AI vs Human LinkedIn Messages: 1.2M-Send Reply Data

The debate over ai vs human linkedin messages finally has enough data to settle — and the answer is not what the AI SDR vendors have been selling. We pulled anonymized reply and acceptance benchmarks across roughly 1.2 million LinkedIn sends run through LinkedCamp in Q2 2026, then normalized by sender, ICP, and sequence position so we weren't comparing operators, we were comparing message stacks.

The short version: at the aggregate level, AI-assisted openers look like they win. Once you hold sender and ICP constant, that lift collapses — and in the segments where deals actually close (C-level, regulated verticals, follow-ups) human copy still beats AI outright. This matches what Expandi's H2 2026 within-account analysis found: AI-hyperpersonalized campaigns performed 12% worse on acceptance than the same accounts' own human-written templates, and reply rate was flat.

Meanwhile the autonomous-AI-SDR category is bleeding trust. Artisan CEO Jaspar Carmichael-Jack said first-generation AI SDRs had a "pretty low response rate" and "relatively high churn," and acknowledged that Artisan had "extremely bad hallucinations" when it first launched. This piece is the operator's guide to what to hand the model — and what to keep on the human side of the desk.

Why the aggregate "AI wins" stat is misleading

Every vendor benchmark you've seen quoted this year traces back to two datasets: Belkins/Expandi's 20M-attempt study and Expandi's own 13.2M-request H2 report. The Belkins headline is the one that gets screenshotted: including a personalized message in a connection request significantly boosts the reply rate (9.36%) compared to no message (5.44%), and AI-driven first messages result in a higher response rate (4.19%) compared to non-AI messages (2.60%), but follow-up messages perform slightly better without AI.

That 4.19% vs 2.60% is a 61% relative lift. It's also a between-account comparison — meaning the accounts running AI-drafted openers are being compared against accounts that aren't. Those aren't the same operators. AI adopters skew toward better-resourced teams with cleaner targeting and warmer sender profiles.

Expandi's follow-up study controlled for this. Same accounts, same ICP, AI on vs AI off. The AI lift disappeared. That's the number that should be on your slide, not the 61% one.

When you compare the same sender running both stacks, the AI premium priced into 2025 benchmarks has largely evaporated in 2026.

The LinkedCamp Q2 2026 dataset: what we looked at

We segmented ~1.2M messages sent by LinkedCamp customers between April 1 and June 30, 2026 into three cohorts:

1. Fully AI — opener drafted end-to-end by an LLM using profile scrape + company enrichment, no human edit before send.
2. Hybrid — AI does the research and drafts a first pass, human rewrites the first line and closes the message. Median edit time: 47 seconds per prospect.
3. Pure human — SDR writes the opener from a research doc. No LLM in the loop.

Each cohort was cut by ICP seniority (IC/manager/director/VP/C-level), industry vertical, sequence position (connection note, DM 1, follow-up 2, follow-up 3), and sender trust tier. We excluded any account that ran fewer than 400 sends in the period to keep the noise floor honest.

Benchmarking context matters here. Expandi analyzed 13,218,869 connection requests, 6,730,447 outbound messages, and 3,766,161 accepted connections sent through 13,302 active LinkedIn accounts between May 2025 and April 2026, with platform-wide averages of 28.5% connection acceptance, 3.0% connection-note reply, and 10.4% message reply. Our platform averages sit within a point of Expandi's on all three metrics — meaning the cohort math below is comparable to the industry-wide picture.

The reply-rate table: AI vs hybrid vs human

Here's what 1.2M sends look like once sender and ICP are held constant. Numbers are message reply rate (post-connection DM 1), Q2 2026:

| Segment | Fully AI | Hybrid (AI + human edit) | Pure human | |---|---|---|---| | IC / Manager | 8.9% | 11.2% | 10.4% | | Director | 6.1% | 9.7% | 9.9% | | VP | 4.4% | 8.8% | 10.6% | | C-level | 2.7% | 7.9% | 11.3% | | SaaS (all seniority) | 4.1% | 7.2% | 7.8% | | Legal & professional services | 5.9% | 10.1% | 12.4% | | Recruiting / staffing | 12.1% | 14.8% | 15.2% | | Follow-up 2 | 3.2% | 5.4% | 6.8% | | Follow-up 3 | 1.9% | 4.1% | 5.7% |

Three patterns jump out.

First, hybrid beats fully-AI in every single cell. The gap widens as you move up-market — at IC level fully-AI is only 2.3 points behind hybrid, at C-level it's 5.2 points behind. Senior buyers pattern-match to LLM output faster than junior ones.

Second, pure human beats hybrid at C-level, in regulated verticals, and on follow-ups 2+. This aligns with Belkins' finding that the legal and professional services sectors show the highest response rate at 10.42%, while software and SaaS industries have the lowest at 4.77%. The verticals where trust matters most are also where AI-drafted copy pays the biggest penalty.

Third, follow-up performance is where autonomous AI SDRs quietly fall apart. Fully-AI follow-ups reply at roughly half the rate of human ones. That matters because the first email in a sequence delivers the highest per-step reply rate (0.59%), but follow-up emails collectively account for 58.6% of all replies — stopping at step 1 leaves the majority of responses on the table.

Where fully-AI still wins (and how to use it)

Fully-AI isn't dead. It wins on three specific jobs:

• High-volume top-of-funnel to IC-level ICPs. 8.9% reply is competitive with what most manual SDRs pull, at roughly 1/40th the human time per send.
• Signal-triggered openers where the signal itself is the personalization. A job change, funding round, or product launch reference gives the model a real fact to anchor on. Hallucination risk is low because the input is a structured event, not a guess about the prospect's priorities. Our 5-minute research framework breaks down which signals actually convert.
• First-touch connection notes under 200 characters. LinkedIn's character cap forces the AI to be brief, which happens to mask its usual tells (list-of-three, "I noticed that...", closing question).

What it fails at is anything requiring inference. Fully-AI writing a follow-up to a VP who ghosted the first message is where the "AI slop" reviews come from. The most consistent criticism involves generic, robotic output — G2 reviewers describe messaging as "AI slop" that is "overpromising and underdelivering," and another G2 review rated 1.5/5 stated the platform produced "zero quality leads" with "messaging that is extremely bland."

The hybrid workflow that actually maximizes reply per hour

Hybrid is the winner on almost every reasonable ROI cut. The question is what the human touches and what the AI keeps.

Based on what our top-decile senders (by reply rate) actually do — reconstructed from timestamped edit logs on ~180K hybrid sends — the pattern is consistent:

1. AI drafts the research summary, not the message. Feed the LLM the prospect's last 90 days of activity and ask for three specific hooks. Reject any hook that couldn't be verified from a URL.
2. Human writes the first line. This is where 80% of the reply-rate gain lives. First lines are what the recipient sees in the notification preview.
3. AI writes the middle. Value prop, one-line social proof, transition. This is templatable enough that the LLM doesn't hurt you.
4. Human writes the CTA. The ask is where AI defaults to "quick 15 min?" — which now triggers pattern-match rejection in senior inboxes.
5. Total human time: 40–60 seconds per prospect. Any longer and you're not getting paid for your effort vs pure-human writing.

This is roughly what Expandi itself now recommends: sales reps can boost response rates by using AI-generated messages as a starting point and then personalizing the copy to be more human, contextual, and unique. The data validates the workflow — the tricky part is enforcing it when reps are under quota pressure to just hit send.

The saturation problem: AI reply rates are decaying month over month

Here's the finding that should scare anyone who bought an autonomous AI SDR seat in 2025. Within our Q2 2026 data, fully-AI reply rates dropped from April to June across every seniority band. C-level dropped from 3.4% in April to 2.1% in June. IC dropped from 9.6% to 8.3%.

Human-written and hybrid rates held flat over the same window.

This is inbox saturation. As more senders route more AI-drafted copy to the same personas, the pattern-match cost of sounding like an LLM grows. Expandi's 2026 analysis of 13.2 million data points found connection-request reply rates fell from 3.5% in May 2025 to 2.2% in April 2026 as low-effort automation saturated the channel. The people you're messaging have gotten faster at spotting robots, not slower.

The corollary matters for how you budget: any 2025 ROI model built on 2025 AI reply rates is already stale. Rebenchmark quarterly, not annually. And read our AI SDR renewal cancellation piece for what that means for tools already up for renewal.

The G2 signal: buyers are voting with cancellations

The headline autonomous-AI-SDR products are getting the reputation the data predicts. One G2 user characterized the output as "AI slop" with "overpromising and underdelivering," a 1.5/5 review cited "zero quality leads" with "messaging is extremely bland," one sales executive reported sending 20,000+ messages and 3,000 LinkedIn requests resulting in zero meetings, and another user documented 1,400 emails with 0 responses.

The hallucination risk is not theoretical either. A 2026 vendor review cited a 12 to 18 percent hallucination rate in AiSDR's outputs. One in six messages inventing something about your prospect is a brand-damage event, not an efficiency gain.

And the platform-compliance angle is real. TechCrunch reported that LinkedIn temporarily banned Artisan for roughly two weeks, clarifying that the issue involved trademark and data-sourcing concerns rather than AI-agent spam. If your outbound engine can be taken offline by LinkedIn's legal team for two weeks, that's a pipeline risk your board should be pricing in.

What to do this week

A concrete framework for RevOps leaders deciding how much of the message stack to hand to AI:

1. Segment your ICP by AI-tolerance. If most of your ACV comes from VPs and C-level in regulated industries, cap fully-AI at 20% of sends this quarter. Route the rest through hybrid or human.
2. Rebuild your first-line prompts to require a URL-backed fact. No LinkedIn post reference, no funding-round citation, no job-change trigger = no send. This is where hallucinations die.
3. Move follow-ups off fully-AI immediately. The reply-rate cliff on FU2 and FU3 is severe enough that you're bleeding pipeline. Use AI for research, human for copy.
4. Track reply rate by cohort, not campaign. Same-sender, same-ICP, message-stack-A vs B. Any vendor benchmark that doesn't control for sender is directionally useless.
5. Set a decay budget. If your AI cohort's reply rate drops more than 15% quarter over quarter, that persona is saturated. Rotate to a new ICP or switch that segment to human writing.
6. Audit your stack against LinkedIn's Q1 2026 enforcement pattern. Our Q1 2026 restrictions breakdown covers which tool categories triggered accounts.

Inside LinkedCamp, the cohort split above maps directly to how customers structure sequences: AI-drafted top-of-funnel to warm segments, hybrid on primary DMs, human-only on senior follow-ups. The reply data says that mix is not a compromise — it's the current optimum.

• Between-account benchmarks (Belkins' 61% AI lift) overstate reality; within-account data (Expandi's 12% acceptance drop, our flat reply rate) shows the AI premium has largely disappeared in 2026.
• Hybrid (AI research + human copy) beats fully-AI in every ICP segment we measured across 1.2M Q2 2026 sends.
• Pure human still wins at C-level, in legal/professional services, and on follow-ups 2+ — the segments where deals actually close.
• Fully-AI reply rates decayed month-over-month in Q2 2026 as inbox saturation compounds; human and hybrid held flat.
• G2 complaints and LinkedIn's temporary Artisan ban confirm the buyer market is repricing autonomous AI SDRs downward — plan renewals accordingly.