Scaling lead research without scaling manual research

Before anyone can send a credible first message to a local business, someone has to run a small research project. Search the name, work past directory listings and similarly named companies, and confirm the site belongs to the right business in the right city. Open it and look for what matters: a working inquiry path, a usable mobile layout, a named owner, an email address that reaches a person. Decide whether it deserves a message, then write it and log it.

In the modeled manual baseline the chain costs 18.4 operator minutes per lead, with inspection and enrichment the largest block. At a few dozen leads a month that is tolerable. At a thousand, it is a job.

Running every lookup, crawl and model call on every imported row trades one cost for another. Spend goes to duplicates and businesses with no usable site. Worse, a wrong website match fails quietly: contacts come from the wrong company, audit findings describe someone else's site, and the draft email is confidently wrong.

Volume is not the underlying problem. Applying the same effort to every record, regardless of what is already known about it, is.

Varilead orders its work by cost and by consequence. Six implemented stages anchor their state to the same persisted lead model (PostgreSQL through Prisma), keeping search candidates, audit findings, selected contacts and CRM activity tied to the business they describe.

The first stages are cheap and deterministic: CSV rows are mapped into the lead model, then names and phone numbers are normalized and checked for same-phone duplicates before anything costs money.

Discovery is where uncertainty enters. Varilead queries SerpApi, stores the search results and ranks website candidates instead of keeping only a winner. A reviewer sees what the system considered, and the decision stays auditable.

The website is the key every later stage depends on, so it is the first gate. Today every site is confirmed by a person. The ranking score shortens that review without replacing it: it is a heuristic rank, not a calibrated probability, and the system treats it that way. A score of 91 earns a faster check, not an automatic pass.

Only after confirmation does Varilead spend on HTML fetches and Chromium-based site checks. Audit findings and operator notes become an outreach angle, and the model-generated outreach draft is created only after a person has qualified the lead. One limitation is worth stating: qualification does not yet gate every crawl.

People keep three decisions: identity, contact choice and outreach judgment. Ambiguous records get additional investigation. Routine extraction runs in the background, which is where the largest block of manual minutes used to go.

The trace follows EX-1042, Cedarline Heating & Air, a synthetic HVAC contractor in Columbus, Ohio. The business is invented; the stages, fields and decision points are the implemented workflow, and elapsed times are illustrative.

At import, the name and phone number are normalized and checked for same-phone duplicates. By 00:14 discovery has ranked three candidate websites, with cedarline.example on top at 91/100. At 00:44 the operator confirms it against the name and location. Nothing is extracted before that decision.

By 01:19 the site checks return three findings: no contact form detected, a missing mobile viewport tag, and visible phone and service details. Morgan Ellis, the owner, comes from the About page as a high-confidence contact. Taylor Reed's service inbox stays a candidate, flagged because the link between a person and a role address needs review. The uncertainty stays on screen instead of being resolved silently.

At 02:34 the operator rates website quality 2/5, chooses a site-issue angle and saves the lead as ENRICHED; Varilead does not compute that rating. At 03:13 a three-email draft sequence is stored. Editing, approval and sending stay with the operator.

The capacity model asks what this workflow demands at 500, 5,000 and 25,000 leads a month, using the economics baseline: 8.0 operator-equivalent minutes per lead, including blocking wait, and 120 productive hours per operator per month.

In the solo scenario, 500 leads produce 3,300 automated task executions and a $101 API and compute budget, plus a $100 fixed allowance. Every lead still needs review, and 100 need extra investigation: 67 operator hours, 58.3 of them active review, or about 25 leads a day for one person.

The human side sets the plan: 667 operator hours at 5,000 leads and 3,333 at 25,000. Manual research grows by 18.4 minutes per lead; this workflow grows by 8.0, and most of those minutes are decisions rather than searches. The model names operator judgment as the likely constraint and recommends batches small enough that confirmation and draft review actually finish.

The larger scenarios need execution infrastructure beyond today's workstation workflow, and throughput, budgets and review-time assumptions still need validation through a timed operator study.

The economics model prices a 1,000-lead month at $35 per operator hour, from list sourcing through draft approval.

Manual research models at 18.4 minutes per lead, or 307 operator hours. The Varilead-assisted path models 8.0 operator-equivalent minutes, including a one-minute blocking-wait allowance: 133 hours, of which 116.7 are active review. That releases 173 hours a month, 56.5% less operator time.

Inspection and enrichment shrink most because extraction moves to the background; finding and verifying shrinks less because confirmation stays human. The assisted path also adds exceptions and wait, the cost of running review as a queue.

Manual labor models at $10,733 a month. The assisted total is $4,969: $4,667 in labor, $202 for API and compute, and $100 for hosting and tooling. The modeled difference is $5,765 (53.7%), or $10.73 versus $4.97 per input lead. Software is about 6% of the assisted cost. That is why the architecture is tuned to save human minutes rather than API calls.

Released hours are capacity, not banked cash. Per operator, modeled throughput moves from 391 to 900 leads a month. Whether that becomes lower cost or more pipeline depends on the team.