Finding, enriching and qualifying leads without losing the evidence

An AI lead generation pipeline is usually drawn as a funnel: lots of companies go in at the top, a few qualified conversations come out at the bottom. That picture hides where things actually break. Most failures happen at the handoffs, when one stage passes the next a record it cannot trust: a duplicate, a stale fact, a contact with no lawful basis, a qualification nobody can explain. This guide treats the pipeline as a series of stages with explicit contracts between them, from finding potential customers on the web to the moment a person approves the first message.

Stages and what each one hands over

Name the stages and, for each, the output it must produce and the check it must pass before the next stage accepts it.

Two things stand out in this table. Automation is concentrated in the middle, where mistakes are cheap and reversible. People sit at the start, where the target is defined, and at the end, where something leaves the company. The rest of this guide goes through the stages that need the most care.

Finding potential customers on the web

Discovery works best when it is driven by the written target from stage 1 rather than by a keyword. If your ideal customer is a manufacturer with several plants that exports to Europe, describe that in criteria and look for evidence of it, rather than searching for “manufacturing companies”. The guides on matching your ICP and on finding companies by technology stack cover two common discovery strategies in depth.

Typical public sources include company websites, business registries, industry association member lists, trade event exhibitor lists, public tenders, job postings and marketplace listings. Whatever the source, keep a “found at” field on every candidate with the address and the date. It costs nothing at collection time, and it is the only way to answer, months later, why a company was ever in the pipeline.

Respect the terms of the sites you collect from. Robots rules and terms of service apply to agents as much as to scripts, and a page behind a login does not become public because your agent can reach it with someone's session. What web scraping is covers the mechanics and the boundaries.

Enrichment that serves qualification

Enrichment is where pipelines quietly bloat. Every extra field costs collection effort, goes stale and adds noise. The test is simple: if no qualification rule, routing rule or message uses a field, do not collect it.

For the fields that pass, store provenance per field rather than per record: the value, the source address, the retrieval date, the supporting passage and a status of confirmed, inferred or unknown. A company's country may come from a registry and its headcount range from a careers page, with different ages and reliability. Field-by-field enrichment is covered in enriching CRM accounts from the public web, so this page will not repeat it.

Quality gates between stages

Each handoff should run a small set of checks, and a record that fails should stop, not slide through with a warning nobody reads.

• Identity: the domain resolves, the company is still trading, and it is not a subsidiary of an account you already work with.
• Duplicates: normalize domains and legal names before comparing, and check against the CRM, not only within the batch.
• Freshness: any fact older than its field's window is re-verified or marked stale, never silently trusted.
• Conflicts: when two sources disagree, keep both with their sources and send the record to a person, instead of letting the last write win.
• Sampling: a person checks a handful of records from each batch against their evidence before the batch moves on.

Watch the counts that tell you the pipeline is healthy: records stuck as unknown, duplicates caught, disqualifications grouped by reason, bounces and opt-outs. A sudden change in any of them usually points at a source that changed, not at the market.

Consent, lawful basis and suppression

Company data and personal data are different materials. A company's address is rarely sensitive. A named person's email address and role are personal data almost everywhere, and many jurisdictions also regulate commercial electronic messages separately. Turkey's KVKK and its commercial electronic message rules, the GDPR in the European Union and similar laws elsewhere set requirements that differ in the details, so get the rules for your markets from counsel rather than from a blog post.

Whatever the jurisdiction, a few practices are almost always needed:

• Record the lawful basis and the source for every contact at the moment you collect it.
• Collect the minimum; a role-based contact is often enough to start a conversation.
• Set a retention period, and delete records that never progress.
• Check suppression and opt-out lists at the moment of sending, not only when the list was built, and carry every opt-out to every tool that holds the contact.
• Be able to tell a person where their data came from if they ask.

The last gate: approving outreach

Everything upstream exists to put a well-supported draft in front of a person. The approver should see the recipient, the message exactly as it will be sent, the evidence it relies on and the reason this company was chosen. They approve, edit or reject, and the decision is recorded with their name. Avoid blanket approvals of whole batches; approving one message should not authorize the next hundred. The per-company research behind each draft is the subject of the companion guide on researching prospects with evidence.

This is the part of the pipeline where GoCux is deliberately strict. Sending something out is one of the actions that pause for a decision by someone on the team, who can approve it, edit it safely or reject it. Exporting data, for example handing a qualified list to a sequencing tool, also pauses, with the files in scope listed. Each decision applies only to the attempt it was given for and is checked again when it executes, and a rejected step stays rejected. Decisions, tool runs and outcomes are written to a single trail you can read back later, and spending draws on visible credits with a daily limit; when that ceiling is reached, the work halts and says why. The seven action classes lists everything that pauses this way.

GoCux does not search the open web itself; web material comes in through an authorized browsing or export step, keeping its page address and retrieval date. To turn collected material into a table of qualified candidates, FindAll takes your criteria and columns, returns each row with its evidence and reports where its coverage stops.

Launch checklist

• Stages, outputs and handoff checks are written down, with an owner for each decision.
• Every candidate carries a “found at” address and date.
• Deduplication runs against the CRM, and customers and do-not-contact records are excluded early.
• Only fields used by a rule or a message are enriched, each with provenance and status.
• Records that fail a quality gate stop; conflicts go to a person.
• Every contact has a recorded lawful basis and source; suppression is checked at send time.
• Each outreach message is approved, edited or rejected by a person, and the decision is recorded.
• Replies, bounces, opt-outs and rejection reasons flow back into the target definition.

The sales solution page shows how these pieces fit together for account research.