Quick answer
Expandi's H2 2026 report on 13.2 million connection requests initially made AI-generated LinkedIn messages look 18% better than the rest of the platform. That comparison is confounded: accounts that adopt AI tools tend to be stronger operators overall. When Expandi controlled for that by comparing each account's own AI campaigns against that same account's human-written campaigns, human-written copy won on acceptance by about 12%, with reply rate coming out effectively identical. Use AI on LinkedIn for research and drafting speed, not as the thing that's supposed to write the message that lands.
The pitch every AI LinkedIn tool is making
I'm Hlib Storchak. I build outbound systems for B2B founders and sales teams, and I've booked 2000+ meetings for B2B clients doing it, which means I've sat through a lot of demos for tools that promise to write a better LinkedIn message than a person would. The pitch is always some version of the same sentence: the AI reads the prospect's profile, their recent posts, their company's latest news, and drafts an opener more relevant than anything a rep could write at scale. It's a reasonable-sounding claim. It's also the kind of claim that usually ships with no source attached, which is exactly why I went looking for one.
The number that makes AI look like the obvious win
Expandi's own State of LinkedIn Outreach H2 2026 report, built on 13.2 million tracked connection requests and 6.7 million outbound messages across 13,302 active accounts from May 2025 through April 2026, is one of the largest independent LinkedIn outreach datasets around, and I've relied on it directly for benchmark figures in other pieces on this blog. Read the surface-level number and AI looks like a clear win: accounts using AI to generate their outreach saw roughly 20.4% acceptance and 6.5% reply, against 17.3% acceptance and 4.6% reply for the rest of the population, an apparent 18% relative lift on acceptance. If you stopped reading there, you'd have every reason to switch your whole seat over to AI-drafted messages tomorrow.
The catch Expandi's own report found
Expandi didn't stop there, and the report is honest about why that comparison is misleading on its own: it's comparing AI-using accounts against everyone else, and accounts that adopt AI tools in the first place tend to be stronger operators across the board, better targeting, better list hygiene, more disciplined sequencing, independent of what wrote the actual message. So Expandi re-ran the comparison the way you'd want a skeptic to run it: within account, meaning each account's own AI-generated campaigns measured against that same account's own human-written campaigns, on the same underlying targeting and operator skill. Across 118 campaigns from 73 accounts between October 2025 and June 2026, the picture reversed. Human-written messages outperformed AI-generated ones on acceptance by about 12%, and reply rate came out effectively identical between the two.
Why this matters. The 18% number and the 12% number come from the same report, the same company, and roughly the same time window. The only thing that changed is which accounts got compared to which. That's the whole lesson: a lift number is only as good as what it's being compared against, and "AI users versus everyone" is a different, much weaker claim than "AI versus human, same operator."
The two comparisons, side by side
Laid out next to each other, it's easier to see exactly where the apparent win comes from and where it disappears.
| Comparison | What's being measured | Result |
|---|---|---|
| AI-using accounts vs. rest of platform | Different accounts, different operators, different targeting | AI side: 20.4% acceptance, 6.5% reply. Rest: 17.3% acceptance, 4.6% reply. ~18% apparent lift for AI. |
| Within-account: AI campaigns vs. that account's own human campaigns | Same account, same operator, same targeting, only the writer changes | Human-written wins acceptance by ~12%. Reply rate effectively identical. |
Source: Expandi's State of LinkedIn Outreach H2 2026 report, fetched directly. The within-account figure is the one worth planning around, because it's the only one of the two that actually isolates the writer as the variable.
Why the reversal makes sense once you think about it
Selection bias like this shows up constantly in vendor benchmarks, and once you see the pattern once, you start spotting it everywhere: a tool's power users tend to be good at the job for reasons that have nothing to do with the tool, and a naive comparison credits the tool for the user's underlying skill. It's the same shape of mistake as crediting a CRM for a sales team's revenue growth when the team that adopted the CRM was already the highest-performing team in the building. Expandi controlling for it and publishing the corrected number anyway is, frankly, more transparency than most vendor benchmark reports offer, and it's the reason I trust this specific report enough to build an article around it.
This isn't the first channel where this showed up
I wrote up a similar finding for cold email a while back: a 10,000-email study found AI-written cold emails replying at 8.2% against 11.7% for human-written ones, with the gap widest in the most relationship-driven verticals. That earlier piece is about email specifically, and this one is about LinkedIn's own, differently structured dataset, but the direction is now consistent across two separate channels and two separate studies: a generic AI draft on a cold or lukewarm list tends to read as generic, and a person who can tell reacts to it accordingly. Two channels agreeing isn't proof of a universal law, but it's a lot more than one anecdote.
What's actually driving acceptance and replies instead
The same H2 2026 report is more useful for what it says moves the numbers than for what it says about AI specifically. Acceptance varies hugely by industry, from 17.5% in Consumer Electronics up to 40.1% in Broadcast Media, and reply rate splits just as wide: Staffing & Recruiting converts at roughly 18.9%, more than double Software's 8.8%, even though Software drives 14% of all tracked volume. None of that has anything to do with who or what wrote the message. It's targeting, industry norms, and how used to LinkedIn outreach a given audience already is.
4 things that moved these numbers more than the writer did
- Sequence length. Campaigns with three messages reply at 9.8%. Push past five messages and reply rate drops to 5.0%, worse than a single message alone, per the same report's analysis of 2,966 campaigns with at least 50 contacts each. Verdict: three is the number to build around, not the number to blow past.
- Message length. Across 458 qualifying single-message campaigns, 150 to 200 characters hit the peak reply rate at 8.5%. Shorter or much longer both underperform it. Verdict: write to the length, not to how much you have to say.
- Send timing. Morning sends between 7 and 11am land roughly 32% acceptance against roughly 24% in the evening, a real and fairly large gap. Day of week barely moves the number by comparison, a tight 26.5% to 30.4% band. Verdict: fix your send window before you touch anything else.
- Lead source. Sales Navigator's own targeting filters produced 19.8% acceptance against 18.5% for a plain CSV import, with reply rate coming out identical either way. Verdict: a real but small edge, not a reason to pay for Sales Navigator on its own.
The "AI voice" problem a template doesn't have
I don't think the within-account result is a fluke, and I think there's a plain reason for it: a well-written human template, reused across a list, still sounds like one specific person wrote it, because it was written by one specific person with a real voice. A lot of AI-generated personalization, even when it correctly pulls a name, a job title, and a recent post, tends to land in the same handful of sentence shapes and the same handful of compliment-then-pivot structures, because that's what the underlying model gravitates toward without heavy prompting and editing. Prospects on LinkedIn see a lot of outreach. A message that reads like the fortieth AI-personalized opener they've seen this month, even a technically well-targeted one, is working against a pattern-match the reader has already built up a resistance to.
The mistake I see most often when I take over an AI-run seat
When a client hands me a LinkedIn seat that's been running an AI-drafting tool for a few months, the pattern I see most often isn't a disaster, it's a slow fade: acceptance and reply both drift down over a few weeks, and nobody notices because there's no single bad send to point at. The fix is rarely "turn the AI off entirely." It's usually rewriting the two or three templates the account actually needs, keeping the AI for the research step, pulling the recent post or the job change worth referencing, and having a person write the actual sentence that uses it. That split, AI for finding the signal, a person for the line that says something about it, is the version of this I run for clients now.
Where AI on LinkedIn is still worth using
None of this is an argument for abandoning AI tooling on LinkedIn. It's an argument for being precise about which job you're asking it to do. Surfacing a prospect's recent activity, summarizing a company's latest news, flagging a job change worth mentioning, drafting a rough first pass a human then edits down to their own voice: all of that is a legitimate use of AI, and it's faster than doing the same research by hand. What the data argues against specifically is letting AI write and send the final line unedited and expecting it to outperform a person who already knows how to write a good LinkedIn opener. Research assistant, yes. Ghost writer with no editor, not yet, not per this data.
How to test this on your own account before you believe either of us
Expandi's number is the best public data I've found on this exact question, but it's still one vendor's dataset over one time window, not a universal constant, and I'd rather you check it than take my word or theirs. Split a real segment of your own list in half, run the same targeting and the same offer through an AI-drafted version and a human-written version, and let both run for at least two to three weeks before you read the result, since a fresh sequence swings a lot in its first few days. The number that should actually change how you run your seat is the one from that test on your own list and your own audience, not a benchmark from someone else's 73 accounts.
Key takeaways
- Expandi's H2 2026 report initially shows AI-generated LinkedIn messages beating the rest of the platform by about 18% on acceptance, but that comparison confounds the writer with the fact that AI-adopting accounts tend to be stronger operators overall.
- Controlling for that with a within-account comparison (each account's AI campaigns versus its own human campaigns) reverses the result: human-written messages win acceptance by about 12%, with reply rate effectively identical.
- A similar pattern already showed up in cold email, where a separate 10,000-email study found human-written emails outreplying AI-written ones, especially in relationship-driven verticals.
- Sequence length (three messages), message length (150 to 200 characters), and send timing (morning over evening) all move acceptance and reply more than the choice of AI versus human writer.
- AI is genuinely useful on LinkedIn for research and drafting speed. The data argues against letting it write and send the final message with no human edit, not against using it at all.
- Treat any single vendor benchmark, mine included, as a starting hypothesis. Confirm it with a split test on your own list before you change how your seat runs.
FAQ
Does AI-written LinkedIn outreach actually perform worse than human-written outreach?
In Expandi's H2 2026 report's within-account comparison, the version that controls for operator skill, yes: human-written messages won on acceptance by about 12%, with reply rate coming out effectively identical. A naive comparison across the whole platform shows the opposite, an apparent 18% lift for AI, but that comparison is confounded because stronger operators are more likely to adopt AI tools in the first place.
Why do two comparisons from the same report give opposite answers?
Because they're measuring different things. Comparing AI-using accounts against everyone else measures the accounts, not the writing, since AI adopters tend to already run tighter targeting and better lists. Comparing each account's AI campaigns against that same account's own human campaigns holds the operator and targeting constant and isolates the writer, which is the comparison that actually answers the question.
Should I stop using AI tools for LinkedIn outreach?
No. Use AI for what it's genuinely good at on LinkedIn, research, summarizing a prospect's recent activity, flagging a job change or a relevant post, and drafting a rough first pass. The data argues against trusting it to write and send the final message unedited, not against using it as part of the process.
What actually predicts a higher LinkedIn acceptance or reply rate, if not the writer?
Per the same H2 2026 report: a three-message sequence outperforms longer ones, messages of 150 to 200 characters hit the highest reply rate, morning sends between 7 and 11am beat evening sends by a wide margin, and industry accounts for some of the largest swings, from 17.5% acceptance in Consumer Electronics up to 40.1% in Broadcast Media.
How do I check a claim like this on my own LinkedIn account?
Split a real segment of your list in half, run an AI-drafted version and a human-written version through the same targeting and offer, and let both run for at least two to three weeks before reading the result. The number from that test on your own audience is worth more than any benchmark report, including this one.