Is it acceptable to use AI to write cold email?

Writing it is not the problem; sending unverified claims is. Use AI for drafting and variation, keep a person responsible for the facts and the tone, and treat every generated detail as unverified until someone has checked the source.

Should I tell recipients that AI helped write the email?

There is no universal requirement to disclose a drafting tool, and a disclosure line in a sales email is usually noise. What does require care is automated decision-making that significantly affects a person, and claims presented as a human's personal judgement when nobody made one.

How many personal details per email?

One specific, checkable claim about their work, plus the problem it points to. Two can work if the second is about the industry rather than the person. Three reads as a dossier, which is the point where accurate research starts to feel like surveillance.

What about scraping profile data at scale?

Collecting what people publish is one thing; using it in ways they would not expect is another. Many jurisdictions require a lawful basis, a privacy notice and a way to object. Using a platform's data against its terms is a separate problem, and it is usually the sender who pays for it.

AI Personalisation That Reads Like Research, Not Surveillance

The short answer

Use what the recipient published about their work: their role, their company's announcements, a talk they gave, a problem their industry has. Leave out anything a stranger could not legitimately know, anything inferred about them personally, and anything that proves you read their private profile. Then review every first line before it sends.

AI made personalisation cheap, and the first thing that happened is that everyone received emails beginning with a compliment about their recent post. The technique stopped working because it stopped being evidence of attention. What still works is unglamorous: a first line that shows you understood the recipient's actual situation, drawn from sources they published themselves, and short enough that a human could have written it in two minutes.

The line is publication, not intimacy

A person's job title, their company's press release, a conference talk, a job posting, a product changelog, an industry regulation: all published, all fair to reference. You are allowed to notice what someone put in public, and referencing it is what a diligent salesperson has always done.

What crosses the line is anything that suggests you know more than they told the world: a reference to a private group, a detail from a personal social account, an inference about their health, finances, family or beliefs, or a comment on something they wrote in a space they consider semi-private. It is not about legality in the abstract; it is about the recipient's reaction, which is a report button.

Fair: role, company news, published writing, a public talk, a hiring post, a regulation affecting their sector.
Not fair: inferred personal circumstances, private-group content, family details, anything scraped from a closed space.
Never: a claim about their feelings or intentions that they never stated.

What AI is genuinely good at here, and what it is bad at

AI is good at turning a page of notes into one clear sentence, at matching tone across a hundred variants, at finding the specific detail in a long company announcement, and at rewriting the same idea for a different role. That is real labour removed.

AI is bad at knowing whether a detail is appropriate, at judging whether a source is public in the sense that matters to the recipient, and at noticing that a fact is stale. It will confidently reference a funding round that closed two years ago, because nothing in the text says it is out of date. The review step is not optional; it is the part that makes the rest usable.

The rule that keeps a campaign honest: one checkable claim

Each email should contain one specific claim about the recipient that they can verify in one second, because it is true and about their work. "You are hiring two support engineers" is checkable. "I loved your recent post" is not a claim about them, it is a claim about you.

One is enough. Emails that stack three observations read as a dossier, and a dossier is what makes people uncomfortable even when every fact is accurate. The second and third details should be about the problem, not about the person.

Where the review happens, and who owns it

An AI draft should be reviewed by a person before the first send of any new variant, and after any change to the prompt or the data source. After that, a sample from each batch is enough, provided someone is genuinely reading rather than clicking approve.

Ownership matters more than process. If nobody's name is on the review, the campaign has no owner and the first embarrassing line goes out at scale. One named reviewer, a sample size written down, and a rule that a bad line pauses the batch is the whole control.

Worked example: two opening lines for the same list

A payroll software company writes to operations managers at 60 mid-sized firms. The first AI variant opens with a compliment: "I loved your recent post about scaling operations." It is generic, it is unverifiable, and four recipients replied to ask which post. The tool had invented it to fill a variable.

The second variant uses one checkable fact: "You are advertising three operations roles in Leeds, which usually means the manual payroll run has stopped scaling." The claim is visible in the job posting, the inference is about the process rather than the person, and the next sentence offers a worked example of the same transition at a similar firm.

The second version produced replies, and the first produced a complaint. The lesson was not that AI writes badly; it was that a tool will fill a blank with something plausible unless the instruction is to leave the blank empty and cite the source.

What to personalise with, and what each source costs you

Source	What it gives you	Risk	How to use it
Their job title and team	The right problem to raise	Low	The opening sentence, no compliment attached
Company announcements	A specific, dated reason to write	Low, if the date is checked	Reference the change, not the achievement
A talk or article they published	Their own words on the problem	Low	Quote the idea, do not praise the person
A job posting	What they are building now	Low	Name the role, connect it to your offer
A personal social account	Details that feel private	High	Do not use it
Inferred attributes	Nothing defensible	High	Do not use it
A stale fact presented as current	Looks careless, not personal	Medium	Check dates before sending

The middle of this table is where most campaigns fail: nothing illegal, but the recipient can tell the difference between research and a guess.

Before an AI-personalised batch goes out

1Every personal detail in the draft comes from something the recipient published.
2No draft references personal circumstances, health, finances, family or beliefs.
3Each email carries one checkable claim, not three observations.
4Dates are checked: the announcement is current, not from two years ago.
5A named person has read a sample of this variant, and the sample size is written down.
6The unsubscribe and reply paths work, and someone reads the replies.
7The list itself is verified, so personalisation is not spent on dead mailboxes.
8A bad line pauses the batch rather than being noted for next time.

Where this connects on this site

For the discovery side rather than the drafting side, see finding business contacts on this site
Approval before sending is a campaign setting on run the batch with a review step
Personalisation is wasted on dead addresses; clean them with verify the list first

What this will not do

AI drafts can state facts confidently that are wrong or out of date. Every claim needs a check before it reaches a recipient.
Personalisation does not create consent, and using published information does not remove obligations under data protection law where they apply.
This page does not promise reply or conversion rates. It describes what makes an email defensible and readable.
No tool can tell you whether a detail will feel intrusive to a specific person. That judgement stays with the reviewer.
Nothing here is legal advice. Rules on business-to-business email, profiling and automated decision-making differ by jurisdiction.

Questions this page gets asked

Is it acceptable to use AI to write cold email?: Writing it is not the problem; sending unverified claims is. Use AI for drafting and variation, keep a person responsible for the facts and the tone, and treat every generated detail as unverified until someone has checked the source.
Should I tell recipients that AI helped write the email?: There is no universal requirement to disclose a drafting tool, and a disclosure line in a sales email is usually noise. What does require care is automated decision-making that significantly affects a person, and claims presented as a human's personal judgement when nobody made one.
How many personal details per email?: One specific, checkable claim about their work, plus the problem it points to. Two can work if the second is about the industry rather than the person. Three reads as a dossier, which is the point where accurate research starts to feel like surveillance.
What about scraping profile data at scale?: Collecting what people publish is one thing; using it in ways they would not expect is another. Many jurisdictions require a lawful basis, a privacy notice and a way to object. Using a platform's data against its terms is a separate problem, and it is usually the sender who pays for it.

Last reviewed 2026-10-26. Written for readers in the United States and elsewhere; rules about access, email and consent differ by country, and nothing on this page is legal advice.