← Back to Blog

This Clay Workflow Scores 10 Million Companies for $140. Here's the Full Outbound Playbook.

By Eric Nowoslawski, PLG Growth Leader at Clay ·

When Eric Nowoslawski ran his outbound agency, 60% of all the tasks his team did were list building. His fix was to stop fighting messy LinkedIn data and let cheap AI clean up after it. In this session he walked through the four-step process his team used to build a full target account list, how he "hill climbs" agents with Claude until a cheap model performs like an expensive one, the three-bucket framework he uses to test offers, and why he stopped building software for lead magnets. If you run outbound and your list is the bottleneck, this one is for you.



Why are company lists built from LinkedIn data so messy?

Most list building starts with LinkedIn company data, and that data is self-reported. Nobody has much reason to fill it in carefully. A real estate software company often calls itself a real estate company, not a software company. LinkedIn also auto-creates pages that nobody ever claims, so you end up with wealth management firms tagged as semiconductor manufacturers and, in one case Eric found, a public school in New Jersey tagged as luxury goods and retail.

Filters built on that data will always leak. Eric's response was to accept the mess and plan for it: use filters to get a list that is too big, then let AI throw out what doesn't fit.

What is the four-step process for building a perfect company list?

Step What you do Tools Eric mentioned
1. Reverse lookup Enrich your best customers and see how they describe themselves on LinkedIn Any data tool; Clay, the Clay CLI
2. Build the search Turn those self-reported filters and keywords into a company search Claude or Codex, Clay Audiences
3. Score with AI Scrape each homepage and have a cheap model decide yes or no Claygent, an HTML-to-text scraper, ZenRows
4. Find what you missed Ask an agent to find lookalikes of your best-fit companies Claygent, Exa, Parallel

Step 1: How do your best customers tag themselves?

Start with your current customers, or just your best ones (Eric's suggestion is highest ACV, or whatever "best" means to you that day). Enrich them and look at the industries, geographies, headcounts and keywords they use on LinkedIn.

This step tends to break assumptions. If you sell to ecommerce brands doing $10M in revenue, you might assume you need companies with around 70 employees. Plenty of brands at that revenue have about 20 people on LinkedIn. You will also find industry tags you never would have guessed.

Eric runs this as a Codex skill. He gives it a customer CSV (or a website to pull customers from), it resolves the companies, looks them up through the Clay CLI, and comes back with the filters and keywords those customers use. It also suggests additions, like a missed industry or a strong keyword, and he approves the changes before it pushes the filters. You can do the same with Claude. It is all API based, so you skip Sales Navigator entirely, and you don't need to store the data. You fetch it, analyze it, and keep the keywords.

Step 2: How do you turn that into a search?

Clay Audiences has unlimited company and contact search, so the filters from step one can go straight in. In his demo, Eric described the target in plain language ("every bank in the United States") and the agent built the search itself:

  • Industry: banking, OR financial services, OR credit intermediation
  • Description keywords: credit unions, community banks, savings bank, commercial bank, national bank
  • HQ: United States
  • Estimated employees: more than 100

He never had to read through the filter menu. The point of steps one and two is to make the search as brainless as possible and let your customer data decide the filters.

Step 3: How do you clean the list with AI?

The search will return too many companies, and that is what you want. Now you score them the way a person would: look at the homepage and decide if the company fits.

To get the homepage text, you can prompt a Claygent inside Clay (and have Claude or Codex write that Claygent for you). If you'd rather stay in Claude or Codex, a simple HTML-to-text tool like html-to-text will scrape most sites as long as they aren't behind bot protection like Cloudflare. For protected sites, Eric's pick is ZenRows, the cheapest scraper he has found that gets past bot protection. ZenRows also has a crawl API if one page isn't enough context.

Then pass the page and your ICP definition to a cheap decision model. It returns a yes or no plus a confidence score. Review a sample by hand, and send anything under roughly 70% confidence to a more expensive model for a second look.

How cheap is it? Eric's math: rescoring every company in the United States in Clay's dataset, about 10 million companies, with no filters at all, would cost around $140 in credits. Your filters cut that down a lot, so in practice it costs far less.

Step 4: How do you find the companies your filters missed?

Some companies have no LinkedIn page or no Google Maps listing, but they do have a website that explains what they do. Once you have a scored list, take a company that is a perfect fit, give an agent like Claygent, Exa or Parallel your ICP definition, and ask it to find 10 more like it. Throw away any domain already in your database. Run anything new through the same scoring step, keep what fits, and drop the rest.

Eric was candid that this step only adds the last few companies. Covering 80% of your market takes about 20% of the effort, and the last 20% takes far more. His agency did it because clients judged them on list completeness. For most teams standing up a new motion, a clean list of 10,000 to 20,000 companies is plenty to start. Treat step four as optional until you run out of accounts.

What is a hill-climbing agent?

This is the part Eric thinks too few GTM teams are doing, and he called it one of the best returns you can get for very little effort.

The idea: Claude is the smartest model you have, and on a $100 or $200 a month plan it is cheap intelligence. But you can't run it across your whole market. Cheaper tools like Claygent, Exa and Parallel can run at scale. Hill climbing means using Claude to tune the cheap tool's prompt until it performs almost as well. (The term is what Anthropic and OpenAI people use for it now. Six months ago, Eric said, it was called auto research.)

How does the process work, step by step?

Eric's most recent project was teaching Claygent to find a business owner's name when the business has no LinkedIn presence. Here is how he did it:

  1. Build the answer key. He had Claude find 1,000 companies with no LinkedIn company page where the owner, CEO or founder is named on the website, for example on the About Us page. The point was an agent that doesn't depend on LinkedIn for contacts.
  2. Check the answer key. If you have both Claude and Codex, have one build the dataset and the other verify it, so you can trust your benchmark.
  3. Let Claude iterate. Point Claude at the Claygent, Exa or Parallel API and tell it to find the prompt with the best coverage at the lowest cost, and to keep tuning until the gains stop.

Eric didn't write any of the final prompt. Claude tested version after version until it stopped improving.

What did hill climbing actually deliver?

Metric Result
Claygent cost per lookup From half a cent down to a tenth of a cent
Claygent coverage From about 60% to 85% (86 to 87% on the benchmark)
Exa coverage 94%, at about half a cent per lookup
Claygent first, Exa only on misses 96% overall

The waterfall is the lesson here. Exa found more people but costs about five times as much, so run the cheap agent first and only pay for Exa when it comes up empty.

One cost tip Eric caught in his own prompt: model providers discount cached inputs, roughly 90% off, but only when the first block of text is identical across calls. Put the static instructions first and the variable data (like a list of target business types) at the bottom. He also noted that the business types listed at the top of his prompt came from the specific client he built it for. For a general agent, deleting that part would probably improve results.

Does hill climbing work for messaging too?

Yes, with a caveat. Eric wouldn't hill climb directly on emails going to real prospects. Instead, build a grading system, give the AI the company website and whatever custom data you have, and keep improving the generated email against your rubric until you're happy with it. Your answer key can be the messages your reps have already written by hand. It takes a lot of reading and checking, but it works well for one-to-one copy.

What if your targets don't have websites or LinkedIn?

The same reverse-engineering logic applies to other sources. For companies that live on Google Maps, look at how your customers show up there and search the same way, even ZIP code by ZIP code. Eric shares a GitHub repository with skills for this, including every ZIP code in the United States: Cold Outbound Skills. If companies have no website at all, your options shrink, and a rep researching by hand would struggle too. For regulated industries, state government registrations are the next place to look.

For contacts, take customers where you already know who works there and have Claude find the public evidence that links each person to the company. Where it can, you have an answer key, so hill climb an agent like Exa or Parallel to find more people the same way.

Is outbound dead?

Eric doesn't think so. What changed is buyer awareness. Ten years ago, an email offering bookkeeping might land with someone who had no idea how to find a bookkeeper. Now they have endless options and will ask ChatGPT the moment they need one. If you interrupt someone's day, lead with value.

How should you test outbound offers?

Eric puts every messaging test into one of three buckets.

Bucket What it is Example from Eric's agency
1. The direct ask A CTA you'd be thrilled to get a yes to Offer to walk a PR company through how the agency generates 100 leads a day for PR firms
2. A little friction for you More work on your side, more yeses from them Book three meetings on their calendar for free before any sales call
3. The uncomfortable one Something you hope isn't the only thing that works Pay them $100 to $200 to take the meeting

Bucket two was his agency's best performer for companies under 100 employees with at least two salespeople. Larger companies didn't care about three free meetings. For them, the agency built what it called a growth hub: custom research for the prospect's team, including mobile numbers, everyone in their ICP newly hired in the last 90 days, warm overlaps (people who live in the same city or used to work at the same company as one of their reps), and open jobs. Prospects could download a sample. Anyone who downloaded got a follow-up, and anyone who requested it got a question about whether they had a plan to use it, which opened the sales conversation.

Why test a paid meeting at all?

Bucket three is a diagnostic. If you can't get a yes even when you're paying for someone's time, you don't have message fit with those cold prospects. It cuts through the usual doubts about whether they even saw the email.

Eric spintaxes the amount: $50, $75, $100, $150 or $200. What he has seen: $50 gets fewer responses, and there is no difference between $100 and $200. People who need an incentive value their time above $100 anyway. The money just tips an already curious prospect into saying yes. Some recipients legally have to receive a gift rather than cash, and you can meter a gift the same way, for example regular AirPods versus AirPods Pro. He doesn't find that people see it as a scam either, since you're sending a gift card, not asking for bank details.

He agrees that incentivized meetings are a questionable thing to optimize for. At the agency, it was mostly protection: if a client insisted buyers would love a message and nobody said yes even with $200 attached, the data settled the argument.

How many variations should you run?

Try three to five different framings in each bucket and send each bucket to about 500 prospects. If the only thing that gets replies is the paid meeting, you have two choices. Check your CAC and decide incentives are worth it, the way companies like Rippling, Deel and Ramp have. Or go back and find a better framing or a better lead magnet. One client of Eric's had a $3,000 cost per lead and an LTV around $100,000, so he suggested sending prospects $400. The math still worked.

Framing matters more than most teams think. For one advertising client, a "try a new channel where your competitors aren't" message got nothing. Rewriting it as "Is your cost per lead starting to climb past $150?", with the dollar figure filled in dynamically by industry, got five times the responses.

Why talk about outcomes instead of services?

Eric's "three free meetings" offer started as "we'll launch a campaign for you for free," which barely worked. The difference was the outcome. His advice: look at your last 10 closed customers or case studies, find their average result, and promise that in the email.

What are the two biggest lead magnet mistakes?

Building in a vacuum. Teams guess what customers will value, automate it, and then find nobody wants it. Eric has seen engineers spend two weeks building a lead magnet that no one wanted. For the growth hub, his team tested 10 different ideas and built none of them first. Write the copy, send it, and see what people respond to. Even if only three people reply, you can get back to them a day or two later, shut the campaign down, and then build the thing they wanted.

Building software instead of agents. Eric used to build custom software for every lead magnet. Software gets brittle, and edge cases are hard to fix. Now he builds a custom skill and a custom agent instead, so the deliverable isn't locked into a path that has to be rewritten every time something changes.

Do ads lift cold email replies?

In Eric's experience, rarely. Several agency clients ran LinkedIn ads to half of an outbound list (with an 80% match rate) for two weeks and left the other half cold. It never lifted email replies. The one time ads clearly helped was an ABM program spending around $20 per person, which shows how much awareness you need before it moves the needle. The only other bump he saw was when Clay announced a big funding round while he was running its outbound as a contractor, and replies jumped that day.

What he recommends instead:

  • Social listening on competitors. Run always-on enrichment on everyone who likes or comments on your competitors' LinkedIn, X or Instagram posts, filter out the interns, and reach out. He expects high reply rates.
  • Gifting. Eric thinks gifting is due for a comeback. One idea he heard from Sendoso is "viral gifting": send two dozen cupcakes to the office, so the whole department shares them and sees your logo. Sendoso costs more than ordering from Amazon, but it works out the office address versus the home address, which may justify the price.

Key Takeaways for GTM Teams

  • Let your customers write your filters. Reverse look up how your best customers describe themselves on LinkedIn, then build the search from that.
  • Pull too many companies on purpose. AI scoring is cheap enough to clean the list after the fact. Scoring every US company in Clay's dataset would cost about $140.
  • Don't chase 100% coverage on day one. 10,000 to 20,000 good companies is enough to prove the channel works.
  • Hill climb your cheap agents with Claude. Build an answer key of about 1,000 records, have Claude tune the prompt, and waterfall the cheap tool before the expensive one.
  • Test offers in three buckets, with three to five framings each and about 500 prospects per bucket. A paid-meeting test tells you whether you have message fit at all.
  • Write the lead magnet copy before you build anything, and when you do build, make it an agent, not custom software.

Want to Attend the Next Session?

GTM Clawllective is a community where go-to-market operators share how they actually use AI in the field. The live discussion stays off the record, and every recap includes the speaker's demo video. To catch the next build live, join the community.


About the Speaker

Eric Nowoslawski is PLG Growth Leader at Clay. He was employee number 11 at Clay, joining in 2022 when the company had about $100,000 in ARR, and left in 2023 to go all in on his outbound agency, Growth Engine X. The agency grew to $340,000 a month in revenue, sent 8 million emails across 60 customers, and ran outbound for companies including HubSpot, ClickUp, Intercom, Notion, Instantly and Clay itself. He has since rejoined Clay.

Most of what he teaches comes from running outbound at agency volume, where list quality and offer testing decided whether a client stayed.

Connect with Eric on LinkedIn or grab his open-source Cold Outbound Skills repo for Claude Code.

Watch the session replay: Learn from Clay: How to Generate 300 Leads a Day