and outbound, for 90 days.
Antispam verdicts
The verdict and score of every checked message, and the reason for a quarantine.
DNS status
MX, SPF, DKIM, DMARC and MTA-STS for your domain, as the authoritative nameservers see them.
Quarantine
What is there, why it is there, and what it can propose to do.
The content of emails? Out of its sight, until you ask it, for that one email.
02 — No AI model
The fast lane, no AI involved.
The most common questions — why it didn't arrive, what's in quarantine, how the DNS is doing — get an answer computed from system data, with no AI model at all: about 0.3 seconds from the click, and no data leaving.
The question
Press “Why didn't it arrive?” on a sent message. Nothing to type: the question is already the right one, and it carries the message's identifier with it.
System data
The platform reads that message's tracking events, the antispam verdict and, when needed, the DNS check. No language model: only the data the system already has.
The answer, with sources
In about 0.3 seconds you get an explanation in plain words, with links to the data it was computed from and the note “Answer computed from system data, without an AI model”. No data has left the platform.
03 — Pseudonymisation
Before anything leaves, your data changes its name.
When a free-form question needs an external model, the text does not leave as it is. Look at what the model actually receives.
What you write
Hello, Jane Rowe never received the invoice I sent to jane.rowe@sample.example. The message had our IBAN GB29 NWBK 6016 1331 9268 19 and my mobile +44 20 7946 0000.
What the model receives
Hello, <PERSONA_1> never received the invoice I sent to <EMAIL_1>. The message had our IBAN <IBAN_1> and my mobile <TELEFONO_1>.
Addresses, names, domains, IPs, phone numbers, tax codes, VAT numbers and IBANs become placeholders and come back in plain text only in the answer you see.
The three ways it can answer
No model
For typed questions — delivery, quarantine, DNS — the answer is computed from system data. No data leaves the platform.
How it reads in the answerAnswer computed from system data, without an AI model.
Pseudonymised
For free-form questions: only the placeholders reach the external model. The real values come back in plain text in the answer you see.
How it reads in the answerPseudonymized data: addresses, names, IPs and domains were replaced with placeholders before reaching the model.
Plain text, only with double consent
The provider must be enabled by the platform administrator and you must turn the option on. It is never on by default.
How it reads in the answerPlain data: you turned on plain sending and the provider is enabled by the administrator.
04 — Actions
It proposes, it doesn't act.
When there is something to do, the assistant prepares a proposal with the real data and stops there. The next step is yours.
- An action, such as releasing a message from quarantine, is proposed and runs only if you confirm it.
- Permissions are decided by the system, never by the model: the tools already receive your permissions.
- Email content is treated as data, never as an instruction: the defence against prompt injection has been tested.
05 — Transparency
It's an AI, and it says so.
Every answer states that it was generated by an AI (AI Act, art. 50) and how your data was handled. Not a footnote: it is inside the answer.
“AI assistant: answers are generated automatically and may contain errors.”The real wording, at the bottom of every answer.
- Every answer says how the data was handled: pseudonymised, in plain text, or with no model.
- Sources are links to the real data — tracking, quarantine, DNS — generated by the system, not by the model.
- If an answer contains numbers the tools did not verify, it is flagged.
- Conversations are encrypted and deleted after 30 days.
- Neither questions nor answers end up in the logs.
- Purple is reserved for the assistant, in the app as on this site: you recognise it everywhere.
- The sources: links to the real data — tracking, quarantine, DNS — generated by the system.
- The “48 hours” come from no tool: the answer gets flagged.
- How the data was handled: pseudonymised, in plain text, or with no model.
- The declaration that it is an AI, in every answer.
Want to see it answer?
A walkthrough of the assistant in the webmail and the panel, with your questions about what it reads, what it doesn't, and how it says so.
Write to info@qmail365.com