What goes at the bottom of an agent's email

Nobody designs the bottom of an email, which is why it is where the mistakes accumulate: a three-image signature block, a legal disclaimer copied from 2009, no way to reach a person, and no indication that the sender is software.

3 August 2026 · 4 min read

An agent's footer has a real job, and it is short.

The three things it must do#

Say who is writing. A name that reads as an assistant, and the company. Not a fabricated human name, which is the one choice here with a genuinely bad failure mode: a recipient who discovers mid-thread that "Sarah from support" is software feels deceived, and says so publicly.

Offer a route to a person. One sentence, always true. "Reply with 'human' and a colleague will take over" works if your system honours it, and is worse than nothing if it does not.

Carry what the law requires. For cold sends, that means a physical postal address and a working opt-out. Both are obligations rather than preferences, per the compliance floor.

Ada, assistant at Northwind
Reply here and a colleague can take over any time.

That is a complete signature for a reply inside an existing thread. Cold sends carry the postal address and the unsubscribe link that MCPmailer attaches to sends classified as cold.

What does not belong#

• Images. A logo in the signature adds weight, breaks in text clients, and nudges a conversational message toward promotions.
• The 200-word confidentiality disclaimer. It has no legal effect on a recipient who never agreed to it, and it makes every message look like a form letter.
• Social links, awards, and taglines. This is correspondence, not a brochure.
• A no-reply address. On an agent that exists to hold conversations, it is self-defeating and a small negative signal.
• Tracking pixels by default. Off unless you have a specific use for the data, per do spam filters catch AI-written email.

Disclosure, and how much is enough#

The direction of travel is toward disclosure for systems interacting with people, and the practical case is stronger than the regulatory one: it performs about as well and removes the worst outcome. The bar is low. A name that reads as an assistant plus a route to a human is enough for most correspondence.

Where to be more explicit: first contact with someone who has not dealt with you before, anything a recipient might reasonably act on financially, and any thread where they ask directly. "Yes, I am an assistant. Would you like me to bring in a colleague?" is the whole answer, and it should never be evasive.

Unsubscribe placement#

For cold sends the unsubscribe must be findable rather than technically present. That means visible text near the bottom, not a light grey word inside a paragraph of legal text, and it must work on the first click.

The important half is what happens after: suppression is enforced at the infrastructure layer, so a suppressed address is refused on later sends rather than depending on an agent remembering, per bounces, complaints, and suppression. Your footer promises something; that enforcement is what keeps the promise.

Consistency across the fleet#

If you run several agents, the footers should look like they came from one company: same shape, same human route, different names. That sounds cosmetic and is not. A recipient who has dealt with your support agent and then hears from your billing agent should recognise the pattern immediately, which is also how they notice when something arrives that does not match, per when someone sends email pretending to be you.

Keep the footer in one place in your code rather than in each agent's prompt. A footer in a prompt drifts, gets truncated, and eventually one agent stops including the opt-out.

Questions#

Should an AI agent say it is an AI?

Yes. A name that reads as an assistant plus a route to a human is enough for ordinary correspondence, and it removes the worst outcome, which is a recipient discovering it mid-thread.

What must a cold send include?

A physical postal address and a working, findable opt-out, in addition to the identity and human route. Those are legal obligations rather than style choices.

Should signatures include a logo?

No. Images add weight, break in text clients, and push a conversational message toward the promotions category.

What about the confidentiality disclaimer?

Leave it out. It has no effect on a recipient who never agreed to it, and it makes every message read as a form letter.

Where should the footer live?

In your code, applied consistently, not in each agent's prompt. Prompt-based footers drift and eventually one agent stops including the opt-out.

Can the agent use a human-sounding name?

A first name is fine as long as it is presented as an assistant. Presenting software as a named colleague is the version that backfires.