Subject lines an AI agent should never use (and 15 it can)

An AI agent should never write a subject line that deceives (a fake reply or forward, a made up urgency), invents a fact about the person, or promises what the email does not contain. The US CAN-SPAM guide bans deceptive subject lines, and Backlinko and lemlist both found the best response at 36 to 50 characters. Plain, specific and short beats clever.

Subject lines are the part of a cold email that an AI is most tempted to overdo. It has read a million of them, and the loudest ones are the ones it remembers. A person reading the inbox sees a sender name and a subject, and decides in a second. This post lists what to block, why, and what to allow.

Why do subject lines need their own rules?

Because the subject is the only part of the email that is read by everyone, and the only part that can mislead a person before they have chosen to read. It is also the part most affected by a model's habit of copying what looks like it works.

The FTC's guide to the CAN-SPAM Act says the law bans deceptive subject lines, and that it makes no exception for business to business email. Other regimes, such as GDPR, add rules on personal data, not on subject lines specifically. For the full legal picture see is cold email legal. The short version for subject lines: do not make the person believe something that is not true in order to get the open.

Which subject lines should an AI never write?

Block these in the persona and the rules, not just in your head.

1. A fake reply or forward. "Re: our conversation" or "Fwd: the proposal" when there has been no conversation. It is the most common trick and the clearest deception.
2. Invented urgency. "Last chance", "Final notice", "Expires tonight" when nothing expires.
3. A made up connection. "Following up on our call" when no call took place, or "Quick question from your colleague" when no colleague asked.
4. An invented fact. A company size, a funding round or a tool they use, when the lead data does not say so. Sellinger's email rules forbid facts that do not come from the lead data, and the subject is no exception.
5. A promise the email does not keep. "Your free audit" when the email asks for a call.
6. A person's name in capitals or with a failed merge field. "Hi {first_name}" is the signature of a broken template.
7. Clickbait and emoji clusters. They trigger filters and read as bulk mail.
8. All caps and repeated punctuation. "URGENT!!!" belongs to another era of spam.
9. A subject that does not match the body. A mismatch is a deception of its own, however mild.
10. A mention of the person's employer's problems. Subjects such as "Is your sales team failing?" read as insults, not interest.

What do the studies say about length?

Two large sources agree on a range. Backlinko analysed 12 million outreach emails and found subject lines of 36 to 50 characters got the best response rate. lemlist's data on reply rates points to the same range. Neither tells you why, and both measure a mix of senders. Our own view is that short subjects are easier to read on a phone, where the inbox cuts off the rest, and that very short ones, one or two words, can look like internal mail.

We have no email subject line test of our own to add. Our case studies publish the results of campaigns, not of single subject lines, and we would rather say that than invent a number.

What makes a subject line good?

Three things, in this order.

• It is true. Everything in it can be checked against the email.
• It is specific. It names something the reader would recognise: their company, their market, the thing you want to ask about.
• It is plain. It reads like a note from a colleague, in lowercase if that is your style, without tricks.

An example of the plain style comes from our own hotel. The agent at Hotel Pirin sent short Bulgarian letters of three to six sentences with one offer and one question: reply with your dates and headcount. Of 5,499 B2B contacts reached by email, 163 replied (3.0%). The case study makes no claim that any single element caused the result, but the shape is the point: a letter a person would not mind receiving. See the Hotel Pirin case study.

What are 15 subject lines an AI can use?

These are plain and make no promise the email cannot keep. Adjust the nouns to your case.

1. A question about your dealer network
2. Group stays at your hotel, quick question
3. How do you handle inbound leads at [company]?
4. Seeing your hiring for a head of sales
5. Your post on pricing
6. An idea for [company]'s onboarding
7. Intro from [mutual contact name], only if true
8. Are you the right person for supplier questions?
9. Four short questions about installers
10. Quick one on your last launch
11. Thoughts on your new office in [city]
12. Can I ask how you find new clients?
13. Rooms for your team events
14. Booking link for next week, if useful
15. One question, no pitch

Lines 3 to 6, 10 and 11 point at something the agent must find in the lead data. If the data is not there, the agent should use a plain line without the specific. Line 7 may only be used when there truly is a mutual contact.

How do you enforce this?

Put the list above in the rules your agent follows, written as plain instructions: "Never write Re: or Fwd: in a subject. Never claim urgency. Never state a fact about the person that is not in the lead data." Our guide on writing rules your agent follows shows how, and the email rules we enforce shows what is already built in. Read the subjects of the first twenty emails before you let the agent send on its own.

What about follow up subject lines?

A follow up in the same thread usually keeps the original subject, with the mail client adding "Re:" because it really is a reply in the thread. That is not deceptive, since the earlier message exists. The problem is typing "Re:" yourself on a first email. For a follow up that changes the angle, a new subject is better than the old one: it tells the reader this is something different. Our follow up template shows one way to write it.

Keep the same rules. The follow up subject must not claim that the person has seen or promised something they have not.

A checklist

• [ ] The subject is true and the body keeps its promise.
• [ ] It is 36 to 50 characters or near it.
• [ ] No fake reply, forward or urgency.
• [ ] No fact about the person that is not in the lead data.
• [ ] No merge field can fail into a visible placeholder.
• [ ] I read the first twenty subjects before sending on.