I sell an outbound tool but only run it from Claude, here's why.

My job is to sell lemlist. Since June, I haven't built a campaign in it by hand. Here's why, in 37 campaigns and 10,947 messages, failures included.

On November 4, Claude told me my campaigns were fighting each other.

28 of them were running at the same time, all on my LinkedIn account. My account sends 20 invites a day so they queued, like ten kids waiting for one football. After a week, one play had reached 21 of its 112 leads, and an older campaign had loaded 3,509 people and reached 62.

I sell outbound software for a living, and I had built a traffic jam, but it's not even the worst thing in this article.

Below: the setup, my real prompts, the numbers, and where it broke. A field report from the guy who sells the tool.

Can you really run outbound from Claude

Yes. This year, the 37 campaigns I built from Claude got more replies per message than lemlist Lab's average of 111,043 campaigns, on email and LinkedIn, alone or combined.

My rule: every campaign starts in a Claude chat.

Claude reaches lemlist through the MCP, a connector that lets it use the tools, and every action waits for my OK. If you're scared of losing control, that's the setting that saved me.

Here's where this sits among the ways to put AI in your outbound:

The last row is how I work. The last column is the rest of this article.

How do you set it up

You need a lemlist account (the free trial works), Claude or ChatGPT, and 10 minutes.

1. Connect the two

In Claude: Customize → Connectors → Search for “lemlist” → Connect.

For ChatGPT and the others, the setup guide is here.

It then asks which team to open. Make sure to pick the one that holds your campaigns, I picked the wrong one twice and had to start over.

How to connect lemlist to Claude

2. Check where you are, without touching anything

Don't change anything. Tell me which lemlist team you're connected to and list
my 10 latest campaigns with their status.

3. Give it your context once

Mine lives in one doc, in a Claude project: ICP, offer, proof, tone, the words I never use. Every prompt starts from it, so I never explain our offer twice.

Here’s a free template if you never built your context file: Download

What does a campaign look like when you build it by prompt

Take the one I built on October 21, for 1,074 people who had downloaded an outbound playbook from one of our ads.

1. Clean the list before writing a word

Claude checked every lead against HubSpot and customer.io, then removed customers, open deals, booked demos and anyone who had already replied. 1,074 became 514.

2. Let it do the homework

Claude read about 479 personalized playbook pages, one per lead, and pulled each person's buyer, pain and buying signal into the copy. Then it split the list into four tracks, including one that writes to an individual contributor's VP.

3. Make it roast its own work

If you think AI copy all sounds the same, you're right about version 1. So before I read anything (the skills are free):

Roast this campaign with /copywriting-analyzer, /no-ai-slop and /outbound-analyst. Score it out of 10. List anything false, missing or risky before I read it.

Version 1 failed its own review. Good.

Version 1 failed. Emails landed before LinkedIn invites could be accepted, five touches instead of three, a meeting ask in every first message, links that hurt deliverability.

Version 2 opened on LinkedIn, waited up to four days, sent one LinkedIn message, then two emails with no links and one question: "want me to run your prompt and send you the campaign it builds?"

4. Review, then launch

Claude then builds everything as drafts, and I read every first message before anything goes out. On launch day I also made Victor and Nico from my team senders. I told them afterwards, in a very calm Slack message.

Me: "good to go?" Claude: "almost." Nothing launches without my OK.

How did the campaigns perform

My AI outbound results, by channel:

Same method as lemlist Lab, on 62,800,800 messages (Aug 2025 to Jul 2026). The 37th campaign is an SMS test, covered below. All 37 are listed in the appendix.

Two caveats. One campaign carries the first line: a follow-up to people who engaged with one of our LinkedIn posts, with 192 replies and 100 people tagged interested. And a reply isn't interest, let alone a meeting. On the campaign aimed at people who liked a competitor's MCP post, I reread 11 conversations: most were some version of "thanks for connecting".

My weekly check is one line, exactly as I type it:

summarize again perf + next steps

What went wrong

The traffic jam

Building went from an hour to minutes. Sending didn't: 20 invites a day is still 20. Claude found the jam, paused seven campaigns and gave unreached leads priority. Now I check sending capacity before I build.

28 campaigns, 20 invites a day. Claude did the math before I did.

Lists faster than checks

One campaign bounced 100 of its 1,349 emails (7.4%). In another, one lead had left his company before the message arrived, and another didn't know the colleagues we'd named as rivals. Claude finds people in seconds. Nobody verifies them in seconds unless you ask.

The silent skip

I gave Claude 307 people for the competitor-post campaign. 157 were skipped at import because they were already in other campaigns, and nothing told me which ones.

The like

While collecting engagers from a LinkedIn post, browser automation liked a stranger's post, then unliked it. I've stopped letting it scroll LinkedIn for me.

The texts

98 SMS to UK prospects. Zero replies.

The worst one

lemlist logged zero meetings across all 37 campaigns, because I never log them there. HubSpot keeps only each contact's last campaign, so the best I have is correlation: 4 won deals worth €1,268.80 a month on contacts from one of my campaigns, a 70-seat POC still open, and 3 new trials since October. Two of those three had also clicked one of our Google ads. I'm VP Growth. Attribution is my job.

What do I still do by hand

Last Tuesday I had 9 Claude chats between 00:42 and 23:01: recruiting, feedback for someone on my team, a job post, customer research in our call recordings and CRM, a rescheduled call, my inbox. One was a campaign. The full prompt, exactly as typed:

https://www.lennysjobs.com/next50 I want to sell lemlist to every company out there with lemlist MCP find people who are responsible for outbound with /lemlist-context Find people, enrich, draft campaign

Claude found the outbound owners at 50 startups and drafted the campaign in lemlist. Then I sent the 33 emails from Gmail, one by one. Claude then pointed out the draft we should send them a second time.

So, by hand: my one-line cold emails, the last read of every first message, and the launch button. My job is supposed to be 80% team and 20% execution. Claude does most of the 20%.

Your first 10 minutes

1. 1.

   Connect lemlist to Claude and run the read-only prompt above

2. 2.

   Then ask it to roast your best campaign before you build a new one.

Connect lemlist to Claude, free for 14 days

Erwan, VP Growth @lemlist

P.S. If you try it, DM me your campaign list. I reply to everyone.

Appendix: every campaign I built from Claude

Pulled from lemlist on 11 Nov 2026. Reply rate per message is shown from 30 messages up; below that, raw counts. Paused campaigns stay in.

Launched, paused before any message (5): Peer swap with growth leaders APAC; CA; US 500+; US 100-500; UK 100-500.

Built, never launched (13): Lenny's Next50 outbound owners; construction firms in the Paris region (for a colleague); post-event follow-up template (for a colleague); 1:1 ABM sprint (account F); users of competing CMS tools; no-reply follow-up; partnership pitch to a hardware brand; SDR managers, sales managers and BDR managers who engaged with a post (3 campaigns); an event follow-up; Cold Call Arena account E (reps and managers).