Alright folks, I've been running a 30-day experiment with Relevance AI to automate a specific part of my workflow: LinkedIn outreach for my DevOps consultancy. I wanted to see if the "AI agents" hype could translate into tangible, quality leads without sounding robotic or burning through my network's goodwill.
My goal wasn't full automation, but to use it as a force multiplier. I focused on personalized connection requests and follow-ups for people posting about specific pain points (e.g., "our Kubernetes deployments are chaotic," "incident response takes hours").
Here's the core setup I used in their visual builder (Studio). Think of it like a CI/CD pipeline but for outreach 😅
```yaml
# My Outreach Agent Workflow (simplified)
Trigger: New LinkedIn post matching keywords (k8s, alert fatigue, deployment failure)
Steps:
1. **Scrape & Analyze**: Agent scrapes post & comment context.
2. **Personalize**: Uses a custom prompt with my background to draft a response.
- Input: Post text, commenter's role/company.
- Prompt: "Draft a helpful, non-salesy reply offering a relevant tip from a DevOps engineer perspective. Mention [specific tool/strategy]. Max 2 sentences."
3. **Human-in-the-loop**: Draft goes to my review queue in Slack.
4. **Execute**: Upon my approval, agent sends via LinkedIn (using a secure connection).
```
**The Results (Good & Bad):**
* **Volume & Efficiency:** I reviewed and sent ~120 personalized messages. Manually, this would have taken 10+ hours. With the agent doing the first draft and routing, I spent maybe 2 hours total. That's a huge win.
* **Quality of Interaction:** The initial connection acceptance rate was ~35%, which is solid. The key was my **human review step**. The AI drafts were good starters but often missed nuanced tech details. I almost always tweaked them.
* **The Pitfall - "AI Drift":** Around day 20, I noticed replies getting slightly more generic. The agent started reusing phrases. I had to go back and refine my core prompt and add more negative examples ("avoid saying 'I saw your post about...'"). This felt like maintaining a CI pipeline—constant iteration.
* **Cost:** For my volume, it was reasonable (~$120 for the month). But scaling up significantly would require careful calculation. It's not a "set and forget" tool.
**My Verdict:**
Relevance AI is a powerful orchestration layer, not a magic bullet. If you approach it like a GitOps pipeline—where you define your desired state (personalized, helpful outreach), let it execute, but always monitor logs and adjust—you can get great results. It's excellent for **sourcing** and **first-draft personalization**.
However, if you're looking for a fully hands-off solution, this isn't it. The need for prompt maintenance and human review is critical, especially in a field like DevOps where technical credibility is everything. I'll keep using it, but strictly as a controlled, monitored system.
Has anyone else tried using AI agents for technical community building or sales? I'm curious how you handle the "prompt drift" problem.
— francesc
— francesc
Interesting approach to structure it like a CI/CD pipeline. The visual builder aspect is key - it forces you to define clear triggers and transformations, which is often where these automation tools fail.
What was your "throughput" and "latency"? Specifically, how long from trigger (post detected) to the personalized draft being ready for your review, and what percentage of those drafts did you actually send vs. modify? I've found the review step is the real bottleneck; if the draft needs heavy editing, the time saved vanishes.
Also, did you track any leading indicator on response quality beyond just connection acceptances? Like reply rate to your initial comment or meeting book rate? That's where you'd see if the personalization was genuinely relevant or just superficially customized.
You're right about the review bottleneck. My latency from trigger to draft in my inbox was 7-12 minutes, which was acceptable. Throughput was roughly 25-30 drafts daily. The critical metric was my **send rate**, which ended up at about 65%. The 35% I modified or scrapped were usually because the agent tried to over-index on a minor detail in the post, making the connection feel forced.
On quality indicators, I tracked the sequence, not just the first touch. A connection acceptance alone is meaningless. The more telling metric was the **reply rate to the first follow-up message**, which was sent manually 2-3 days after connection. That held steady at my historical manual rate of ~22%, suggesting the initial personalization was effective enough to open the conversation. If it had dropped, the automation would have been counterproductive.