Skip to content
Solution 04 · Pipeline Generation

The research is what does not scale. So automate that.

Outreach fails because nobody had time to find a real reason to write. A representative can research five accounts properly in a day, or send fifty templates nobody answers. This pipeline discovers qualifying accounts from public records, researches each against real data, resolves and verifies a contact, and drafts the message — then stops, and hands it to a person.

The line this system does not cross

It writes drafts. It does not send. Every message lands in a drafts folder where a person reads it, edits it, and decides. Automating the send is what burns a sending domain and produces the outreach everyone deletes — so the architecture stops one step short, on purpose.

Emails sent without a human reading them 0
Addresses guessed from a name pattern 0

The problem

Volume and relevance trade against each other.

Every outbound team eventually picks a side, and both choices lose.

FAILURE 01

Research properly, reach nobody

A representative who genuinely researches an account — reads the filings, finds the operational pressure, identifies the person who owns it — manages five to ten in a day. The messages are excellent. There are not enough of them to build a quarter on.

FAILURE 02

Scale the template, burn the domain

The alternative is volume: buy a list, merge a first name into a template, send thousands. Reply rates collapse, spam complaints accumulate, and the sending domain's reputation degrades until legitimate mail stops arriving. The damage outlasts the campaign by months.

FAILURE 03

Automate the send, lose the judgment

The current wave of tools closes the loop and sends autonomously. That removes the one step where a person notices the research misread the company, the contact left last year, or the timing is wrong — and it is the step that costs almost nothing to keep.

Architecture

Nine stages, and a dead end at every failure.

A lead that cannot be qualified, contacted, or verified does not fall through to the next stage with a guess in place of the missing fact. It exits, labelled with why.

  1. 1 · DISCOVER Four public sources, each source record retained Public company filings · federal award records · higher-education registries · web research
  2. 2 · ENRICH Firmographic waterfall, four tiers Real size data, never inferred. Each tier is tried only if the previous one returned nothing.
  3. 3 · QUALIFY In scope, yes or no Fast model tier — the question is binary, so it is cheap to ask.
pass
4 · RESEARCH Deep model tier Operational pressure, waste estimate, intent signals, and a hook.
fail
EXIT Disqualified, reason recorded
  1. 5 · RESOLVE CONTACT Ranked provider waterfall Never a guessed address pattern. No matching of first.last@company.com.
  2. 6 · VERIFY Three-tier deliverability firewall Tier 1 syntax · tier 2 MTA probe · tier 3 catch-all scoring
verified
7 · DRAFT Four sentences, grounded in the research
unverifiable
PIVOT A second channel is drafted and saved
no person found
EXIT Nothing is sent
QUEUE FOR REVIEW A person reads it and decides The draft lands in your drafts folder with contact pinning applied and a confidence score attached. Nothing leaves without that decision.
Seven stages, three exits, and a person at the end of the only path that sends anything.

What makes it trustworthy

Four mechanisms, each closing a specific failure.

MECHANISM 01

Two model tiers, not one

"Is this organisation large enough to qualify?" is a yes-or-no question. Running it through the same expensive model that writes the research is waste — at scale it is most of the bill.

A fast tier handles the binary decisions and the short-form writing. A deep tier handles the nuanced research and the first draft, where the quality difference is actually visible. The routing is what keeps per-lead cost in cents rather than dollars.

MECHANISM 02

The deliverability firewall

Sender reputation is the asset outbound runs on, and it is destroyed by bounces. Three escalating checks stand between a discovered address and a draft: syntax and disposable-domain rejection, a live mail-server probe, and — only for catch-all domains, which accept everything and confirm nothing — a confidence score that must clear a threshold.

Fail any tier and the lead routes to a different channel instead. It is never given a fabricated address to try.

MECHANISM 03

Contact pinning

Re-running a pipeline over an account you already know is where data quietly degrades: a fresh extraction returns something plausible and overwrites a contact somebody had verified by hand.

Every contact carries the trust rank of the source that produced it. A higher rank may overwrite a lower one; a lower rank never overwrites a higher one. A manual lock blocks all automated writes permanently — so a human correction is final, which is the property that makes people willing to correct anything at all.

MECHANISM 04

Confidence scoring

Every lead is scored from 0 to 100 across four axes: the trust rank of the contact source, the deliverability verdict, how many intent signals the research actually found, and whether a profile was located.

It deliberately does not predict reply likelihood — nothing in the data supports that claim. It measures how much the system actually knows, so a queue can be worked in evidence order and the weakest leads can be discarded rather than sent anyway.

What it does not do

The limits, stated up front.

These come from production deployments, so they are observed rather than theoretical.

It does not replace a seller

It removes the research and drafting hours. The conversation, the qualification call, and the judgment about whether an account is worth pursuing remain entirely human work.

Estimates are benchmarks

When research quantifies an operational cost, it applies industry benchmarks to observable facts. It is a defensible opening figure, not an audit of that specific company, and the drafts say so.

Extraction is non-deterministic

First-time contact extraction can return different people across runs when the underlying web results shift. Pinning stops that from corrupting a record once verified, but it does not make the first attempt deterministic.

It is throughput-bound

Accounts are processed sequentially, and third-party providers impose rate limits. This is a steady daily pipeline, not a tool for blasting a list of ten thousand — which is the correct shape for the outcome anyway.

FAQ

What revenue leaders ask.

No. The pipeline ends by placing a draft in your drafts folder. A person reads it and decides whether it goes. Automating the send is the step that destroys sender reputation and produces the outreach everyone deletes — and it is the cheapest step to keep human, because reviewing a researched draft takes seconds while writing one takes half an hour.

Pattern guessing is disabled by design. Most tools infer first.last@company.com and send hopefully; that is what generates bounces and complaints. Addresses here come only from ranked commercial providers and must then clear a three-tier verification firewall. If no verified address exists, the lead routes to a second channel rather than receiving a fabricated one.

It survives, permanently. Every contact carries the trust rank of the source that produced it, and a lower-ranked source can never overwrite a higher-ranked one. A manual lock blocks all automated writes outright. This matters more than it sounds: if reps learn that corrections get silently reverted on the next run, they stop correcting anything, and the database rots.

Public records, not a purchased list. Public company filings, federal award records, higher-education registries, and web research. The advantage is that the source document is retained alongside the account, so research works from the actual filing rather than from a company name and a guess — which is what makes the resulting message specific enough to answer.

Because they are measured, not estimated. The limitations section above is specific rather than generic: those are the constraints these deployments actually hit, recorded against real run volume. The same instrumentation is what puts the per-lead cost in cents rather than dollars — model-tier routing sends the binary qualification decision to a fast tier and reserves the deep tier for research that needs it.

Bring us your worst-performing sequence.

Not the one that works. The one with a 0.4% reply rate that nobody wants to look at — because the reason it fails is almost always visible in the first sentence, and it is almost never the subject line.