The problem
He writes that every article about AI-assisted outbound reports cost per lead, and the conclusion “writes itself”: outbound is now essentially free, so send more. His cost per contactable company is “roughly three US cents.” He has come to think that number “is one of the most misleading figures in outbound, because it describes the one part of the process that stopped being a constraint.”
The blueprint (as he describes it)
- Source wide, then keep only complete records. A complete record is a company plus a named decision-maker plus an email. He asked for 500 companies and got 130–150 complete records, roughly 28%. In his words, “almost all of them died on the person, not the company”.
- Never guess an address. He keeps everything “sourced and unconfirmed until a verifier rules.” Guessing address formats is a prohibited operation in his system. Of about 119 collected addresses, 74 (62%) passed bulk verification, which left 57 distinct contactable companies.
- Spend where the customer looks. Sourcing cost about ¥8 and 25 minutes on a low-cost model. Research cost about ¥4, mostly from non-LLM APIs. Drafting is “the one place I deliberately spend more”: a mid-tier model writes the first draft and a frontier model reviews and rewrites it.
- Require a source for every claim. His drafting rules require “every factual statement to carry a traceable source”, which is how he keeps a language model “from inventing a decision-maker.”
- Cap the sends. Five mailboxes per domain, five emails per mailbox on weekday mornings, so 25 sends a day, after roughly 24 days of warmup on a new domain. Every email gets a human pass before it goes out.
- Wait seven days before judging a batch. One positive reply came four to five days after sending. His rule now: “a batch is not evaluable until at least seven days after the last send.”
- Filter before research, not after. An agent asked for 300 companies found 300 and more than 170 email addresses, then scored them and recommended seven. The research money was already spent. He moved the filters forward: geography, product-line overlap, authorised-distributor status, sanctions exposure and adjacent products are checked “before a single research call”. After a human replies, “there is no scoring at all.”
- Instrument instead of rebuilding. He tried to replace his agent with a small app that itemizes every call. After “Ten-plus versions” and three days, it still would not run end to end, and he stopped. He now adds per-stage record counts, per-stage cost and a source trace on every claim to the system he already has.
Stated results
- Yield: 500 requested, 130–150 complete records, about 119 addresses, 74 verified, 57 contactable companies. 11.4% end to end, about ¥14 in total, ¥0.25 (three cents) per contactable company.
- Speed gap: 25 minutes to generate the 57, a bit over two working days to send to them, 24 days to make a fresh domain ready. “The generation side is running about a hundred times faster than the delivery side.”
- Early batches: 128 sends brought three positive replies, the first around send 60. 44 sends brought one positive reply, a C-level contact in the Philippines who asked for the product catalogue.
- First full round: 282 emails, 4 positive replies from decision-makers (1.4 percent) and 1 explicit rejection. He compares this with a published benchmark’s effective interested-reply rate of roughly 0.64 percent.
- Planning rule: “If I need 100 companies contacted, I source 900.”
Caveats
- These are his own numbers on his own site. He calls the batch figures “running totals taken mid-experiment, not results” and “anecdotes with denominators.”
- His sending setup changed between posts: 25 sends a day in the yield post, and a ceiling of about 40 a day across five rotating mailboxes in the later 282-sends post.
- The agent framework and the models are not named.
- He compares his 1.4 percent with a benchmark built on two million emails across every sector, and says himself the comparison “is not clean.”
- Haoli Global offers a free cold email template pack for an email signup, and its disclosure page says some links may be affiliate links.
- Handles: LinkedIn in/lihao-design3 and GitHub lihao88927-arch (both in the author data on his article pages). No verified X, Bluesky, Threads, Instagram or TikTok profile was found.
FAQ
How many of 500 requested leads could Li Hao actually contact?
- Complete records came to 130–150, about 119 addresses were collected, and 74 passed verification across 57 distinct companies, an end-to-end survival rate of 11.4%.
What did the pipeline cost?
About ¥14 in total for that run, a little under two US dollars, or ¥0.25 (three cents) per contactable company. Drafting is where he spends more on purpose.
What limits his outbound if leads are this cheap?
Sending. Five mailboxes per domain at five emails each means 25 sends a day, a new domain needs about 24 days of warmup, and every email gets a human pass.
Does every email get human review?
Yes. He writes that every email gets a human pass before it goes out.
Sources
- https://haoliglobal.com/blog/ai-outbound-pipeline-yield (primary: “I Asked My AI Pipeline for 500 Leads and Got 57”, Jul 21, 2026, updated Jul 28, 2026)
- https://haoliglobal.com/blog/positive-reply-rate-vs-reply-rate (his follow-up: 282 sends, Jul 28, 2026)
- https://haoliglobal.com/about/ (author page)
Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).
Get one teardown a week.
Same format, a new operator every issue. Double opt-in.