Skip to content
BlueMarlin
All guides

Email

AI email support agent: how it works and where it stops

An AI agent can answer the email your customers send to an address like info@ without changing mail provider: you forward the address to the tool. It must only answer people, never automatic email, and it should say it is an AI, which the EU AI Act has required since 2 August 2026.

Published
November 16, 2026
Last reviewed
November 16, 2026
Next review
March 31, 2027
Current

In 30 seconds

  • Yes, an AI can answer your support mailbox, and you keep your mail provider. You forward an address such as info@ to the tool, and the agent reads what arrives there.
  • It only answers people. Automatic emails, bounces and mailing lists get no reply: the standard for automatic replies (RFC 3834) says so, and it is what prevents two machines writing to each other forever.
  • Replies should leave from your own domain. That takes a few DNS records (SPF and DKIM). Gmail and Yahoo require every sender to authenticate with at least one of the two.
  • The From of an email can be forged. The agent should only use a customer's data when the sender passed DMARC, the check that confirms the domain in the From.
  • It has to say it is an AI. Since 2 August 2026 the EU AI Act requires it, at the latest at the first interaction, unless it is obvious.
  • A person stays behind it. Complaints, negotiations, sensitive data and anything the agent does not know go to the team.

What does an AI agent do in a support mailbox, and where does it stop?

It reads each email a customer sends, looks for the answer in what your business has given it, and replies in the same thread. When it does not know, or the matter is not for a machine, it hands the email to a person.

What it does well is the repeated question: opening hours, prices, the state of an order, how to book, what documents to bring. It answers at any hour and in the customer's language, and it asks for what the team will need before a person steps in.

Where it stops:

  • It does not make things up. It should answer only with the information you gave it, and say so when the answer is not there.
  • It does not decide. A refund, a discount or the terms of a contract are a person's answer. The agent takes the request down.
  • It does not answer everything that arrives. A mailbox receives far more than customers' emails: receipts, newsletters, notifications, out-of-office replies. None of that should get an answer, as explained below.
  • It does not write first. Answering someone who wrote to you is service. A promotion by email needs consent, the same as on WhatsApp: the rules are in the guide to WhatsApp and the GDPR.

If you receive five emails a week, you do not need an agent: a person answers them faster than a tool is set up. It pays off when the same questions arrive every day, or when they arrive while nobody is there.

How does email reach the agent, and from which address does it reply?

Email reaches the agent by forwarding: your address stays where it is, and your mail provider sends a copy of each message to an address the tool gives you. Replies leave from the tool's address at first, and from your own domain once you add its DNS records.

Getting in. You do not move your domain or change provider. In Gmail, for example, you add the tool's address as a forwarding address, Gmail sends it a verification link, and from then on it forwards every new message except spam. The other mail providers have an equivalent setting.

Going out. There are two stages:

From the tool's shared address From your own domain
What the customer sees in From An address on the tool's domain An address on your domain, such as info@yourbusiness.com
What you set up Nothing A few DNS records the tool gives you: SPF and DKIM
Who the email is authenticated as The tool Your business
Good for Trying it out Day-to-day use

Why your own domain matters. Gmail and Yahoo set requirements for whoever sends them mail, enforced since 2024:

All senders Bulk senders
Gmail (personal accounts) SPF or DKIM SPF, DKIM and DMARC. A bulk sender sends close to 5,000 messages or more a day to Gmail accounts
Yahoo SPF or DKIM, at a minimum SPF and DKIM, plus a DMARC policy of at least p=none, with the From domain aligned with the SPF or the DKIM domain

Three terms, in one line each:

  • SPF is a DNS record listing the servers allowed to send mail for your domain.
  • DKIM is a signature on each message that the receiving server checks against a key published in your DNS.
  • DMARC passes when SPF or DKIM passes and, besides, the domain it passed for is the one in the From. That match is called alignment.

A support mailbox is far from being a bulk sender. Even so, Google recommends always setting up all three, and warns that messages that are not authenticated "might be marked as spam or rejected". A reply the customer never sees is worth nothing.

Which emails should an AI never answer?

Any email a person did not write. The standard for automatic replies, RFC 3834, starts from one rule: an automatic responder "MUST NOT blindly send a response for every message received".

What arrives Why it gets no reply
An automatic email: a receipt, a notification, an out-of-office reply It carries the Auto-Submitted header. RFC 3834 says not to answer a message that has it with any value other than "no"
A bounce ("your message could not be delivered") Its return address is empty. RFC 3834 forbids answering it: the reply would reach nobody
A mailing list or newsletter It carries List- headers or Precedence: list. RFC 3834 lets a responder ignore them
A no-reply address Nobody reads what you send to it
Spam Answering confirms your address exists. Gmail does not even forward it

The risk has a name: a mail loop. Your agent answers an out-of-office reply, the out-of-office reply answers your agent, and so on all night. That is why the same standard asks every automatic reply to carry a header of its own, Auto-Submitted: auto-replied, which tells the other side "a machine wrote this, do not answer it".

A good tool adds a second brake: a limit on how many replies it sends in one conversation. BlueMarlin's agent, for example, is off on every mailbox until a person switches it on, answers only emails written by a person, marks its replies with Auto-Submitted: auto-replied, and sends at most 3 replies an hour in one conversation, with a daily cap per mailbox. Past either limit, a person takes the email.

How does the agent know who is really writing?

It does not know from the From alone: an email's From can be forged. RFC 3834 already said it in 2004: "the vast majority of email is unauthenticated, and return addresses are easily forged".

On WhatsApp, a customer writes from their phone number. In email, anyone can write "From: your-customer@example.com". If the agent looked up that address among your customers and answered with their order, their appointment or their invoice, it would be handing a stranger someone else's data.

The check that exists is DMARC. An email passes it when the domain in the From authorised that message, through SPF or DKIM. The current standard (RFC 9989, of May 2026) is precise about what that proves: only that the use of the domain was authorised by its owner. It does not say the message is trustworthy, but it does say the address was not forged.

The practical rule:

  • DMARC passed: the agent can treat the sender as the customer with that address.
  • DMARC did not pass, or the domain publishes none: the agent can still answer general questions, such as opening hours or prices, but nothing from a customer's record.

This is how BlueMarlin does it: a contact's record is used only when the sender's DMARC passed. Bear in mind that forwarding has a cost here: Google warns that forwarded messages often fail SPF, so what usually survives the trip is the DKIM signature. An email from a domain that does not sign its mail may arrive unverified.

Do you have to say that an AI is answering?

Yes. Since 2 August 2026, Article 50 of the EU AI Act requires that people who interact with an AI system are informed of it, "unless this is obvious", and "at the latest at the time of the first interaction".

The duty falls on the provider of the system: it has to be built so that the notice is given. In an email nothing is obvious, since a well-written reply looks like a person's. The simple way to comply is a line in the first reply, or a signature that says who answers: "I am the AI assistant of (your business). If you would rather talk to a person, just say so."

Data protection is separate, and it is yours. What follows is general information, not legal advice:

  • Your business is the controller of the emails and of the data in them. The tool that reads and answers them is a processor: it works on your behalf.
  • You need a contract with the processor (Article 28 GDPR): the data processing agreement. Among other things, it obliges the tool to process the data only on your documented instructions.
  • Tell your customers in your privacy policy what you use their data for and who receives it (Article 13).
  • Health data is a special category (Article 9). If your mailbox receives it, decide what the agent may answer and what goes straight to a person.

How is email different from WhatsApp for an AI agent?

Email has no 24-hour window and no buttons, moves more slowly, and arrives wrapped in signatures and quoted text. The agent is the same; the manners are different.

WhatsApp Email
When you can reply Within 24 hours of the customer's last message; after that, only an approved template At any time
Options to choose from Buttons and lists A numbered list the customer answers in words
Rhythm Short messages, within minutes One complete answer; the customer may reply the next day
How a conversation is kept together By the phone number By the In-Reply-To and References headers, which every reply should carry
What comes with the text Little Signatures, legal footers and the quoted text of earlier messages
Who writes A phone number An address that can be forged

Three consequences for the agent:

  • One answer, complete. On WhatsApp it can ask and wait. By email, each round may cost a day, so it should answer everything in one message and ask for all it needs at once.
  • It reads only what is new. The signature and the quoted history are not part of the question.
  • It replies in the thread. A reply without In-Reply-To and References shows up as a new conversation in the customer's mail.

The rules of the other channel are in the guide to an AI agent for WhatsApp. With BlueMarlin, WhatsApp and email share one inbox and one agent: a mailbox is created by forwarding an address, and replies leave from your own domain once it is verified.

When does a person take over, and how do you measure what the AI resolved?

A person takes over when the customer asks, when the agent does not know, and when the matter is not for a machine. The measure that counts is how many conversations ended without a person, and whether the customer really got their answer.

Hand over to a person:

  • When the customer asks for it. Always, and without making them insist.
  • Complaints and negotiations. They need someone who can decide.
  • Decisions with consequences. The GDPR (Article 22) gives the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects.
  • When the limits are reached. If the agent has replied several times in an hour to the same person, something is not working.

On handing over, the agent should tell the customer that a person will answer, and leave the team a summary, so nobody has to ask again.

To measure, three figures a week are enough:

  • Resolved without a person. Separate those where the customer confirmed it from those where they simply did not write again: silence is not always a yes.
  • Handed to the team, and why. Those emails tell you what the agent still needs to know.
  • Unanswered, and why: automatic, unverified, over the limit.

Frequently asked questions

Do I have to change my email provider to use an AI agent?

No. You keep your address and your provider, and set up forwarding to the address the tool gives you. In Gmail it is a setting with a verification link; the other providers have an equivalent one. To send replies from your own domain, you add the DNS records the tool gives you.

Can the AI reply from my own address?

Yes, once your domain is verified with the tool, which means adding SPF and DKIM records to your DNS. Until then, replies leave from an address on the tool's domain. With your own domain, the customer sees your address and the email is authenticated as yours, which is what Gmail and Yahoo look at.

Will the AI answer newsletters and automatic notifications?

It should not. RFC 3834 says not to answer messages with the Auto-Submitted header, and lets a responder ignore mailing lists. A well-built agent answers only emails written by a person and marks its own replies with Auto-Submitted: auto-replied, so other systems do not answer them.

Do I have to tell customers that an AI is answering their email?

Yes. Since 2 August 2026 the EU AI Act requires people to be informed that they are interacting with an AI system, at the latest at the first interaction, unless it is obvious. The duty falls on the provider of the system. A line in the first reply or in the signature is enough.

Is it safe for an AI to give customer data by email?

Only when the sender is verified. An email's From can be forged, so the agent should use a customer's record only when the message passed DMARC. Otherwise it can answer general questions, and anything personal waits for a person.

Sources

Everything above was checked on 16 November 2026 against the standards of the IETF, the official documentation of Google and Yahoo, and EUR-Lex:

Change history

  • 16 November 2026: first version.

No contracts. No sales calls. Start free this afternoon.

Create your account, set up your agent, try it from your phone on our test number and connect your number when you’re ready. Move up when your business does.