AI-generated first lines mostly don't work. What does.
The problem isn't the AI. It's using it to generate the part of the email that sounds most obviously AI-generated.
The first-line personalisation pattern (pull a LinkedIn post, feed it to Claude, generate a compliment about their thought leadership) is detectable now. Enough people have received it that the “I loved your recent post about…” opener immediately signals automation, regardless of how well it’s written. The structure gives it away: compliment + bridge + ask is a pattern everyone recognises, whatever the content of the line.
What works better: use AI to find the specific, non-obvious detail that a human researcher would find if they spent 20 minutes on the account. Then use that detail to anchor a relevant question, not a compliment.
How do you personalise with tech stack and role data?
Open with the tech stack and role combination. “You’re using Salesforce and Outreach with a 12-person SDR team” is more useful than any compliment: it shows you know their context, and it is the setup for a highly relevant point instead of an opener that signals automation.
What company events give you a reason to reach out?
Anchor outreach timing to a recent company event: a funding round, a new product launch, a leadership hire, or a hiring spike in a specific function. These are factual, verifiable, and create natural reasons to reach out without manufacturing one.
What can you infer from a job title and company stage?
Name a specific operational problem implied by the job title and company stage. A VP of Sales at a Series B with 15 reps has predictable infrastructure gaps. Naming one of them specifically (without claiming to know for certain) is more engaging than generic benefits.
What is the AI’s job in the personalisation stack?
The AI’s job is to extract the useful signal from unstructured data (a job description, a news article, a LinkedIn profile) and format it as structured input that a template can use, rather than to write the personalised opener. Better research, simpler templates. The output reads more human because it’s based on real specifics, not generated warmth.