Quick answer
Use AI for the research and the first draft, and write the line that gets sent yourself. LinkedIn allows AI-drafted messages: it sells its own AI drafting in Premium and Sales Navigator, and its User Agreement asks you to review and edit generated content before you share it. What it bans is bots and other unauthorized automated methods that send messages for you. On results, the best public test is Expandi's: comparing each account's AI-personalized campaigns with its own human-written ones, human copy won on acceptance by about 12% and reply rates were flat. LinkedIn claims AI-assisted recruiter InMails get 44% more acceptances, without publishing a method. Lab studies point the same way as Expandi: people trust a message less once they think AI wrote it. Checked on 15 October 2026.
The short answer: AI for the work, you for the thinking
I set up LinkedIn outreach next to cold email for B2B teams, and "can AI just write the messages?" is a question I hear a lot.
By AI-personalized outreach I mean a tool that reads each prospect's profile, posts and company news, writes a custom opener for every person and often sends it too. My rule for it is the same as for any automation: automate the work, not the thinking. AI owns the speed, you own the judgment.
What LinkedIn's rules say about AI-written messages
LinkedIn has no rule against AI-written text. It sells it. Premium Business comes with "AI message drafts", and Sales Navigator has Message Assist, which drafts first InMails from Lead IQ and Account IQ data. What LinkedIn's rules care about is how a message gets sent and whether it's spam.
| What you do | What LinkedIn says | Allowed? |
|---|---|---|
| Draft with Message Assist or Premium's AI drafts, then edit and send | LinkedIn recommends you review and edit the draft. The recipient isn't told AI helped. Message Assist allows 50 AI-assisted messages a month, one lead at a time | Yes |
| Draft in ChatGPT or another tool, edit it and send it yourself | Section 3.6 of the User Agreement: review and edit generated content before sharing it. You're responsible for it under the Professional Community Policies | Yes, it's your message |
| Let a tool write and send messages from your account | Section 8.2 bans "bots or other unauthorized automated methods" to add contacts or send messages | No |
| Send near-identical AI messages to hundreds of strangers | The Professional Community Policies ban "untargeted, irrelevant, obviously unwanted" and "gratuitously repetitive" messages | No, whoever wrote them |
Sources: LinkedIn's User Agreement (effective 3 November 2025), Professional Community Policies, Message Assist and the Premium vs Sales Navigator page.
So drafting with AI is fine. The risk sits in the sending. AI SDR tools and automation platforms that log into your account and send for you fall under the same ban as any other automation, which I go through in is LinkedIn automation safe.
What the evidence says about AI-written outreach
Public data on AI-written LinkedIn messages is thin, and most of it comes from vendors. Here's what I could check, with how each study was run:
| Study | How it was run | What it found | How close to LinkedIn outreach |
|---|---|---|---|
| Expandi, State of LinkedIn Outreach H2 2026 | 118 campaigns from 73 Expandi accounts, October 2025 to June 2026. Each account's AI-generated campaigns compared with its own human-written ones. Observational | Human-written campaigns won on acceptance by about 12%. Reply rates effectively the same | Closest: real LinkedIn campaigns. Small sample, all from one automation tool |
| LinkedIn, AI-Assisted Messages in Recruiter | "LinkedIn data". No method published | 44% higher acceptance, accepted over 11% faster by job seekers than non-AI messages | Recruiters messaging candidates, who often want what's offered. A vendor claim |
| Jakesch et al., CHI 2019 (Cornell and Stanford) | Three online experiments. People rated real Airbnb host profiles after being told some or all were written by AI | When people thought they saw a mix of AI and human profiles, they trusted the ones they believed were AI-written less | Not sales. But a LinkedIn inbox today is exactly that kind of mix |
| Schilke and Reimann, 2025, Organizational Behavior and Human Decision Processes | 13 experiments with more than 5,000 participants, across work tasks from communication to analysis | People who disclosed AI use were trusted less. Being exposed by someone else hurt trust even more | Workplace trust, not outreach |
| Liu, Kang and Wei, 2023, Journal of Social and Personal Relationships | Online experiment with 208 adults. A friend's message written alone, with AI help or with another person's help | AI help made the friend look like they'd put in less effort, which lowered satisfaction. Help from another person wasn't significantly different | Personal messages, not sales. Shows effort is what readers judge |
Sources: Expandi, LinkedIn, Jakesch et al., Schilke and Reimann (participant count from the University of Arizona), Liu, Kang and Wei.
None of these is a clean experiment on cold LinkedIn outreach. Read together, they point one way: a message works partly because the reader believes a person took the time to write it, and anything that makes it look generated costs you some of that. For the cold email side, see does AI personalization actually lift reply rates.
Why AI looked 18% better in Expandi's data, then lost
Expandi's report shows how easy it is to fool yourself on this question. The first comparison looked great for AI: AI campaigns got 20.4% acceptance and 6.5% replies, against 17.3% and 4.6% for everyone else on the platform. That's an apparent 18% lift on acceptance.
Then Expandi checked who uses AI. Accounts that adopt it tend to be stronger operators, and their own non-AI campaigns already beat the platform average. Compared with their own human-written campaigns, the same accounts did about 12% worse on acceptance with AI and the same on replies.
Expandi gives three possible reasons, and all three are fair to keep in mind:
- AI copy "may read as personalized to the sender but feel formulaic to recipients" who have seen the pattern before.
- Teams may use AI on their coldest audiences, where lists are weaker anyway.
- The AI campaigns were newer, so late acceptances may not have come in yet. Expandi says the feature launched in late 2025.
LinkedIn's 44% figure for recruiter InMails may hide the same trap. LinkedIn doesn't say whether it compared the same recruiters with and without AI.
So the fair reading is that unedited AI personalization didn't beat the same people's own templates, and I found no public data showing that it does. The same report found bigger gaps elsewhere: 9.8% replies for three-message campaigns against 5.0% for five or more, and about 32% acceptance for morning requests against about 24% in the evening.
Where AI helps in LinkedIn outreach and where it hurts
| Job | Give it to AI? | Why |
|---|---|---|
| Building and filtering the list | Yes | Matching companies and roles to your criteria is work, not judgment |
| Spotting a reason to reach out now: a new role, hiring, funding, a post | Yes | AI reads profiles and news faster than you can |
| A first draft | Yes, as a draft | Both Expandi and LinkedIn say to review and edit it |
| The line about the person that actually gets sent | No | This is where readers spot a generated shape |
| Deciding who gets a connection note or an InMail | No | You get a few notes and a limited number of InMails a month. Spending them is judgment |
| Answering a reply | No | A real person wrote back. Answer as one |
| Clicking send | You, by hand | Tools that send for you break LinkedIn's automation rule |
The best lead is the one who just did something. AI is good at finding what they did. Deciding whether it matters, and saying something true about it, is your part.
Same prospect, two messages
Here's the shape most AI-personalized openers fall into, next to one written after the same research.
AI-personalized, unedited
"Hi Marta, I came across your profile and was really impressed by your journey to Head of Sales at Nordlys! Your recent post about building high-performing teams really resonated with me. I'd love to connect and explore how we help sales leaders like you drive more pipeline."
Written by a person after the same research
"Hi Marta, saw you joined Nordlys as Head of Sales last month. Is outbound running on lists someone else built, or are you starting from scratch?"
The first one is full of personal details and still says nothing about Marta. Swap the name, the title and the post topic and it fits the next 50 people. The second could only go to her, and it asks one question she can answer in a line.
To get from the first to the second step by step, use my checklist for editing an AI-drafted LinkedIn opener. If you sell into Sweden, Norway, Denmark or Finland, tone matters even more, as I explain in why AI-personalized outreach backfires with Nordic buyers.
Rule of thumb. If the opener would still make sense after you swap in another prospect's name and post, it isn't personalized. Rewrite it before it goes out.
If you want AI to write anyway, test it on your own list
Expandi's result came from comparing the same accounts with themselves. You can run the same test:
- Split one segment at random into two halves: same targeting, same offer, same sender.
- Send half the AI-personalized version and half the version you wrote.
- Run both for two to three weeks, so late acceptances come in.
- Compare acceptance and replies, then read 20 replies from each side.
Be realistic about sample size. By my math, at the usual 80% power and 5% significance, reliably seeing a 12% drop from a 28.5% acceptance rate (Expandi's platform average) takes roughly 2,600 requests per side. That's far more than one profile should send in a few weeks. A small test will catch a big difference, not a 12% one, so if your volume is low, read the replies and judge them as a reader would.
Key takeaways
- LinkedIn allows AI-drafted messages and sells its own. It bans bots and other automated methods that send them, and it bans spam whoever writes it.
- The closest public test, Expandi's within-account comparison, found human-written campaigns about 12% ahead on acceptance, with replies flat. It's observational and small.
- LinkedIn's 44% acceptance lift for AI-assisted recruiter InMails comes with no published method.
- Lab studies show people trust text less once they believe AI wrote it, and read it as less effort.
- Give AI the list, the research and the first draft. Write the line that gets sent, choose who gets it and click send yourself.
FAQ
Can I use AI to write LinkedIn messages?
Yes. LinkedIn sells AI drafting in Premium and Sales Navigator, and its User Agreement asks you to review and edit generated content before you share it. You're responsible for what you send.
Does LinkedIn flag AI-generated messages?
Not that LinkedIn describes. Its Message Assist page says the recipient won't be told AI helped draft the message. LinkedIn's rules target automated sending and spam, not who drafted the text.
Is it against LinkedIn's rules to use AI to send messages?
Using bots or other unauthorized automated methods to send messages is against section 8.2 of the User Agreement, and that includes AI tools that send from your account. Drafting with AI and sending by hand is allowed.
Do AI-personalized LinkedIn messages get more replies?
Not in the best public test. In Expandi's within-account comparison, human-written campaigns got about 12% more acceptances and the same reply rate. LinkedIn reports a 44% acceptance lift for AI-assisted recruiter InMails but doesn't publish how it measured it.
What's the best way to use AI for LinkedIn outreach?
Use it for list building, research and first drafts. Keep the final opener, the choice of who to contact and the replies for a person. Automate the work, not the thinking.