Short answer: let AI produce a first draft, then make the message yours. The parts that make a message worth answering are the specific reason you are writing and the decision to send it. A model cannot supply either one for you. It can only guess, and guesses read like guesses.
Why AI messages are easy to spot
Most AI outreach fails in the same few ways. Once you know them, you will see them in your own inbox.
- A compliment that could go to anyone: 'I was impressed by your impressive journey in the industry.'
- A summary of the person's profile read back to them. They already know their job title.
- Three benefits in a row, each one vague.
- An opening line that sounds like a greeting card, then a pitch in the second sentence.
- A closing question that asks for a call before giving a reason to take one.
- Length. Models are generous with words. Readers on a phone are not.
None of these comes from the model being bad. They come from the model having nothing specific to say, so it fills the space. Your job is to give it something specific, or to write that part yourself.
Before and after
Here are three drafts of the kind a model produces when it is given nothing but a job title, each followed by an edit. The people and companies are made up for the example.
| Draft | Message |
|---|---|
| AI draft | Hi Priya, I was impressed by your impressive career in logistics software. At GrowPython we help companies like Cobalt Freight unlock growth with cutting-edge solutions. Would you be open to a quick 15-minute call this week? |
| After the edit | Hi Priya, I saw Cobalt Freight is hiring two dispatch planners. Where do replies from carriers land for your team today, one inbox or several? |
| AI draft | Hello Daniel, I hope this message finds you well. I wanted to reach out because I believe our platform can revolutionize your agency's outreach. Let me know if you would like to learn more! |
| After the edit | Daniel, your agency's site lists LinkedIn and email as separate services. We run both from one inbox, so the replies land together. Worth showing you how that looks on one of your own campaigns? |
| AI draft | Hi Mei, just following up on my previous message. I know you are busy, but I would love to connect and share how we can help. Looking forward to hearing from you. |
| After the edit | Hi Mei, one more note and then I will leave it. I put together a one-page checklist on setting up a booking link for replies. Want it? If not, no problem. |
The edits share a shape. Each one names one fact the writer actually noticed, says one thing, asks one easy question and leaves the reader free to say no. None is clever. All are shorter.
A five step process
- Write down the facts first. Before you ask for a draft, know the one true reason you are writing to this group. It might be a hiring signal, a post, a shared tool or a question you hear often. If you cannot write that reason in one sentence, the message will not have one either.
- Teach the tool your voice once. In GrowPython this is the Business profile, under Settings. It holds your company description, your audience segments, your voice and tone, words to avoid, and your red lines: claims never to make and topics never to discuss. The AI parts of the product read the same description, so you do not repeat yourself in every prompt.
- Let variables carry the data you really have. In the Message library, a variable such as {{first_name}}, {{company}}, {{headline}} or {{occupation}} is filled in for each person. Use them for facts that are on the profile. Do not ask a model to invent a personal detail that a variable cannot supply.
- Ask for options, not an answer. In the Inbox, Suggest gives three drafts, shown as chips marked as AI drafts and not sent. Pick the closest one and rewrite it in your own words. Having three also shows you the pattern the model leans on, which is useful to avoid.
- Read the preview with a real person filled in. The Message library previews your message with an actual contact, so you see it as the reader would. Gaps are marked: if a person has no company on their profile, the company variable has nothing to fill, and the preview tells you so.
Rules that make the edit fast
- Open with them, not you. One short reason that is about the other person.
- Say one thing. One idea, one question.
- Make it easy to say no. A line that tells the reader to say so if now is not the time keeps the tone friendly.
- Cut every adjective that does not carry a fact. 'Powerful', 'seamless' and 'cutting-edge' go first.
- Keep follow-ups short and each one useful. A follow-up that only says 'just checking in' adds nothing.
- Do not invent facts. Only write what is true. Do not claim results, customers or numbers you cannot back up. If a model adds a statistic, delete it.
Limits that shape the draft
On LinkedIn, the connection note has a length limit. In GrowPython's Message library a connection note can run to 300 characters, and the counter warns past 200 because free LinkedIn accounts allow 200. The note itself is available for Premium accounts in the campaign flow. With a free account the request goes out without a note, so your reason to talk belongs in the first follow-up instead. Plan the shape of the sequence with that in mind. For examples of short connection notes, see LinkedIn connection request message examples, and for the follow-up stage, LinkedIn follow-up message templates.
Where GrowPython helps, and where it does not
The Message library keeps your messages in one place, by type: connection note, follow-up, meeting, sales, InMail or email. It checks length, flags variables it does not recognise, previews with a real person, keeps version history and shows where each message is used. The Copilot can draft messages as part of a campaign plan, and everything it makes is a draft that you read and approve. The Company Brain is where your voice and red lines live, so every AI helper starts from the same description.
What it does not do: decide whether your message is good. It has no way to know whether your reason is true or whether the person cares. AI can be wrong, and it works from what is in your workspace. Read every line the way you would read a colleague's work. A message leaves only when you press Send, and a campaign starts only when you press Start. The same goes for email: replies from the Inbox go out after a short undo window of about ten seconds, which is a good last moment to reread.
It also promises no result. A better message can lift replies, and a poor list can sink the best message. If you want to know whether a change helped, change one thing at a time and compare against your own earlier numbers, as described in LinkedIn outreach metrics. For teams who write for clients, the same approach applies in the agency workspace.
A checklist before you send
- The first line is about the reader, not about us.
- One reason, one question, no pitch deck.
- Every variable filled in the preview for a real person.
- No claim I cannot back up.
- I would be comfortable if the reader saw the draft I started from.
- I read it once, out loud, on a phone-sized screen.
Questions
- Can I use AI to write outreach messages?
- Yes, as a first draft. Give it a specific reason to write, edit the result into your own voice and read it before sending. A message that leaves only when you press Send gives you that last check.
- How do I make an AI message sound like me?
- Write your voice, tone, words to avoid and red lines once in the Business profile, then edit each draft. Cut vague adjectives, shorten it and keep one idea and one question.
- Does GrowPython send AI messages by itself?
- No. Suggest and the Copilot only put drafts in front of you. Campaigns start when you press Start, and replies leave when you press Send.
- What should I never let AI write?
- Facts you cannot back up: results, customer names, numbers and personal details about the reader. If a draft contains one, delete it or check it yourself.