AI personalization in outbound: what works and what reads as fake

AI personalization works in outbound when it's built on real evidence about the person and explains why you're writing now. It fails when it pads a template with profile trivia, compliments or details the model made up. The difference comes from the input more than the model: give the AI a specific reason, like a post the person wrote or a question they asked, and it writes a message worth reading. Give it only a name, a title and a company, and you get something every prospect has seen a hundred times.

Here is where AI helps, where it hurts, and how to set it up so the messages sound like a person who did their homework.

Why does most AI personalization sound fake?

Because it personalizes the wrong thing. The typical AI opener takes something from a profile and builds a sentence around it:

Hi Dana, I see you went to Michigan. Go Blue! I noticed you've been at Acme for three years, which is impressive.

Every detail is true and none of it matters. It doesn't say why the sender is writing, and prospects have learned that a compliment about their school or their tenure means a template filled in automatically.

The other common failure is invention. A model asked to "reference their recent work" with nothing to go on will produce something plausible: a project they didn't lead, an article they didn't write. One invented detail ruins the message and the sender's credibility.

What kind of personalization actually works?

Personalization that answers "why you, and why now?" The useful question isn't how much you know about someone, but whether the detail explains the message.

The bottom two rows are where AI earns its keep. It can read a person's post and comments, pull out the idea that matters, and draft an opener that responds to it, faster than any rep.

What should AI do in outbound, and what should it not?

Good uses:

1. Summarizing evidence. Reading what a prospect posted or engaged with and pulling out the one relevant point.
2. Drafting from that evidence. Writing a short opener that responds to the point.
3. Adapting tone. Matching the register of the person's own writing, formal or casual.
4. Writing variations. Giving you options to test instead of one message sent to everyone.

Bad uses:

1. Filling gaps with guesses. If there's no evidence, the model shouldn't invent some.
2. Writing without review at the start. Early drafts need a human check before they go out.
3. Flattery on demand. Instructions like "make it warm" tend to produce compliments nobody believes.
4. Deciding who to contact. Targeting is a judgment call about your market. AI can score fit, but you should set the criteria.

How do you write instructions that produce good messages?

Put the rules in the instructions, in plain words, and give the model the evidence directly rather than asking it to find some:

• Give the reason for writing. "This person commented on a post about X and said Y."
• Ban invention. "Only mention facts that appear in the evidence. If there is no evidence, write a short, plain message without personal details."
• Ban the usual filler. No compliments, no "I hope this finds you well", no "I came across your profile", no claims of familiarity you don't have.
• Set the shape. One observation, one sentence on why it matters, one easy question. Under 80 words for a first message.
• Show examples. Two or three messages you would actually send teach the model more than a page of adjectives.

How much human review do AI messages need?

More at the start, less over time, and never none. Early on, read every draft and correct what's wrong. Each correction should become a standing rule, like "don't mention funding unless it happened in the last three months" or "never use the word 'quick'". As the rules build up, the corrections get rarer.

Once most drafts go out untouched, move to sampling: read a share of what goes out each week and look for new failure patterns. Models change, evidence sources change, and the mistakes change with them.

Can prospects tell a message was written by AI?

Often, when it follows the usual patterns: an overly polished opener, a compliment, three-part sentences and a vague benefit. They're much less likely to notice a short message that responds to something they actually said, because that's what a thoughtful person would write.

The goal isn't to hide the tool. It's to send a message that would be worth reading whoever wrote it. For examples of what that looks like on LinkedIn, see our connection request examples and follow-up messages.

FAQ

Does AI personalization improve reply rates? It can, when it's based on real evidence and tied to a reason for writing. Personalization based on profile trivia or invented details tends to do no better than a plain template, and sometimes worse.

What data should I give an AI to personalize a message? Something the person did or said recently: a post, a comment, a question in a community, a job change or a company event. The more specific and recent, the better.

How do I stop AI from making things up in sales emails? Tell it to use only facts in the evidence you provide, give it that evidence directly, and review drafts. Also tell it what to write when there is no evidence: a short, plain message.

Should every message be AI-personalized? No. If you have no real evidence for a lead, a short, honest message is better than a personalized one built on nothing.

Is it okay to send AI-written messages without reading them? Not at the start. Review every draft until the corrections become rare, then keep sampling.

Where Saava fits

Saava's Full AI outreach writes each message from the evidence it found for that lead, such as the post they engaged with and why they surfaced, and puts the drafts in a review queue where you can edit, approve or reject them. Corrections you make become standing guidance for future drafts. If you'd rather use your own AI, Saava can hand the evidence to it through its MCP connector.

Get 100 free leads →