An outreach pipeline needed deep research on every firm and a personalized opener. I replaced the expensive AI audit with deterministic evidence, then found the real risk hiding in the CRM push chain.
Measured shift
$4 per lead → approximately $0.009 per lead.
The email sequence is built and wired, but still in warmup mode while reputation and reply rates are measured.
A professional-services business needed a way to identify and reach potential partner firms at scale. The manual version was slow: find firms, research their websites, decide what kind of outreach might land, write an opener, and push the contact into the CRM.
An earlier version used an AI model to audit each firm's website and generate an opener. It worked, but at $4 per lead the variable cost would compound quickly. The openers also came out generic—the kind of language that reads like a template even when each one was generated fresh.
The harder problem was reliability. Data had to move through discovery, research, email verification, CRM upsert, and outreach sequencing across five workflows and an external CRM. Every handoff could silently go wrong.
I designed the pipeline as independent workflows with a shared data layer. Schedules are staggered to avoid rate limits, and each workflow has its own error handler.
The research workflow reads public pages, counts form fields, checks booking links, classifies the outreach channel, and writes from evidence. No AI model is needed for that stage. Two additional workflows handle alerting and staleness monitoring. The only manual step is approving which leads receive outreach—a deliberate gate.
The interesting problems were in the seams between systems: one workflow's output becoming another's input, and subtle data loss or coupling hiding in between.
The email-verification request replaced the entire data item with the API response body, wiping out the firm name, website, research scores, opener, and everything else. Three downstream nodes broke from the same root cause. The fix was tracing all three failures back to that single node and anchoring references to the last node that still held the original record.
After resetting leads for re-research, the selector kept skipping them. A second status field, set during the original research, independently blocked re-entry. Clean tests passed; real reprocessing exposed the dependency.
Sites returning 403 or 503 were marked blocked, but the next hourly run picked them up again. A retry counter capped attempts at two so unreachable sites could not consume the pipeline's capacity.
The deterministic scorer reads public pages, counts form fields, checks for booking links, and classifies the best outreach channel from evidence. Variable cost dropped to zero, while the openers became more specific because they cite what the system actually found.
The email sequence is running in warmup mode—five emails per day maximum—with reputation tracking before scale. Early open and reply rates will determine whether the templates need refinement. That work is in progress, not hidden behind a launch claim.