AI personalization in cold email: does it actually work?

There's a lot of hype around AI in cold email — and equally a lot of skepticism. Some people claim AI personalization is a game-changer; others argue that recipients can't tell the difference and that it doesn't move the needle on reply rates.

The honest answer: it depends entirely on how you do it. AI personalization done well can meaningfully lift reply rates. AI personalization done poorly produces emails that feel more robotic than a simple merge tag ever would.

Merge tags vs. AI personalization: what's the real difference?

Merge tags do find-and-replace. 'Hi {{First Name}}' becomes 'Hi Sarah'. That's useful for basic customization, but it doesn't require the email to be relevant to Sarah as a person — it just requires knowing her name.

AI personalization works differently. When Claude reads your contact data — name, company, role, industry, any custom notes in your sheet — it generates a unique opening that references something specific about that person. The output isn't a substitution. It's a sentence written for that individual.

That distinction matters because relevance is what actually drives replies. An email that references something genuinely specific about the recipient — their role, their company's recent announcement, a shared context — signals that you actually know who you're emailing. A name swap doesn't.

When AI personalization works well

• B2B outreach to senior buyers (VP and above) — where a generic pitch is most likely to be filtered or ignored. Senior buyers get a lot of email; specificity stands out.
• Smaller, high-quality lists where contact data is rich — the more context you give Claude, the better the output.
• Follow-up emails — a personalized second touch after no reply can re-engage prospects who missed or skimmed the first email.
• Any situation where standing out in a crowded inbox is the primary challenge.

When it doesn't help

• Thin contact data — if your sheet only has a name and email, Claude has little to work with and the output won't be meaningfully better than a merge tag.
• Very high-volume, broad lists — when you're blasting tens of thousands of loosely targeted contacts, the data quality typically isn't there to support genuine personalization.
• When the rest of the email is weak — a personalized opening can't rescue a bad pitch. It buys you a few extra seconds of attention; the value proposition still has to close.

What makes a good AI personalization

The quality of the output depends almost entirely on the data you feed in:

• More columns = better personalization. Company, role, industry, and a custom note column all give Claude more to work with.
• Specific beats vague. 'VP of Sales at Acme Corp (50-person SaaS)' gives Claude something to anchor on. 'Executive' does not.
• Custom notes are the highest-leverage column. A 'Context' field with values like 'Spoke at SaaStr about RevOps' or 'Their team just closed Series B' produces remarkably specific openings.
• Keep the system prompt simple. A clear instruction — 'write a 1–2 sentence opening that references something specific about this person, in a warm and direct tone' — outperforms over-engineered prompts.

How CMass handles it

CMass runs Claude (Anthropic's AI) on our infrastructure. You don't need a Claude account or API key — AI personalization is included in the Pro plan. Your contact data is used only to generate the email and is not stored beyond that step or used to train AI models.