By I run a company that automates LinkedIn outbound, so you'd expect me to tell you to automate all of it. I won't. About half of outbound should stay with a human, and it's the half that decides whether the campaign works.
Some context on where this comes from. We published benchmarks from real Linkdify campaigns run in August and September 2026: 8,400 connection requests and 2,511 messages. Average acceptance was 25.2%. Reply rate was 14.6%. Those are averages, and averages hide the campaigns that did far better and far worse. When I look at what separates them, it's rarely a software setting. It's usually something a person did before the automation started or after it finished.
So here are the five things I'd keep.
1. Deciding who goes on the list
Software can pull 500 people matching a title filter in seconds. Whether that filter describes people who have your problem is a judgment call, and it's the most expensive one to get wrong. A bad list gives you a bad week of sending before you notice anything.
My rule: before a campaign starts, open the first 30 profiles and look at them. Actual profiles, one by one. If you wouldn't personally message a good chunk of them, the filter is wrong, and no amount of clever copy downstream will rescue it. This takes fifteen minutes. Most founders skip it because the list looks right in a spreadsheet.
2. The offer
AI is good at rephrasing an offer. It has no idea whether you have one.
The test I use is simple. Say out loud, in one sentence, why this specific kind of person should care about what you do. If you'd feel awkward saying it to a stranger at a conference, it will read just as awkwardly in a DM, only now it's going to 20 people a day. The model will happily polish a vague pitch into a well-written vague pitch. Founders blame the tool, rewrite the prompt, and never touch the sentence that was broken from the start.
Write the offer yourself. Then let the software handle the variations.
Want LinkedIn outreach that runs itself?
Linkdify finds your leads, writes personalized messages, and sends them at safe, human-paced limits.
3. The decision to say nothing
This one surprised me. In our data, connection requests sent without a note were accepted 48.6% of the time. Requests with a note: 8.6%.
A caveat, because I'd rather be honest than dramatic: this wasn't a controlled experiment. People who attach notes may also be targeting colder audiences or writing pitchy notes. But a gap that size is hard to explain away, and it matches what a request with a note looks like from the receiving end. It looks like a sales email wearing a name tag.
The funny part is that the connection note is the first thing people ask AI to write for them. The most requested piece of automated personalization is the piece that seems to work best when it's blank. Knowing when a step shouldn't exist is a human call. A tool will generate whatever you ask it for.
4. Every reply
Across those 2,511 messages, the median reply arrived in about 14 hours, and most replies landed within the first day. That means the replies show up in a tight window, and each one is a person who chose to answer a stranger.
I would not let software answer them. Let it draft if you want. But the questions people ask in a first reply are the ones only you can answer: whether you'd handle their odd use case, what it costs for a team their size, whether you know their competitor. A wrong or generic answer here wastes everything upstream that earned the reply. At a 14.6% reply rate you are not drowning in conversations. A solo founder sending 20 requests a day can answer every one in a few minutes over coffee.
5. The weekly read-through
Dashboards tell you the reply rate. They don't tell you what people said.
Once a week, read the actual text of every reply, including the ones that say no. The rejections are the useful part. "Not my area, try our ops lead" means your list is off by one job title. "We already use something for this" means your offer isn't saying what's different. "Who is this?" means your profile isn't doing its job. Those are three different problems with three different fixes, and they all show up as the same number on a chart.
I build reporting for a living and I'll still tell you: the report is where you find out that something changed. The inbox is where you find out why.
What I'd hand over without a second thought
The sending. The pacing. The follow-up that goes out on day four whether or not you remembered. Tracking who accepted and who didn't.
Humans are bad at this work in a specific way. We do it in bursts. Forty requests on a motivated Monday, nothing until Thursday, then a guilty catch-up on Friday. Steady, modest daily volume is boring, and boring is what software is for. This is the part of Linkdify I use most on my own outbound, and I don't miss doing it manually at all.
Want LinkedIn outreach that runs itself?
Linkdify finds your leads, writes personalized messages, and sends them at safe, human-paced limits.
The question to ask before automating a step
Ask what happens when the step is done badly at volume.
If a step is repetition, doing it by hand just means doing it late and unevenly. Automate it. If a step is judgment, automating it means you get the same bad judgment twenty times a day, delivered on schedule, to people you only get one shot with. Keep that one. The founders I see getting replies have drawn that line in the right place, and most of them spend less than half an hour a day on the human side of it.