Personalised Cold Email with AI: What Works and What to Avoid

Updated

Personalised cold email with AI works when the AI writes from facts you already hold about each person, such as their company, role and a note on why you're writing, and leaves the line blank when there's nothing specific. Never let it guess or invent details. Review every line before sending, and keep the rest of the email useful without the personal line.

AI has made it cheap to add a “personal” line to every email. It has also filled inboxes with openers like “I loved your recent post about innovation”, sent to people who never wrote one. Recipients notice. This guide covers what useful personalisation looks like, the data it needs, and how to stop an AI from inventing things in your name.

What good personalisation actually is

Personalisation isn’t a compliment or a mention of someone’s name. It’s the reason this email is for this person rather than everyone. A good personal line answers the reader’s first question: why are you writing to me?

Good personalisation is:

  • Specific. It refers to something true about the person or their business.
  • Relevant. That detail connects to why you’re writing.
  • Short. One sentence. The rest of the email does the work.

Compare:

Weak: Hi Tom, I hope this email finds you well. I came across Bramble Café and was really impressed.

Better: Hi Tom, you mentioned at the Harrogate food fair that Bramble Café was struggling to get sourdough delivered before 7am.

The second line gives Tom a reason to keep reading. The first could have been sent to anyone.

Which data columns to collect

The AI can only be as specific as your data. Before thinking about prompts, think about columns.

Column Example Useful for
first_name Tom Greeting
company Bramble Café Naming the business
role Owner Choosing the angle: cost for owners, quality for chefs
city or area Harrogate Local relevance, delivery areas
website bramblecafe.co.uk Your own research before sending
notes Struggles with sourdough deliveries before 7am The opening line
reason Met at Harrogate food fair, October Context for why you’re writing now

The notes and reason columns carry most of the value. If you can only fill those columns for some of your list, that’s fine; see the section on blank rows below.

For setting up the sheet itself, see our guide to mail merge from Google Sheets to Gmail.

The risk of invented facts

Language models generate text that sounds right. When asked for a personal opening line about someone and given only a name and company, a model has nothing true to say, so it may produce something plausible instead: a congratulations on an expansion that didn’t happen, praise for an article the person never wrote, a reference to an award they didn’t win.

In a cold email this is worse than no personalisation:

  • It’s a false statement sent under your name. The recipient knows immediately that it’s wrong, and the rest of your email loses credibility.
  • It shows the email was automated and not checked, which is the opposite of what personalisation is for.
  • It can create data protection problems. Under UK GDPR, personal data you hold has to be accurate. If invented details end up in your records or CRM, they are inaccurate personal data about a real person, and the ICO’s guidance on the accuracy principle expects you to correct or delete it.

How to prevent it

  1. Give the AI only your data. The instruction should be to write from the facts in the row and nothing else. No guessing about the company, no claims about recent news.
  2. Allow an empty result. If a row has no specific detail, the right output is nothing. A model that must always produce a line will fill the gap.
  3. Review every line before sending. Read each one and ask: is this true, and do I know it’s true?
  4. Don’t ask it to research. “Find something interesting about this company” invites the model to make things up if it can’t actually look anything up.

Replyable is built around these rules. Its AI writes a one-sentence opening line ({{personal_line}}) for each contact using only the facts in that contact’s row, such as company, role and notes. If the row has nothing specific, it leaves the line empty rather than inventing one. You then step through every contact, see exactly what they’ll receive, and edit any line before anything is sent.

Write the email so it works without the personal line

Some rows will have no notes. The email has to read naturally either way. Put the personal line first, as its own sentence, and make sure the next sentence doesn’t depend on it.

{{personal_line}}

We’re a bakery in Leeds supplying sourdough, pastries and cakes to independent cafés across West Yorkshire, delivered by 6.30am six days a week.

With a personal line, the email opens with Tom’s delivery problem. Without one, it opens with the second paragraph, which still makes sense.

Examples: from row to opening line

A recruiter

Row: first_name Priya, company Northfield Health, role Head of Design, notes Hiring two product designers on LinkedIn since September

Good: I saw Northfield Health has had two product designer roles open since September.

Invented: Congratulations on Northfield Health’s recent Series B funding.

Nothing in the row mentions funding. That line should never reach Priya.

A design agency

Row: first_name Alex, company Harlow & Co, role Marketing manager, notes Site not mobile-friendly; booking form breaks on iPhone

Good: I tried booking a table on the Harlow & Co site from my phone and the form wouldn’t submit on iPhone.

Note that “I tried booking” is only acceptable if someone at the agency actually did. If the note came from a colleague, write “We noticed the booking form doesn’t submit on iPhone.” The AI shouldn’t claim actions you didn’t take, and you should check for that in review.

A UK bakery

Row: first_name Tom, company Bramble Café, notes empty

Good: (no personal line; the email starts with the standard opening)

Invented: I love what you’re doing with Bramble Café’s seasonal menu.

An empty line is the correct output here.

Use AI for the draft, too, but in your voice

AI is also useful for the main email and the follow-ups, especially if you give it a brief: who the list is, what you’re offering, and what you’d like them to do. Generic AI phrasing is easy to spot (“I hope this finds you well”, “I wanted to reach out”), so give it examples of how you write. In Replyable you can paste two or three of your own sent emails in Settings, and drafts follow your voice. The same tool can write automatic follow-up emails from your first email.

Personal, plain emails also tend to do better in the inbox. Google’s sender guidelines ask for accurate subject lines and clear sender information, and short emails that read like normal correspondence avoid many of the patterns associated with bulk mail. Our guide on how to stop bulk emails going to spam covers the rest.

How much does personalised cold email with AI cost?

In Replyable, one personal line or one email draft uses one AI credit. The Free plan includes 50 credits a month, Starter 1,000 and Pro 5,000, so a 300-contact campaign with personal lines needs at least the Starter plan. Plans are compared on the pricing page.

On data handling: email briefs, voice samples and contact rows are sent to Anthropic’s API to generate the text, and aren’t used to train models under Anthropic’s commercial terms. If your notes column contains anything sensitive, take it out before you paste the list.

A short checklist

  • Collect a notes column with real, specific facts.
  • Restrict the AI to the facts in each row.
  • Accept blank lines where there’s nothing to say.
  • Write the email so it works without the personal line.
  • Read every line before sending.

If you’d like to try this on your own list, you can install Replyable and generate opening lines from inside Gmail.

Questions people ask

Can AI personalise cold emails at scale?

Yes, if each contact's row contains something specific to work from. AI can turn a note like "opened a second café in Leeds" into a natural opening line for hundreds of contacts, but it can't create relevance that isn't in your data.

Why do AI-written emails sometimes contain made-up facts?

Language models produce plausible text. Asked to write something personal about a person they know nothing about, they can fill the gap with a believable but false detail. Restricting the AI to the facts in your data, and allowing it to return nothing, prevents this.

What data do I need for AI email personalisation?

At minimum: email, first name and company. The column that makes the most difference is a short notes column with something you actually know, such as where you met or what they asked about.

How long should an AI-written opening line be?

One sentence. Its job is to show why you're writing to this person; the rest of the email explains the offer.

Sources