Learn / Lead lists
The Clay cold email workflow behind 100+ calls a month
Clay books us over 100 sales calls a month, and most people still think it's a fancy spreadsheet that happens to integrate with AI. Used right, it's the data layer of a cold email operation: it finds, cleans, verifies, and researches information at a scale no human team matches by hand.
One thing up front, because the discourse got silly: the hype around Clay hyper-personalization is complete BS. List precision plus a real offer win, and the "saw your post" trick stays dead. What Clay earns its keep on is everything below: the boring data work, and research that makes an email genuinely relevant. This is the full workflow from the training, step by step.
Step 1: source the list in Apollo, not Clay
You can build lists inside Clay, but a targeted first pull is still easiest in Apollo: decision-maker titles (owner, founder, CEO, director), person and company location, an employee band that matches your ideal customer, then stacked keywords with exclusions for the companies that sneak through. Include every email status, even "unavailable"; the waterfall in step 3 exists precisely to rescue those rows. Export the result as a CSV and bring it in.
Step 2: clean and normalize
Apollo data is messy. Company names arrive as "Acme Logistics LLC" and "BRIGHT INC", and if those strings land in your copy as variables, the email reads like a mail merge from 2011. First enrichment in Clay: normalize company, run on every row. Thirty seconds of setup, and every downstream variable is usable.
Step 3: waterfall the missing emails
Filter to rows where email is empty: these are contacts Apollo knew everything about except the inbox. Add a work-email waterfall, which queries data providers sequentially until one finds the address, then stops. Order the providers you hold API keys for first, turn the providers' own validation off (you'll verify once, properly, in the next step), run it, and merge the found emails into one column.
Step 4: verify before you spend another cent
We verify through MillionVerifier, wired into Clay over its HTTP API since there's no native integration: one GET request per row, filter to result equals ok. In the training run, 900 of 1,540 scraped emails came back deliverable. The other 640 would have been bounces, and bounces are how good domains die. Everything after this step runs only on the 900.
Step 5: personalization that's actually research
Here's the part worth copying. Instead of complimenting a LinkedIn post, an AI column visits each prospect's website and pulls their top case study, so the copy can say "prospects like {their best client}": the exact kind of company they brag about landing. That's relevance a prospect actually feels, built from public data.
The mechanics that keep it cheap and unbroken:
- Small models. A create-or-modify prompt on a mini model runs about $0.001 per row. Web research costs more; use it only when the data isn't already in the table.
- Conditional fallbacks. Some sites have no case studies. The prompt's fallback swaps in "prospects looking for {their service}", so no row ships broken. Clay's Sculptor assistant will even build the only-run-if-empty logic for you.
- Formulas for cleanup. Stripping trailing punctuation so a variable sits mid-sentence is a formula, not an AI call. Formulas cost $0. Never pay a model to delete a period.
Personalize after you verify. Never spend AI credits on an email that bounces.
Ship it
Download the CSV or push straight to your sequencer, line the variables up with the campaign, and go. From here the work is the copy and the offer: the 3 scripts that book meetings and 15 offers by industry are the other half of this system. If you want the sourcing side fully automated too, Apollo + Claude Code turns step 1 into one prompt, and Apollo vs Prospeo covers whether Apollo should even be your database.
tl;dr: source in Apollo, normalize in Clay, waterfall the gaps, verify everything, and spend AI credits only on research a prospect can feel. That's the whole trick.
FAQ
Is Clay worth it for cold email?
For data operations at scale, yes: importing, normalizing, waterfall enrichment, verification hookups, and research columns across thousands of rows at once. For the hyper-personalization hype alone, no. List precision plus a real offer beat fake personalization, and Clay's job in this workflow is making the data precise.
What does Clay's AI enrichment actually cost?
In this run, about $0.001 per row using a small model on create-or-modify prompts. Web research columns cost more because they browse. Formula columns cost nothing at all, which is why punctuation cleanup and name formatting should always be formulas, never AI.
Do I need both Apollo and Clay?
They do different jobs. Apollo (or another database) sources the raw list; Clay cleans it, finds the missing emails, verifies, and researches. This workflow uses Apollo for sourcing because building a targeted list is still best done there, then hands everything to Clay.
Does cold email personalization still work in 2026?
The fake kind is dead: 'saw your post on LinkedIn' raises guards instead of lowering them. What works is relevance built from real data, like referencing the kind of client a prospect actually showcases. That's research, not flattery, and it's the only personalization this workflow spends money on.
What is waterfall enrichment in Clay?
A column that queries multiple data providers in sequence until one returns the email, then stops. You order the providers (your own API keys first), turn their built-in validation off, and run one verification pass yourself at the end. You only pay providers that actually find something.
We run this exact system for clients end to end, from infrastructure to qualified replies: book a callif you'd rather skip the table-building. The full training is on YouTube, with every video by topic on the trainings hub.
PS - 587 positive replies once came from one ideal buyer profile in a single month with ZERO Clay. List precision and a plain email did it. Keep that in your pocket before you build a 40-column table.