ICP Enrichment Pipeline: Automate Lead Enrichment with Clay

Before building anything in Clay, define exactly what data points you need to qualify a lead as ICP-fit. This is not a generic list. It should be specific to your product and buyer profile. Start with firmographics: company size range (in employees and revenue), funding stage, industry vertical, and geography. Then add technographics: what tools does your ICP use that signal fit with yours? Finally, add contact-level fields: seniority, department, and function. Write this schema out as a column list before you open Clay. Teams that skip this step end up enriching fields they never use and missing the signals that actually matter. Treat the schema as a decision framework, not just a data checklist.

Create a new Clay table and configure the input sources that will feed it. You have several options here depending on your stack. If you are pulling from a CRM, use Clay's native HubSpot or Salesforce integration to sync new leads automatically when they are created or updated. If you are working from CSV exports or third-party lists, use Clay's webhook intake to accept files programmatically. If you are generating leads from scratch, use Clay's company and people search with filters matching your ICP schema from Step 1. The key architectural decision here is whether this table is a one-time batch or a live-updating pipeline. For a real automated system, configure it as a live pipeline: new rows in, processed rows out, on a recurring schedule. This is what separates a workflow from a spreadsheet habit.

This is the core of the pipeline. For each enrichment field in your schema, add a waterfall column in Clay. A waterfall tries multiple data providers in priority order and stops when it finds a valid result. For email finding, configure the waterfall to try Apollo first, then Hunter, then Snov.io, then Clay's own email finder. For company data, try Clearbit first, then PeopleDataLabs, then Clay's web scraping fallback. The priority order should reflect both data quality and cost. Put cheaper providers later in the waterfall so they only activate when premium sources fail. For each column, set a confidence threshold. If you are email sending, only accept emails with 90%+ confidence. If you are doing research, 70% may be fine. Do not over-engineer the waterfall. Start with two providers per field, measure the fill rate after the first batch of 500 leads, then add more providers where gaps exist.

Once each lead has its enrichment data populated, add a Claude-powered AI column that synthesizes all the signals into a single ICP fit score. The prompt should reference your schema fields directly. A good structure: give Claude the enriched row data, remind it of your ICP definition (company size, funding stage, industry, tech stack, contact seniority and function), and ask it to return a score from 1 to 5 with a one-sentence justification. Score 5 means the contact matches every ICP dimension. Score 1 means clear disqualifiers are present. Score 3 means partial fit, worth reviewing manually. Make the prompt specific. Vague prompts produce useless scores. Include examples of what a 5 looks like versus a 2. Claude will apply that pattern consistently across thousands of rows. The scoring column becomes your primary sort key for outreach prioritization.

Configure the output routing so leads flow out of Clay without manual intervention. Use Clay's native export automations to push enriched contacts directly to your CRM or sequencing tool based on their ICP score. Score 4-5 contacts go straight to an active outreach sequence in Smartlead or your sequencer of choice. Score 3 contacts get created in your CRM as a deal or contact with a 'Review' tag for manual qualification. Score 1-2 contacts get added to a suppression list to avoid repeat processing. Set up a weekly notification to yourself with a summary: how many leads processed, average fill rate per enrichment field, score distribution, and how many went to each route. This closes the feedback loop and tells you whether your ICP schema and scoring logic are working or need tuning.