AI personalization in cold outreach: what actually works, and what backfires
I'm Chay. I work on growth at Prosp. I've spent a lot of time watching people argue about AI in cold outreach, and most of the argument is wrong on both sides. One camp says AI personalization is a cheat code. The other says it's slop that's killing deliverability. The data says neither.
Here's the position I'll defend the whole way down: AI personalization is neither good nor bad. It works when it's grounded in real signal, and it backfires when it produces detectable, generic slop. Real signal means the prospect's actual profile and their recent activity. Slop means a model dressing up the same template for ten thousand people.
The honest version is messier than the LinkedIn hot takes. In one large study, AI helped. In another, it hurt. Same technology, different channel, different execution. That tension is the whole point. Let me show you the numbers, by channel, and where the line actually sits.
Does AI personalization actually lift reply rates?
It can, when it's built on real signal. In Belkins' 2025 B2B LinkedIn study (over 20 million attempts), a personalized first message replied at 9.36% versus 5.44% for no message at all. AI-assisted first messages also beat non-AI ones, 4.19% to 2.60%. The lift is real, but conditional.
Look closer at what moved the number. The Belkins lift comes from a personalized message, not from a merge field. Dropping someone's first name into a template is not what produced 9.36%. A message that reflects who the person actually is, and what they're actually doing, is.
There's a second lever in the same data that nobody talks about. Sending a message and visiting the prospect's profile lifted replies to 11.87%, versus 4.88% for a DM alone (Belkins, 2025). The profile visit is a small signal of real attention, and it more than doubles the response. That fits the thesis cleanly: signal beats automation.
So the answer is yes, but only with a condition attached. The condition is where most people fail.
When does AI in cold outreach backfire?
When it produces generic output that scales detectably. In a 100,000-email paired study, AI-generated cold emails replied slightly worse than human-written ones, 4.1% to 5.2%, and got flagged as spam far more often, 8% versus 3% (Digital Applied, "AI SDR Real Performance", 2026). The spam gap was widening over time, not closing.
Sit with that for a second. The reply gap is small. The deliverability gap is not. The real risk of AI in cold email isn't a slightly worse reply rate. It's that more of your sends never reach the inbox. Spam filters are getting good at spotting machine-generated patterns at volume, and once your domain takes the hit, no clever line saves you.
Now hold the two studies side by side. AI helped on LinkedIn (Belkins) and hurt on email (Digital Applied). Same technology, opposite result. The honest read isn't "AI good" or "AI bad." It's that channel and execution decide the outcome. Email is a deliverability game where generic AI patterns get punished by filters. LinkedIn is a relevance game where signal-grounded messages get rewarded by humans. Treat them the same and one of them will quietly bleed.
What actually works in cold outreach right now?
Tight targeting, brevity, and follow-through. The boring fundamentals still outperform clever tricks. In Belkins' 2025 cold email study, sending to a single contact per company replied 7.8%, versus 3.8% for ten or more contacts at the same company. Spraying a whole org doesn't multiply your odds. It halves them.
Brevity wins. Lemlist's 2026 data shows messages around 120 words booked far better than 300-word emails. Nobody on a cold channel reads three paragraphs from a stranger. Short, specific, one ask.
Follow-ups matter, within reason. Most replies still come from the first touch, but a consistent sweet spot lands around three to four follow-ups. Past that you're annoying people who were never going to answer.
Channel choice is a real lever. LinkedIn message replies run about 10.4% (Expandi LinkedIn Benchmarks 2026, 13.2M data points), roughly double cold email's 3.4% to 5.8% (Instantly and Belkins). If your buyer lives on LinkedIn, that's where the math favors you.
If you're running LinkedIn at volume, the bigger risk isn't reply rate. It's account safety, and I wrote about staying inside the limits in how to do LinkedIn outreach without getting restricted.
What's decaying, and what to stop doing?
Merge-field personalization and brute-force volume. Token-merge "{firstName}" is table stakes now, not an edge. Remember, the Belkins lift came from a personalized message, not from a name swap. If your whole personalization strategy is merge tags, you're doing what everyone does and expecting an above-average result.
Reply rates are also in secular decline, which raises the bar every year. Belkins logged cold email dropping from 6.8% in 2023 to 5.8% in 2024. Expandi reports LinkedIn connection-note replies fell about 37% year over year. The channels are getting noisier and buyers are getting numb to outreach. What worked in 2023 is the new baseline you have to beat, not match.
Here's the uncomfortable read. The decay means generic AI makes your problem worse, not better, because it pours more average-quality volume into channels that are already saturated. The escape isn't more sends. It's fewer, sharper, genuinely-relevant messages. Targeting and signal are appreciating assets. Volume is a depreciating one.
Are AI voice notes worth it?
Maybe, but treat it as a test, not a fact. Voice notes are hyped hard right now, with "+30% to 40% more replies" claims everywhere. I want to be straight with you: I can't find a real study behind those numbers. They get repeated because they're sticky, not because they're proven.
The largest dataset I trust on this, Expandi, explicitly says it cannot yet benchmark voice and AI features. So when someone quotes you a precise voice-note lift, they're quoting a vibe, not a measurement. That doesn't make voice notes bad. It makes the specific number unreliable.
My honest take: a voice note is plausibly a strong signal of real human attention, the same category of thing as a profile visit, which we know worked in the Belkins data. That makes it worth A/B testing on your own list. It does not make "+40%" a number you should plan a campaign around. Run it, measure your own reply rate, and trust your numbers over the hype.
How should you actually use AI personalization?
Ground every message in real signal and verify it before you scale. The data points one direction: personalization that reflects the prospect's actual profile and recent activity earns replies, and generic output earns spam flags. The whole game is which kind your AI produces.
A practical checklist from everything above:
- Pull from real signal. The prospect's profile, their recent posts, a specific trigger. Not just a first name.
- Keep it short. Around 120 words. One clear, specific ask.
- Target tight. One right person per company beats ten wrong ones.
- Pick the channel honestly. LinkedIn relevance game, email deliverability game. Don't run them identically.
- Watch deliverability on email. If spam flags climb, your AI output is too generic. Pull back.
- Test voice notes, don't trust the hype number. Measure your own lift.
This is the gap we built Prosp to close. Prosp writes AI-personalized messages from each prospect's profile and recent activity, which is the signal-grounded kind the data actually rewards, not merge fields wearing a costume. It also supports voice notes if you want to run that A/B test yourself. If you're weighing options, I compared the field honestly in Prosp vs HeyReach, Waalaxy, and Dripify, and you can see the product at prosp.ai.
FAQ
Does AI personalization actually beat human-written outreach? It depends on the channel. On LinkedIn, AI-assisted first messages beat non-AI ones, 4.19% to 2.60% (Belkins, 2025). On cold email, AI replied slightly worse and got flagged as spam more than twice as often, 8% versus 3% (Digital Applied, 2026). Channel and execution decide it.
Are merge fields like {firstName} enough personalization? No, not anymore. Merge tags are table stakes that everyone uses. The Belkins 2025 lift to 9.36% reply rate came from a genuinely personalized message reflecting the prospect, not from a name swap. If merge fields are your whole strategy, expect an average result, not a strong one.
Do AI voice notes really boost replies by 30 to 40 percent? There's no traceable study behind that figure, and the largest dataset, Expandi, says it can't yet benchmark voice features. Voice notes are plausibly a strong attention signal worth A/B testing, but treat the "+40%" claim as hype, not a number to plan around. Measure your own results.
Is LinkedIn or cold email better for outreach? By raw reply rate, LinkedIn wins. Message replies run about 10.4% (Expandi 2026, 13.2M data points) versus 3.4% to 5.8% for cold email (Instantly, Belkins). But the right answer is wherever your buyer actually pays attention. Match the channel to the audience, not to the benchmark.
Why are my reply rates dropping over time? Because the whole channel is decaying. Belkins logged cold email falling from 6.8% in 2023 to 5.8% in 2024, and Expandi saw LinkedIn connection-note replies drop about 37% year over year. The fix isn't more volume. It's tighter targeting and messages grounded in real signal.
The honest takeaway
The "is AI good for outreach" debate is the wrong question. The data refuses to give it a clean answer, because AI helped on LinkedIn and hurt on email in two large studies running at the same time. The right question is narrower and more useful: is my AI grounded in real signal, or is it generating detectable slop at scale?
If it's grounded, the numbers reward you. A personalized message replied at 9.36% against 5.44% for none. If it's slop, the numbers punish you, mostly through deliverability, with spam flags more than doubling. The technology is the same. The execution is everything.
So don't ask whether to use AI. Ask whether your messages would still make sense if a real person read them one at a time. If yes, scale it. If no, no model will save you. Build on signal, keep it short, target tight, and measure your own results over anyone's hype.
Chay
About the author: Chay Patil works on growth at Prosp, focused on outreach strategy, onboarding, and content. Reach him at chay.p@everis.ai.