AI in your outreach
Can AI write cold emails that do not sound like AI?
Yes, and not by asking it to sound more human. Almost nobody can reliably identify AI prose. What they can identify is the fourth email this month that opens by admiring their new website. The tell is sameness, and sameness comes from every draft being written off the same public page. Change the input and the problem disappears. Change the adjectives and it does not.
What people are actually noticing
They are not detecting a model. They are recognising a format.
Worth separating, because the two problems have different fixes and only one of them is about writing.
Ask someone why an email felt automated and they will point at a word. Press a little and the real answer comes out: they have seen this email before. Not this sentence, this shape. An opening compliment about the company, a pivot to a problem the sender has decided they have, an abstract benefit with no object attached to it, and a closing question nobody can answer in one word.
That shape is not a quirk of language models. It is what you get when the only thing the writer knows about the recipient is their public website, which is the one input every competitor running the same play also has. Ten senders, one source, one shape. The recipient is not detecting AI. They are noticing that they are on a list.
Which is why instructions about tone do so little. A warmer version of the same email from the same input is the same email. The thing to change is what the writer knows.
A test you can run before you send
The deletion test
One pass, thirty seconds, and it tells you whether you personalised the email or decorated it.
Delete every sentence that would still be true if you sent this email to a different company. Not every sentence that mentions them by name, which merge fields handle and which fools nobody. Every sentence whose content is actually about them.
Whatever survives is your personalisation. For most AI drafts, and plenty of human ones, nothing survives. That is the diagnosis, and the useful thing about finding it this way is that you find it before two hundred people do.
| The sentence | Survives? | Why |
|---|---|---|
| "I was impressed by what you are building at Acme." | No | True of every company on the list. The name is a merge field, not a fact. |
| "Most operations teams are struggling with manual processes." | No | Category language. It describes a market, not a reader. |
| "You posted two site supervisor roles in Geelong this month." | Yes | Checkable, specific, and false of almost everyone else. |
| "Jo Smith suggested I get in touch after we spoke last week." | Yes | Comes from your own records, which no competitor has. |
| "Would you be open to a quick chat?" | No, and that is fine | The ask is allowed to be generic. It is the only sentence that is. |
What to do instead
Give it a fact, not a personality
Every one of these is a change to the input or the structure. None of them is a change to the tone.
- One checkable detail, in the first two lines. A role they are hiring for, a site they have opened, a certification they have renewed, a thing they said. Specific enough that being wrong about it would be embarrassing, which is the test for whether it is real.
- One sentence of consequence. Why that detail means you are writing. This is the sentence almost every draft skips, and it is the one that converts a fact into a reason.
- An ask they can answer in one word. Not "would you be open to exploring". A question with a yes or a no in it, answerable from a phone, in traffic.
- Nothing about your product in email one. If the fact and the consequence are right, the product is the second email. If they are wrong, no description of the product rescues it.
- Feed it your own records, not the open web. Prior conversations, why the account is on your list, who referred them, what the last sequence found out. This is the only input your competitors cannot copy.
- Then let it be plain. Once there is a real fact in the email, plain language reads as confident. Without one, plain language reads as a form letter, which is where the instinct to add warmth comes from in the first place.
What not to do, because it keeps being suggested: deliberate typos, forced lower case, and "write like a human" in the prompt. The first two read as contempt once noticed. The third produces the most recognisable output there is, because warmth with no specific subject is flattery, and flattery is the opening everyone has learned to delete.
Where this gets easier
The input problem is a systems problem
Drafting is not the hard part. Knowing something nobody else knows is.
In Signal, ANI drafts a sequence step from scratch using your ideal customer profile, or polishes one you wrote, and you keep the final word on every step. The part that matters for this page is where it is drafting from: the system that already holds the account, the activity history and the reason the company turned up on your list, rather than a public page anyone can read.
Daily Spark matters here for the same reason. A company that arrived because it matched a description you wrote comes with the reason attached, and the reason is the first line of the email. A list bought from a database arrives with nothing, which is exactly when people start reaching for adjectives.
For the two adjacent questions: deliverability decides whether any of this is read at all and should be checked first, and reply rates sets what to expect once it is. How AI is changing cold email outreach covers the wider question of using it as a quality tool rather than a volume one.
Questions people actually ask
AI cold email voice, answered
Can AI write cold emails that do not sound like AI?
Yes, but not by asking it for a better tone. What readers recognise is not AI prose, it is a template: the fourth email this month that opens by admiring their new website. That happens because every AI draft is written from the same input, the prospect’s public site, which every competitor running the same play also has. Change the input and the problem goes away. Change the adjectives and it does not.
What actually gives an AI-written email away?
Sameness rather than syntax. In practice: an opening compliment about the company, a pivot to a problem the sender has decided they have, an abstract benefit with no object such as streamlining operations, and a closing question the recipient cannot answer in one word. Any one of those is survivable. All four in order is a format people have learned to delete on sight.
What is the deletion test?
Delete every sentence in the email that would still be true if you sent it to a different company. Whatever is left is your personalisation. For most AI drafts nothing is left, which tells you the email was decorated rather than personalised, and it tells you before you send it to two hundred people.
Should I tell the model to write casually or add typos?
No. Deliberate errors read as contempt once noticed, and an instruction to be warm produces the most recognisable output of all, because warmth without a specific subject is flattery. Ask for shorter instead: a specific fact, one sentence of consequence, and an ask the person can answer yes or no to.
Is it the copy or the deliverability?
Check deliverability first, because no amount of voice work fixes an email sitting in a spam folder, and the two failures look identical from the outside. If your replies fell off a cliff rather than drifting, it is almost certainly delivery. If they declined slowly as volume grew, it is the template. Cold email deliverability in 2026.
How long should a first email be?
Short enough that the specific detail is the first thing read. Length is not the variable people think it is: a long email with one checkable fact about the recipient outperforms a short one made entirely of category language. If you are cutting, cut the explanation of your product and keep the reason you wrote to them. What reply rates actually look like.
Can ANI write my sequence steps in Signal?
It can draft a step from scratch using your ideal customer profile, or polish something you have written, and you keep the final word on every one. The useful part is that it is drafting inside the system that holds the account, the prior conversations and the reason the company appeared on your list, rather than from a web page anybody can read. Sequences in Signal.
Does any of this change what I am allowed to send?
No. The rules on consent, identifying yourself and including a working unsubscribe apply the same way whether a person or a model wrote the words. Australia has no CAN-SPAM, so American advice about disclosure does not transfer. Is cold email legal in Australia?.
Question not answered here? Ask ANI and get a straight answer.