What 10,000 AI-led LinkedIn conversations taught us
Patterns from a year of model-driven outbound - the openers that landed, the ones the model retired itself.
We crossed 10,000 model-led LinkedIn conversations in November. With that volume comes a useful luxury: the ability to look at what actually works across cohorts, geographies and developer types rather than reasoning from a handful of standout threads. Some of the patterns surprised us. Most of them confirmed what the data was already whispering at 2,000 conversations.
Openers that landed
The single highest-performing opener class is what we internally call 'specific-and-quiet': one sentence acknowledging a specific recent project, one sentence asking a small, easy-to-answer question. No pitch. No company name in the first message. No CTA.
Reply rates on that class sit at roughly 4x the next-best opener type. The conversation only mentions Capital Edge in message three or later, after the developer has self-qualified into a real conversation about a real project.
Openers the model retired
The model has discretion to deprecate phrasings that underperform across a statistically meaningful sample. In the last twelve months it has retired - among others - any opener referencing 'capital partners', any opener using the phrase 'quick question', and any opener that started with the prospect's first name as a standalone line.
All three felt right to humans when we wrote them. None of them performed.
Conversation depth predicts booking, not reply rate
A reply rate is a vanity metric. What actually correlates with booked calls is conversation depth - how many turns the developer is willing to spend before either booking or disengaging.
Threads that reach four turns convert to booked calls at roughly 28%. Threads that stall at one or two turns convert at under 3%. The implication: optimise the second and third message for depth, not the first message for replies.
Time-of-send no longer matters much
We used to believe send-time was a meaningful lever on LinkedIn. With 10,000 conversations and proper controls, the effect is real but small - on the order of 8-12% variance between best and worst send windows. That is dwarfed by message quality, which moves the needle by 200-400%.
We've stopped optimising send-time and started spending that engineering effort on opener generation.
Ten thousand conversations is enough to start trusting the data and stop trusting the heuristics. Most of what we believed about LinkedIn outbound in 2024 turned out to be wrong, or at least much less important than message-level craft. The next 10,000 will tell us where the new heuristics are wrong too.
Capital Edge publishes one note a month on UK bridging finance, paid acquisition, and AI-led outbound. Written for brokers, by the team running the playbook.