AI for LinkedIn messages: what to let it write, and what stays yours

The question about AI for LinkedIn messages is no longer whether it can write them. It can, fluently, in any language. The question is which messages it should write, and which parts of a conversation you should never hand over. Get that split wrong and your prospects receive the same polished, empty paragraph they already get from everyone else.

Here is the split that works, and what the AI in Swimlane actually reads before it writes a word.

What AI is good at in LinkedIn messages, and what it is not

AI earns its place on the tedious half of messaging:

• Continuity. Reopening a thread from three weeks ago and remembering where it stood, what was asked and what was promised.
• Register. Matching the language, the formality and the length the other person uses.
• The blank composer. A first draft you react to is faster than a sentence you have to invent.
• Summaries. Turning forty messages into three sentences.

It is bad at the half that decides whether a deal happens:

• Why now. The reason you are writing to this person today — their post, their new role, a customer you share — is rarely in the data the AI sees.
• Facts outside the thread. What you said on the call, the price you quoted, the date you promised.
• Stakes. When to push, when to wait, when to stop.
• Sounding like no one else. Generic openers are easy to recognise, and a prospect who spots one stops reading.

The rule that follows: let the AI draft sentences, and keep the reason, the facts, the ask and the Send button.

A reply draft in your tone of voice

The most useful AI in a LinkedIn inbox is the one that drafts your next reply. This is how Swimlane’s works, because the details decide whether you can trust it.

When it drafts. Only when the contact wrote last with a real message and you have both already spoken. A “thanks!” does not trigger anything. When nothing useful can be added, it proposes nothing rather than filler.

What it reads:

• the recent thread, up to its last 16 messages with text;
• the contact’s name and headline;
• the contact analysis: its summary, the recommended next step and the suggested follow-up date;
• your private notes on the contact and the reminders still open on them;
• what you wrote in Settings > Personalization: about yourself, your business, your ideal customer and how you help.

How it writes. In your name, to move the conversation toward your goal rather than simply agree with the prospect. One or two sentences, in the language and formality of the thread, with an instruction never to invent facts.

Whose voice. Swimlane learns your tone of voice from the messages you type yourself in its composer — not campaign sends, and not drafts you sent without changing a word, so it never learns from its own writing. From your recent messages it builds a short style guide: register, sentence length, greetings and sign-offs, recurring expressions, emoji habits. Names of people and companies are deliberately left out. It is refreshed as you keep writing.

What you do with it. The draft appears greyed out in the composer. Accept it and it becomes an ordinary draft you edit. Nothing is sent until you press Send.

What to keep for yourself in a reply

• Facts and numbers. The draft only knows the thread, your notes and your reminders. If the call is not in a note, it did not happen for the AI. Write the note; the next draft improves.
• The ask. Decide the next step yourself. The draft follows the recommended next step and your reminders, which is a suggestion, not a strategy.
• Anything with stakes. Pricing, objections that matter, a breakup message, bad news. Draft it, then rewrite it.

Icebreakers: a first line, not a reason

When you open a contact you can message and there is no conversation yet, Swimlane drafts an opener. It reads their first name, headline, current role and company, location and About section, plus what you wrote about your business, and writes one short, casual sentence in your language, with no pitch. On a thin profile it stays generic and invents nothing.

That is exactly what an opener should be, and exactly why it is not enough. A hook taken from a profile tells the prospect you read their headline. The reason you are writing — the post, the event, the problem you solve for people like them — has to come from you. Keep the AI’s first line if it is good, and add your second.

Templates the AI rewrites, and the sequence question

Message templates come in two types. Static templates are sent as written, with variables filled in. AI templates are reworked for each contact: your text becomes the example, the AI keeps the idea and the structure, and adapts it using the contact’s first name, last name, headline and location, the last five messages of the thread, your business context and your tone of voice. In the inbox, the result lands in the composer for you to check.

In a campaign it is different. A message step linked to an AI template is written for each contact at send time, without the thread, and goes out with no one reading it. Our advice:

• Use static templates with variables for first touches, where one wrong claim costs a relationship.
• Try an AI template in the inbox on a handful of contacts before linking it to a step, and read every result.
• Link steps to saved templates. Campaign steps are locked once a campaign is launched; editing the linked template is how you change a message later.

For the sequence side, see how to automate follow-ups without losing the personal touch.

Contact analysis: a second opinion on the relationship

The contact panel has an analysis section with three parts:

• Summary: where the relationship stands, in one to three sentences.
• Recommended Next Step: one action, with a “Get a reminder” button that creates the reminder on the suggested date.
• Suggested pipeline: a stage from your own pipelines, with a “Move” button.

It reads up to the last 40 messages, recent activity such as profile visits, accepted invitations and link clicks, the contact’s role and company, the connection degree, your context and your pipeline stages. It needs a little history — more than two messages or two interactions — and refreshes itself after new activity, at most once a week, with a button to force it sooner. It is told to choose the earliest stage that fits and to move forward only on explicit evidence in the messages.

What to keep: accept reminders freely, because a wrong reminder costs nothing. Move stages yourself when the move changes your forecast. The analysis sees LinkedIn, not the phone call.

Fit levels: triage, not a verdict

Once you describe your ideal customer, Swimlane places each contact in a campaign audience on one of five levels, from very low to excellent, judged from their profile against your paragraph. In a People Search, each profile also comes with a one-sentence reason it fits or doesn’t. Use the level to decide who gets your time first and which contacts a campaign enrols; do not let it decide a deal. The full method is in how to prioritize LinkedIn prospects by ICP fit.

The split, in one list

Let the AI:

• re-read threads and summarise them;
• draft replies you will edit;
• write the first line of an opener;
• suggest a reminder date and a stage;
• sort contacts by fit.

Keep for yourself:

• the reason you are writing now;
• facts, prices, dates and commitments;
• the ask and the next step;
• negotiation, objections and breakups;
• stage moves that change your forecast;
• the Send button.

The LinkedIn inbox in Swimlane is built on that split: the AI writes drafts, you send messages. For the craft of the message itself, our LinkedIn messaging best practices still apply — AI or not.