Skip to content
← All posts

AI Agents

AI personalization in cold email: what works and what reads like a robot

How to use AI to personalize cold emails without sounding automated: levels of personalization, research inputs, prompts, guardrails, examples and QA.

Victor Mendonça
Victor Mendonça

Founder, OretonGTM · ·6 min read

AI made personalization cheap. Unfortunately, it also made bad personalization cheap. Inboxes now fill with emails that open with “I loved your recent post about leadership!” or “As a fellow innovator in the SaaS space…” followed by a generic pitch.

Buyers have learned to spot these instantly, and they now read as less personal than a plain, honest email.

AI personalization still works, but only when it’s done for the right reason: to make the email more relevant, not to make it look personalized. This guide covers what separates the two, how to set up the research and prompts, and how to keep quality high at scale.

Relevance vs. decoration

The key distinction:

  • Decorative personalization mentions something about the prospect that has nothing to do with why you’re writing. “Congrats on 5 years at Acme!” followed by a pitch for data tools.
  • Relevant personalization uses a fact about the prospect to explain why the email matters to them. “Acme is hiring three SDRs, so list building is about to become a bottleneck” followed by an offer about lists.
Decorative (robotic) Relevant
“Loved your post on leadership!” “Your post about ramping SDRs mentioned list quality; that’s exactly the problem we work on.”
“Impressive growth at Acme!” “With Acme’s sales team doubling this year, routing replies to the right rep gets harder fast.”
“Fellow Stanford alum here!” (Unless the email is actually about something related, skip it.)
“I see you’re the VP of Sales at Acme.” (They know. Skip it.)

The test: if you removed the personalized line, would the email lose meaning? If not, the line is decoration.

My take: The best personalized line is one where the prospect thinks “fair point” rather than “how did they know that?”

The levels of personalization

Level What changes When to use
Segment Problem, offer and proof fit everyone in a narrow segment Always; this is the foundation
Account One fact about the company connected to the problem When a relevant fact exists
Person One fact about the individual connected to the problem Tier A accounts, senior buyers

AI is most useful at the account level, where it can read websites, job posts and news faster than any human.

What makes a good personalization input

AI can only personalize from what you give it. Good inputs:

  • Job posts: what they’re hiring for and why (hiring signals)
  • Company website: what they sell, to whom, how they position themselves
  • Recent news: launches, funding, expansion, partnerships
  • Tech stack data: relevant tools in use (technographics)
  • LinkedIn profile: role, tenure, recent role change, relevant posts
  • Signals you track: new leader, funding, intent (signal-based outbound)

Weak inputs:

  • Generic company descriptions from databases
  • The model’s own knowledge of the company (often outdated or wrong)
  • Personal trivia unrelated to your offer

The workflow

Here’s the setup I use:

Step 1: Write the segment-level email first

The core message, structure and offer, written by a human. Leave one slot for an account-specific line. See how to write a cold email.

Step 2: Collect research

For each account, gather source content: website text, recent job posts, news. In Clay, this can be done with scraping and research columns. See AI research agents.

Step 3: Extract relevant facts

First ask the model to extract facts relevant to your problem area, with sources. Keep this separate from writing. It’s easier to check facts than prose.

Step 4: Write the line

Then ask the model to write one line connecting the best fact to your email’s problem, under strict rules.

Step 5: Review

Human review for Tier A; random sampling for the rest.

Step 6: Assemble and send

Insert the line into the segment template. If the model returned “NONE”, use a segment-level opener instead.

A prompt that works

Here’s an example for the line-writing step:

You write one opening line for a B2B cold email.

Email topic: helping B2B SaaS teams build verified lead lists so new SDRs
don't build their own.

Facts about the account (from research, with sources):
{facts}

Rules:
- Max 25 words.
- Use ONE specific fact from the list above. Do not invent facts.
- The line must connect that fact to the email topic.
- No flattery, no compliments, no exclamation marks.
- Do not start with "I noticed", "I saw", "I came across" or "I hope".
- Do not mention the person's title or company tenure.
- If no fact is relevant to the email topic, return exactly: NONE

Return only the line.

What makes it work:

  • A clear topic so the model knows what to connect to
  • Facts passed in, not left to the model’s memory
  • Banned phrases that signal automation
  • A length limit
  • An escape hatch (“NONE”) so the model doesn’t force a weak line

More prompts in cold email personalization prompts.

Examples: before and after

Fact: Acme posted three SDR roles last week; job post mentions “building our outbound motion from scratch.”

Robotic: “I noticed Acme is growing fast — impressive work!”

Relevant: “Building outbound from scratch with three new SDRs usually means someone has to own lists before anyone can send.”


Fact: Brightline recently announced expansion to the UK.

Robotic: “Congrats on the UK expansion! Exciting times!”

Relevant: “Expanding into the UK usually means building a new account list from zero, and most databases cover UK startups less well than US ones.”


Fact: No relevant fact found.

Robotic: “As a leader in the software space, Acme is surely focused on growth.”

Relevant: NONE → use the segment-level opener.

Guardrails against hallucinations

Models can invent facts: a funding round that didn’t happen, a product the company doesn’t sell. At scale, even a small error rate means many embarrassing emails.

Guardrails:

  1. Source-only instructions: “Use only the facts provided.”
  2. Separate extraction and writing, so facts can be checked.
  3. Require sources for each extracted fact.
  4. Allow “unknown” / “NONE”.
  5. Check dates: old news presented as recent is a common error.
  6. Sample review every batch.

See how to stop AI hallucinations in sales.

Quality control at scale

For each batch:

  • Review a random sample (say, 20 to 50 lines). Mark each as good, generic, wrong or awkward.
  • Track the NONE rate. Very low NONE rates may mean the model is forcing lines; very high may mean poor research inputs.
  • Track factual errors. Any error rate above a very small share is a problem.
  • Watch negative replies. “This is clearly automated” or corrections of wrong facts are red flags.

Measuring results

Personalization should improve outcomes, not just look nice. Test:

  • Variant A: segment-level email only
  • Variant B: same email with AI personalized opener

Compare positive reply rate and meetings per 1,000 contacts on enough volume (a few hundred contacts per variant). If B doesn’t clearly win, your personalization isn’t relevant enough, or your segment message was already doing the work. See A/B testing cold email.

When not to personalize

  • When there’s no relevant fact. Forced lines are worse than none.
  • When the segment message is already highly specific.
  • When the cost of research outweighs the value of the account.
  • When you can’t review quality, for sensitive or high-value audiences.

Common mistakes

  1. Flattery as personalization.
  2. Personal trivia unrelated to the offer.
  3. No source data, leading to generic or invented lines.
  4. No NONE option, forcing weak lines.
  5. Restating the obvious (“As VP of Sales at Acme…”).
  6. Never reviewing outputs.
  7. Personalizing a bad message. A personalized line can’t save a weak offer.

Where to start

Pick your best segment, write a strong segment-level email, add one research source (job posts are a great start), set up the extraction-and-writing prompts above, review 50 outputs, then A/B test against the segment-only version. If it wins, expand to more sources and segments. For the broader picture of AI in prospecting, see AI for sales prospecting.

Frequently asked questions

Does AI personalization improve cold email reply rates?

It can, when the personalization is relevant to the problem you solve. Superficial personalization, like complimenting a post or mentioning a school, rarely helps and can make emails feel more automated.

What should AI personalize in a cold email?

Usually the opening line or a sentence connecting a specific account fact (a hire, a launch, a tech change, a market) to the problem your email addresses. The core message should stay consistent at the segment level.

How do I stop AI personalization from sounding robotic?

Ban flattery and stock phrases like 'I noticed' or 'I was impressed', limit the line length, require a specific verifiable fact, require a connection to your offer, allow the model to return 'NONE' and review samples regularly.

What data does AI need to personalize well?

Real source content: the company's website, recent news, job posts, product pages and sometimes the person's LinkedIn profile. Without source content, models fall back on generic statements or invent facts.

Should every cold email be personalized with AI?

No. Personalize when there's a relevant fact. When there isn't, a clean segment-level email is better than a forced personalized line.

Which AI model is best for cold email personalization?

Modern models from Anthropic and OpenAI all perform well. The prompt, source data and guardrails matter far more than the model choice.

How do I quality-check AI personalization at scale?

Review a random sample of outputs every batch, track error and 'NONE' rates, check facts against sources, and monitor positive reply rate and negative replies for signs of problems.

Ready to stop doing outbound by hand?

Book a strategy call. We’ll map your ICP, review what’s working (and what isn’t) and show you exactly what we’d automate first.

Book a strategy call

Keep reading