Use AI to research, not to write. Give it a company’s site and ask for the facts relevant to your specific offer, then write the line yourself. Asking it for the whole email produces fluent copy that reads exactly like the other AI emails in that inbox.
Published by emailcampaign.ai. Our product includes AI drafting, and this page argues for using it narrowly, which is worth weighing when reading it.
What it is actually good at
| Task | Suitability | Why |
|---|---|---|
| Summarising a company site into relevant facts | Strong | Reading and compression are what these models do best |
| Spotting a trigger in a job posting or announcement | Strong | Pattern recognition across unstructured text |
| Drafting three angles for you to choose between | Good | Useful raw material, with you as the filter |
| Tightening a message you wrote | Good | Editing beats generating |
| Writing the opening line unsupervised | Weak | It cannot judge whether the observation matters |
| Writing the whole email | Poor | Fluent, generic, and recognisable as AI |
The failure mode
The characteristic bad output is accurate and pointless. "I noticed Acme recently expanded into the Nordics" is true, retrievable, and tells the reader nothing about why you wrote. It signals that something automated read their website, which is the opposite of the impression personalisation is supposed to create.
Real personalisation connects an observation to a consequence for that person. The difference is not the fact, it is the inference drawn from it, and drawing the right inference requires knowing what your offer does and who feels the problem. That is the part a model is not positioned to supply.
So the division of labour is: the model finds candidate facts, and you decide which one implies something worth writing about.
The stylistic tells
Recipients now receive several AI-written cold emails a week and have learned the pattern. The most recognisable markers:
- Em dashes, which appear far more often in generated text than in ordinary business writing.
- Openers like "I hope this email finds you well" and "I noticed that".
- Uniform sentence length and rhythm across a whole message.
- A compliment that would apply to any company in the sector.
- Tricolons, where every list has exactly three items.
Any one is survivable. Three in a short message reads as automated, and the reply rate reflects it.
Verify anything concrete
Never let a model write a number, a customer name, or a claim about the recipient’s business without checking it. A confidently wrong detail about their own company is worse than no personalisation at all, because it proves nobody looked.
The same applies to claims about your own product. Fluent text is persuasive to the person reading it and to the person who generated it, which is exactly why unchecked specifics slip through.
It does not touch deliverability
This is worth stating plainly because the two get conflated. Whether your message reaches the inbox is decided by the sending domain’s history, whether SPF, DKIM and DMARC actually resolve rather than merely exist, the hard bounce rate, the complaint rate against Google’s published 0.1 percent threshold, and how many sends per mailbox per day you run.
A perfectly personalised message from a cold, unauthenticated domain still lands in spam. If placement is the problem, check authentication first; it takes about thirty seconds and frequently finds a duplicate SPF record or an include chain past ten lookups.
A workflow that holds up
- Model reads the company site, recent news and any job postings.
- It returns three candidate facts relevant to your specific offer.
- You pick the one with a real implication, or discard all three and skip the prospect.
- You write the opening line. The model may tighten it.
- You verify every concrete claim before it sends.
Step three is the one that matters. Being willing to drop a prospect because nothing relevant was found is what keeps the list narrow, and list quality moves reply rate far more than copy does.
Questions, answered straight
- How do I use AI to personalize cold emails?
- Use it for research rather than for writing whole messages. Feeding it a company's site and asking for the two or three facts most relevant to your offer is where it performs well. Asking it to write the full email produces fluent, generic copy that reads like every other AI email a prospect received that week.
- Does AI personalization actually improve reply rates?
- It improves them when it surfaces something specific and true that you would not have found manually at that volume. It reduces them when it produces observations that are accurate but irrelevant, which is the common output and reads as automation showing its working.
- Can recipients tell an email was written by AI?
- Increasingly, yes. The tells are stylistic: uniform sentence rhythm, em dashes, phrases like 'I noticed that' and 'I hope this finds you well', and a compliment that could apply to any company. Recipients now see many of these per week and pattern match quickly.
- Does AI personalization help deliverability?
- No. Placement is decided by the sending domain's history, whether authentication resolves, bounce rate, complaint rate and volume per mailbox. Content is the smallest factor, and varying it does not change any of the others.
- What should AI never write in a cold email?
- Anything factual it could invent: figures, customer names, claims about the recipient's business, or specifics about your own product. Verify every concrete claim before it goes out, because a confidently wrong detail about their company ends the conversation immediately.
Published 20 September 2026. Updated 20 September 2026. Written by the team that runs the infrastructure; numbers come from the platform's own provisioning and sending, and from the providers' published documentation at the time of writing.