How do you stop AI-written outreach sounding like slop?
Start from a signal you can name in one sentence. If you cannot say why this person and why now, no amount of rewriting saves the message. Then hold it to four lines and one idea, cut every compliment without a referent, and give every merge variable a fallback.
The mechanism: the signal comes first and has to be nameable
Slop is not a style problem. It is what a message looks like when there was no reason for it. The writer had nothing specific, the model filled the space with warmth, and warmth with no referent is the loudest tell in outbound.
So the test happens before drafting. Say out loud, in one sentence, why this person and why now, in terms the recipient would recognise as true. They posted about a specific thing last week. They are hiring three of a specific role. They moved from one company to another in March. If you cannot produce that sentence, the repair is not a better prompt, it is a different person on the list.
This is also why targeting outranks copy. Across two of our audiences with comparable copy, acceptance was 32% and 2.6%. The writing did not differ enough to explain that. A reason to write existed in one case and not in the other.
Four lines, one idea, zero flattery
Line one is the signal, stated plainly, with enough detail that it could only apply to this person. Line two is what that signal means for them, not for you. Line three is the ask, small enough to answer in a sentence. Line four is the out, which is the line that makes the other three credible.
One idea, because a second idea reads as a pitch. Zero flattery, because a compliment is what people write when they do not have line one. A useful edit: if the message still works with the first line deleted, that line was decoration and should go.
A note on length, since people argue about it. Four lines is not a magic number. It is what fits in a preview and what a stranger will read without first deciding to read. If the signal genuinely needs a fifth line, use it. If you are at eight, the message is doing two jobs, and the second one is what will get it ignored.
The banned list, and why each one is a tell
"Pick your brain" asks for unpaid work. "Hope this finds you well" is a template clearing its throat. "Quick question" is never a quick question. "Huge fan" and "love what you're doing" are compliments with no referent, which is to say no evidence that anyone looked. "Synergy", "touch base" and "circle back" are noises a template makes.
Also banned: any compliment that does not name the specific thing being complimented, and emoji in a first touch. Both try to buy warmth the message has not earned yet.
Ban them literally, as strings, not as tone guidance. A rule the model can check beats an instruction about voice, because voice guidance degrades over a long generation and a banned-string check does not.
Every variable needs a fallback
A message that ships with a raw placeholder is worse than a message with no personalisation at all, because it proves nobody read anything and that this went out to a list. One of those in a hundred sends is enough to define you to everyone who sees it.
So every merge variable gets a fallback, and the fallback has to produce a sentence that still stands up on its own. If the fallback version of the sentence reads as generic, the sentence should not be in the message at all. That rule is stricter than it sounds, and it deletes most of what usually gets merged in.
The same logic applies one level up, to the sentences that are not variables at all. If a sentence would be true of everyone on the list, it carries no information and costs you a line of attention to say. Read your own draft and cross out every sentence that would survive being sent to a different person. What is left is the message.
The cost of getting this wrong went up
LinkedIn shipped a report option for AI-generated spam, which means recipients now have a labelled button for precisely the thing bad outbound produces. Volume without quality burns the sender twice: the recipient is gone, and the report is a durable signal attached to your account rather than to your campaign.
That sits on top of the older signal, where people mark that they do not know you. Both are behavioural, both are scored, and neither is addressed by sending from a different address or a cleaner IP.
The practical implication is that quality and account safety are now the same project. Fewer, better-aimed messages protect the account, which is the same thing that produces a higher acceptance rate. Across 861 conversations over three years on five seats, ours averaged 41%.
How to make an agent hold the line
Put the rules where they persist, rather than in a chat instruction you retype every session. Banned strings, the four-line shape, the required signal, the fallback rule. An agent that re-reads the rules on every draft holds them far better than a person does at message forty.
Then keep a human on the approval gate and review drafts rather than results. A message you would not have sent yourself should not go out because the queue was long, which is an argument for keeping the queue short enough that approval is a real decision and not a formality.
The measurement that keeps this honest is acceptance rate by audience rather than in aggregate. An aggregate hides the split that matters, which is that one list is working while another is quietly spending the account. Two of our audiences sat at 32% and 2.6% at the same time under comparable copy, and an average across them would have described neither.
The tell, and what to write instead
| The tell | Why it reads as slop | What to do instead |
|---|---|---|
| "Hope this finds you well" | A template clearing its throat | Open on the signal |
| "Quick question" | It never is | Ask the actual question in line three |
| "Huge fan" / "love what you're doing" | Compliment with no referent | Name the specific thing, or cut the line |
| "Pick your brain" | Asks for unpaid work | Ask one question you would answer yourself |
| "Touch base" / "circle back" / "synergy" | Noises a template makes | Say the thing in plain words |
| Emoji in a first touch | Buys warmth the message has not earned | Earn it with line one |
| A raw placeholder that shipped | Proves nobody read anything | A fallback for every variable |
Questions people ask next
Why does AI-written outreach sound generic?
Because it usually starts without a reason to exist. With no specific signal, the model fills the gap with flattery, and flattery with no referent is the loudest tell there is.
What phrases should I ban from cold outreach?
"Pick your brain", "hope this finds you well", "quick question", "huge fan", "love what you're doing", "synergy", "touch base", "circle back", any compliment with no specific referent, and emoji in a first touch.
How long should a first LinkedIn message be?
Four lines: the signal, what it means for them, a small ask, and an easy out. One idea. A second idea reads as a pitch.
Does personalisation still work now that everyone automates?
Specific relevance does. Merge-field personalisation does not, and a raw placeholder that ships is worse than sending no personalisation at all.
Why this page exists: LinkedIn's report option for AI-generated spam, read against our own 861-conversation message history
LinkedBoost is the LinkedIn MCP server: your agent sources, drafts, sends and works the inbox, inside caps the server enforces rather than suggests.
Related answers
Fully autonomous AI SDRs mostly did not hold. The teams still getting value moved the agent from decision-maker to operator: it researches, drafts and paces, a human approves, and volume stayed where a human could review it. The failure was never the writing.
Five checks, in this order: where the approval gate sits, whether limits are enforced or merely suggested, what happens when a cap is reached, whether the vendor publishes how it measures account health, and whose data the leads come from. A ranked list ages badly. A rubric you can run yourself does not.