On this page
Extrovert runs its own LinkedIn outreach from three team accounts. One part-time person spends at most an hour a day on it, and the campaigns currently get a 49% reply rate and a 47% acceptance rate. Oleg Sobolev, Extrovert's founder, splits the work into three layers: automation, an AI agent, and a person who makes the judgment calls.
Personal, long-term, low effort
Good LinkedIn outreach rests on two things. It's personal: you comment on someone's posts before you reach out, and you make small talk in DMs before you pitch. And it's long-term: you keep following up for as long as the buyer needs to decide, which often means two or three months, well past the end of a two-week sequence.
The third thing is low effort, and usually you can only have two of the three. The three layers below are how Extrovert gets all three.
What sequencers get right, and where they break
Sequencers run simple if-then-else logic that has barely changed in 10 to 15 years. Agencies still choose them over AI for two good reasons:
- They're predictable. A client approves the sequence once and knows every message that goes out. When AI writes every message, the agency answers for anything the client dislikes.
- They're easy to scale. Build the sequence, load prospects, send.
They also have two problems. Prospects recognize a sequence, and reply rates show it. The exception is a short, on-point one, like a follow-up after a webinar or an offer too good to ignore. And a sequence stops at the first reply.
Real conversations don't run in a straight line. A prospect replies, goes quiet, picks up again, then goes quiet again. With a sequencer, every reply lands in the inbox for a setter or BDR to move forward by hand. When Extrovert still used sequencers for its own outreach, the team sometimes made first DMs harder to reply to, just to keep prospects inside the automation longer.
Your outreach tool should keep working after the reply, so a person can step in and then hand the thread back to automation.
Why an agent with every tool fails too
The opposite approach gives an AI agent LinkedIn access and a pile of tools and lets it run everything. Extrovert tried this as well. It breaks in three places:
- Too many small decisions. Who needs a follow-up today, which posts to comment on, when the next touch should land. The context fills up and the agent starts making poor calls, even with new models.
- Memory. Tomorrow, the agent has to remember what it said to each prospect yesterday.
- Weak common sense. Models handle big chunks of work well. Choosing the audience, the strategy and the messaging angle still needs a person.
The three layers
| Layer | What it handles |
|---|---|
| Automation | Comment and DM cadence, high-volume first DMs and follow-ups, storing prospect context |
| AI agent | Reviewing what the automation isn't sure about, acting where it can, escalating the rest. Extrovert's built-in agent or your own Claude or ChatGPT |
| You | Audience, strategy, voice and final calls |
Say you have 35 conversations due today. Most of them should go out automatically. A few need judgment, because the prospect said "not now" or their situation changed. The agent handles the ones it can and escalates the rest, so you read one thread instead of 35.
Volume belongs to the automation layer because outreach is a funnel. Even at a 40% reply rate, 60% of prospects never reply. With three follow-ups in the cadence, you send far more first DMs and follow-ups than replies.
Auto DM: first DMs and follow-ups on autopilot
Extrovert's Auto DM can send first DMs, follow-ups or both. You set how many follow-ups to send and how many days apart. Oleg prefers long gaps, such as 7, 14 and 30 days. When a prospect posts something relevant, their next follow-up comes up sooner so the message can mention it.
Messages adapt to each prospect unless you tell a template to send your snippet word for word. Each DM template has an adaptation setting, from word for word to adapting freely to the conversation, and you can describe in plain language when a template applies.
Extrovert's own campaign shows how this works. Prospects are warmed up with comments first. The first DM opens with a joke or a reference to something they posted, and doesn't pitch. If the prospect posts on LinkedIn, it also says Oleg is happy to support their content. If they don't post, that line is left out.
To shape how Auto DMs sound, edit your voice: add selling points such as your value prop, differentiators and objection handling, then add instructions to your DM templates. The built-in agent can ask you questions and write these instructions for you.
When Auto DM holds a message for review
Autopilot works until it doesn't. A follow-up gets awkward when:
- you already told the prospect in a native LinkedIn DM that this was your last message;
- a fact is missing: the thread ends with "let's talk next week", and the automation can't know whether the call happened;
- the prospect posted about something deeply personal, such as a family tragedy.
When Auto DM isn't sure, as in cases like these, it doesn't send. The message goes to the For review tab, where you or your agent decide. The aim is for automation to send around 90% of messages and park the rest.
Your agent can do everything you can do in the app
Many users already connect Claude to Extrovert through MCP and review their comment feed and DM threads with it. The agent sees everything you see in the interface and can take the same actions.
This works even with automation off. You can tell Claude: "I have 44 DM threads to review. Send the ones you're sure about and ask me about the rest." The agent can also snooze a conversation for months when a prospect says "follow up next year", move a prospect to a nurturing campaign or delete them.
Extrovert's built-in agent has the same skills and tools as an external one, and it's always there with nothing to connect. Your own agent adds two things: it probably runs a stronger frontier model, and it already has context from your other tools.
Prospect Notes as a memory layer
Prospect Notes let you attach notes to any prospect. Extrovert uses them as the place its agent stores what happens outside LinkedIn: calls, emails and CRM stages.
Without that, a follow-up after "I'll book a call" is a guess. Did they book? Did the call happen? What did you discuss? A daily agent run answers those questions. Ask Claude or ChatGPT to pull your Extrovert prospects, check your calendar, call recorder (Fathom, for example), email or CRM, update each prospect's notes, sync the state back to your CRM and suggest next steps.
The agent might come back with: "You had a call with John, and he asked you to revisit next year. Move him to a nurturing campaign?" Or: "Sally didn't book. I've noted that in Extrovert." The next Auto DM to Sally knows she hasn't booked and can offer another time.
The proactive agent
Extrovert's agent can also come to you first. Turn it on in agent settings, and it sends daily or weekly emails with two kinds of output:
- Suggestions. It looks at your campaigns, your voice, how you edit comments and DMs, and how your metrics compare with averages, then recommends changes to your strategy, campaigns or voice.
- Daily DM review. It reads everything in For review and Needs reply and proposes next steps, such as moving one prospect to another campaign or replying to another.
You can set up Claude or ChatGPT the same way. Extrovert's goal is that you run outreach through your agent and rarely need to log in.
How manual should your outreach be?
If a deal is worth more than 1 million a year, or your whole market is 500 companies, you can afford a lot of manual work per prospect. A low-ticket SaaS like Extrovert can't, so it sits near the automated end.
Settings are per campaign, so you can mix approaches: full auto in a broad warm outreach campaign, manual in a small VIP or "slow cooking" campaign for people you want to nurture yourself.
Oleg's suggested rollout:
- Start manual. Adjust the cadence and voice until the drafts are good enough to send unchanged.
- Hand the review to your agent.
- Move to full automation, for the whole flow or just the parts you trust.
How comments stay human
Extrovert uses frontier models and makes four or five LLM calls for each comment. If the post links an article or includes a video, a model reads or watches it first. Then the comment is drafted, checked for AI tells (phrasing that people recognize as machine-written) and repaired if it has any.
Your voice has three parts:
- Tone and structure: a prompt describing how your comments should read.
- Examples: comments you wrote or like. The model copies their structure, punctuation, emoji use and capitalization.
- Talking points: facts, observations and points of view that make a comment yours. People skip this part most often. If you post on LinkedIn, the setup agent can pull talking points from your posts.
Longer comments carry a little more risk of reading as automated. Oleg prefers longer comments backed by talking points that match the topics you target. When the posts you comment on are far from your talking points, short supportive comments work better.
FAQ
Can Auto DM replace a LinkedIn sequencer?
Yes. It runs like a soft sequence with a set cadence, and you control how closely each message follows your template, from word for word to fully adapted to the prospect. It also holds messages it isn't sure about for your review.
Can Auto DM open with something personal instead of business?
Yes. Auto DM sees the same prospect context you see when you write by hand. Tell the template to reference something from their profile when there's a natural hook, and to use a safe opener when the profile is blank. You don't need to switch to manual.
What's a good LinkedIn reply rate?
40% is on the high end. When Extrovert last measured across its users, a couple of months ago, the average was around 25%. Commenting first lifts acceptance: Extrovert's own acceptance rate is 56% after comments and 39% without.
How do I find a hard-to-reach ICP, such as auto body shop owners?
Some industries are thin on LinkedIn, but it's worth testing. If the client serves the whole country, keep the prospect list as narrow as your market allows; with high reply rates, a small campaign can still work. Extrovert's Lead Finder has filters Sales Navigator doesn't, such as profile badges like Premium (a good sign someone is active on LinkedIn). You can also exclude Open to Work and filter by follower or connection count. You can also ask the agent to build the search. Extrovert does all its own prospect filtering with AI now.
Can I run a campaign on my existing connections?
Yes. In a campaign, click Add Prospects and choose recent connections. There's no ICP filter on that import yet, so it works best when your connections already match your ICP. When Extrovert reactivated its own connections, roughly 30–40% of them came back into conversation.
Do I need my own AI agent?
No. Extrovert's built-in agent has the same skills and tools. Connecting Claude or ChatGPT adds a stronger model and the context your agent already has from your other tools.
Want to try this setup on your own campaigns? Start with Extrovert.
