Want the free template? Reach out to me on LinkedIn and I’ll send it over.

Here is my LinkedIn profile

If you’ve ever tried to build a list of SaaS companies using Clay, you’ve probably run into the same frustrating problem I did: the AI agents simply don’t work at scale.

Sure, they might correctly identify 7 out of 10 companies in your test batch. But when you scale to hundreds or thousands of companies? The accuracy falls apart. You end up with tech-enabled consulting firms, staffing agencies, and marketplace platforms all mixed in with actual SaaS companies.

After scraping more than 100,000 companies and achieving 95% accuracy, I’ve figured out why this happens—and more importantly, how to fix it.

The Core Problem: Data Isn’t Detailed Enough

Here’s what I discovered: the primary industry descriptions and company data you get from LinkedIn, Apollo, or Clay’s native enrichment simply aren’t detailed enough for AI to make accurate decisions.

Let me show you what I mean with real examples:

• IT Services & IT Consulting companies that started as global services firms but now generate most revenue from a software product

• Software companies that are actually tech-enabled consulting firms

• Technology, Information & Internet companies—a category so broad it’s essentially useless (you can’t just exclude it because many legitimate SaaS companies fall into this bucket)

Take hiring platforms as an example. A company’s LinkedIn description might say “hiring platform” and look exactly like a SaaS description. But when you visit their website, you discover they’re actually a staffing service that hires people from different countries—not a pure SaaS company you’d want to target.

The Solution: A Multi-Agent Approach

I spent a lot of time talking to AI (specifically Claude) to understand exactly how it determines whether a company is SaaS or not. Through this process, I identified the key information AI actually needs:

• Detailed explanation of what they offer

• Core service or product

• Main revenue model (out of all revenue streams)

• Where their revenue comes from

• Who they target

• How they serve their target customers

The trick isn’t getting one AI agent to figure this all out at once. You need a systematic, multi-step approach.

The Two-Agent System That Works

Agent 1: The Offering Agent (Claude)

I tested multiple models and approaches. Claygent Argon won for quality and price ratio, even though many people use GPT-4o mini for enrichment.

This first agent goes to the company website and extracts specific information:

• Is it subscription-based?

• Detailed description of their offering

• Core service or product

• Revenue model

• Target customers

I use Claygent Argon here because I need consistent, detailed reports—not just quick answers.

Agent 2: The Scoring System (GPT-4o Mini)

This is where the magic happens. I built a point-based scoring system that Claude analyzes to determine if a company is truly SaaS.

Here’s how the scoring works:

Core Definitions Matter I had to create extensive definitions to catch edge cases:

• Staffing agencies (even tech-enabled ones) aren’t SaaS

• Financial services and real estate tech companies often aren’t true SaaS

• Marketplaces require special consideration

• And more…

Pattern-Based Scoring AI is excellent at catching patterns. I created a scoring rubric based on specific keywords and patterns:

• Plus points for SaaS indicators: (examples - more in the prompt)

  • “Web-based platform” (+3 points)

  • API architecture mentioned

  • Subscription or recurring revenue models

• Minus points for non-SaaS indicators:

  • “Hourly consulting rates” (-10 points—major red flag)

  • “Ongoing service relationship” (consulting, not software)

  • “Project-based” work (-8 points)

The scoring is weighted intentionally. For example, anything mentioning hourly billing gets -10 points because it strongly indicates a consulting company rather than SaaS.

The Secret Sauce: Examples

This is the most important part of the entire system: you need to provide examples.

I worked with Claude to structure my expected responses as JSON, then included multiple examples in the prompt showing:

• The company data

• The analysis

• The final score

• The reasoning

AI learns from observation. These examples are arguably more important than the scoring system itself.

I included numerous examples to ensure consistent results across different company types and edge cases.

The Results

After all this work, I use a simple threshold: anything scoring above 5.5 is considered a SaaS company.

You can adjust this threshold based on your own tolerance for false positives vs. false negatives.

With this system, I’ve consistently achieved 95% accuracy across 100,000+ companies. (I aim for 95% rather than 100% because I probably make mistakes as a human too when manually verifying.)

Key Takeaways

If you’re building SaaS company lists in Clay:

1. Don’t rely on single-agent solutions—they don’t work at scale

2. LinkedIn/Apollo data alone isn’t enough—you need to enrich with website data

3. Build a scoring system—binary true/false doesn’t capture the nuance

4. Define your edge cases explicitly—staffing agencies, marketplaces, tech-enabled services

5. Use examples—this is the most important part of prompt engineering

6. Test before scaling—verify with 50 companies before running thousands

Want the Template?

I’m sharing this Clay table and complete prompt templates for free. Just reach out to me on LinkedIn and I’ll send them over.

Connect with me on LinkedIn

Have you struggled with SaaS verification in Clay? What’s been your biggest challenge? Drop a comment below.