How to review an AI-written cold message (anti-slop list)
Six checks that separate an AI-written cold message that gets replies from one that gets deleted: recipient view, plain speech, evidence, ask.
Namanyay Goel · · 5 min read
Contributors: Namanyay Goel
Original source · 2026-08-28 · Adapted and expanded for PumpGTM
Most AI-written cold messages fail the same way. They are polite, grammatical, and completely forgettable, because the model optimised for sounding like a message instead of saying something to a specific person. People call it slop. The recipient just calls it "delete".
I wrote the original rules after reviewing hundreds of drafts with PumpGTM customers. This is the review process built from them. Run it on every message before a campaign goes live, whether a model wrote it or you did.
Check 1: Read it from the recipient's position
Strip the subject line and your company bio and read what is left. The recipient arrives with zero context about you. Does the remaining text describe something they might care about this week?
Then check the problem belongs to the person's role. "Save engineering time" means nothing to a head of finance unless you connect it to a number they own. Personalisation is not their first name. It is making the business reason specific to what they do.
Check 2: Read every sentence aloud
If you would not say a phrase across a table at work, replace it. The words that fail this test are always the same: streamline, leverage, unlock, seamless, empower, "I hope this finds you well". They sound polished and make an ordinary problem harder to understand.
Keep the nouns concrete. Name the task, the person who does it, and what changes. "Your SDRs spend Friday afternoons rebuilding the same list" beats "teams struggle with manual prospecting workflows". Once the details are concrete the message usually gets shorter on its own.
Check 3: Check the evidence behind every claim
Every result in the message needs support you can show. If you cite a customer outcome, confirm you are allowed to share it and that you describe the situation accurately. One customer's result does not tell a stranger what they will get.
When you do not have measured results, say what the product does and what you expect, framed as an estimate. Never let the model fill the gap with an invented percentage. The reader will ask "says who?" and you will not have an answer.
Check 4: Remove the opener that does not earn its place
"I came across your profile." "I see you're the VP of Sales at Acme." "Congrats on the funding round." Delete each one and ask whether the recipient understands any less about why you wrote. They do not.
Use the space for the actual trigger. If there is a real reason (a public question about a workflow you understand, a comment on a competitor's post, a hiring signal), say it plainly and keep it faithful to the source. If there is no real trigger, that is a targeting problem, and no opener will fix it. See finding buyer intent in competitor comment sections for where real triggers come from.
Check 5: Make the next step easy to answer
Do not ask a stranger for thirty minutes. Ask them to confirm the problem, or to react to one specific piece of information. A yes or no they can type in five seconds.
One ask per message. Three questions at the end means none of them gets answered. If the message has already explained why it matters, a direct request is fine. If it has not, no request will land.
Check 6: Separate writing quality from campaign results
Good editing cannot fix a bad list. When a campaign underperforms, decide first whether the targeting or the copy is at fault, and change one at a time. For each test, record the audience, the message version, the send window, and what you counted as a result.
Then read the replies, not just the reply rate:
- "Not my team" means the targeting is off by one role.
- "We already use X" means the problem is real and the timing is wrong.
- "How is this different from Y" is a qualified conversation.
- A polite no is just a no.
That reading is how the sequence improves. It is the same loop we run in the multi-touch LinkedIn sequence checklist.
A before and after
Before (typical AI draft): "Hi Sarah, I hope this message finds you well! I came across your profile and was impressed by your work at Acme. At PumpGTM we help teams streamline their prospecting workflows and unlock pipeline growth. Would you be open to a quick 20-minute call next week?"
After: "Hi Sarah, you asked under the Lemlist launch post whether it can pull people who commented on a competitor's thread. That is most of what PumpGTM does. Is that a workflow your team runs by hand today?"
Same length. The second one has a trigger, a concrete task, and a question she can answer in one word.
How PumpGTM uses these checks
PumpGTM drafts the first message for each prospect from the reason that person was selected, then puts the draft next to the prospect for you to approve. The six checks above are the review we expect you to do in that screen, and the reason the draft carries the trigger instead of a compliment. If you want to see drafts for your own prospects, start at pumpgtm.com.
Questions people ask about AI cold messages
Should I tell the recipient AI wrote it? If the message passes these six checks, it is your message. You reviewed it, you chose the trigger, you are accountable for it.
How long should a LinkedIn cold message be? Two to four sentences. If it does not fit on a phone screen without scrolling, cut it.
Is personalisation worth the time? Personalisation that changes the business reason, yes. Personalisation that mentions their alma mater, no.
What reply rate is good? It depends on the list more than the copy, which is the point of check 6. Fix targeting first, then measure copy changes against a stable list.