We Replaced a $5k/Month SaaS Stack With Claude Code. Here's What Changed.

We were paying $5,200 a month in SaaS tools to run our GTM operation. Today we spend under $300 on the same workflows -- better output quality, more customization, and full ownership of the code. Here is exactly what changed, what we kept, and the tradeoffs nobody talks about when they write "we replaced X with AI" posts.

This is not a pitch for Claude Code. It is a transparent breakdown of a real migration, including the parts that were harder than expected.

---

The Old Stack and What It Cost

Before the migration, our monthly SaaS bill looked like this:

That is the tools and direct labor cost. It does not count the time spent managing these tools -- updating zaps that broke when an API changed, debugging Make scenarios, maintaining Clay tables, or training new team members on four different platforms. Conservative estimate: another 10-15 hours per month of overhead at the same rate is another $750.

Total real cost: closer to $3,500/month.

The frustration was not the money, though. It was the ceiling. Clay is powerful, but it runs on Clay's logic. When we needed enrichment to do something Clay did not support -- custom scoring based on enrichment results, cross-referencing against our own historical data, conditional branching based on company type -- we hit walls. Zapier and Make have the same problem at a different layer. These tools are designed for common use cases. Our use cases were increasingly uncommon.

---

What We Replaced and What We Kept

Before going into the numbers, I want to be clear about what this migration actually is. We did not rip out our entire stack and replace it with an AI. We replaced specific components where custom code gave us better outcomes than SaaS tools. The distinction matters.

What Claude Code Replaced

Enrichment orchestration (Clay, $500/month): The enrichment waterfall -- try source A, if no result try source B, apply verification, write to database -- is fundamentally a decision tree. Claude Code, backed by a set of Python scripts that call each enrichment API, replicates this with more flexibility and no per-row pricing. Our enrichment scripts call the same APIs Clay was calling, but with custom logic we control entirely. Workflow automation (Zapier + Make, $300/month): Most of our Zaps and Make scenarios were data transformation and routing tasks -- take output from script A, clean it, route to destination B. Claude Code with a well-configured directive structure handles these natively. The "if-then" logic that required visual workflow builders now lives in Python scripts that are faster, more reliable, and easier to debug. Manual research (20+ hours/month): This was the biggest win. The manual research line -- reviewing companies, checking websites, writing personalization notes -- is now handled by Claude Code using Playwright for web research and our enrichment scripts for data retrieval. Output quality is higher than what junior researchers were producing because the process is consistent. No variance in how thoroughly each lead gets researched. Report generation: Our weekly client reports used to require someone pulling data from multiple sources, pasting into a template, writing a summary paragraph, and sending. Now Claude Code pulls from Supabase, formats into a Google Sheet, writes the summary, and drops a link in Discord. Nobody touches it.

What We Kept

Instantly ($300/month, unchanged): We still use Instantly for cold email sending. Claude Code is not an email sending platform -- it does not manage inbox rotation, domain health, sending schedules, or deliverability monitoring. Instantly does these things well and the cost is reasonable for what it provides. Replacing dedicated email infrastructure with a general-purpose AI would be the wrong trade. Supabase ($50/month): We still need a database. Supabase is cheap, reliable, and the Supabase MCP server makes it natively accessible from Claude Code sessions. This is not a tool we are trying to replace -- it is infrastructure. Domain infrastructure (unchanged): DNS, warmup services, inbox rotation -- these are not things Claude Code touches. Email deliverability is its own discipline. Apollo ($400/month, partial): We reduced our Apollo usage significantly but kept a lower-tier plan for prospecting and company search. The contact database is genuinely hard to replicate from scratch. We use Apollo more surgically now -- for the initial company and contact pull -- rather than as the center of the enrichment workflow.

---

The New Cost Structure

Down from $3,500/month (tools + labor) to $775/month. A 78% reduction.

The comparison that matters most: we eliminated $1,000/month in manual research labor and $800/month in orchestration tooling (Clay, Zapier, Make) and replaced both with approximately $175/month in Claude API tokens and about 40 hours of upfront development to build the execution scripts.

That development investment paid back in month two.

---

The Honest Tradeoffs

This is the part most "we replaced X with AI" posts skip. Let me be direct about what we gave up.

Pro: 10x cheaper with better customization

The cost reduction speaks for itself. But the customization improvement is arguably more valuable. We can now build enrichment logic, scoring models, and reporting workflows that are specific to our operation -- not constrained by what a SaaS tool's feature set supports. We added a cross-client deduplication check last month that would have been impossible in Clay. It took about three hours to build in Python.

Con: Technical maintenance is now our responsibility

Every SaaS tool we replaced was handling its own reliability. When Zapier's API changed, Zapier fixed it. When Clay added a new enrichment source, Clay shipped it. Now when an enrichment API changes its response format, we update the script. When a library breaks, we fix it. The maintenance burden is real and ongoing.

We estimate 4-6 hours per month of maintenance and improvement work on the execution scripts. At our cost basis, this is still a favorable trade. For an agency without technical capacity, it might not be.

Pro: We own the code

When you build your enrichment and automation logic in SaaS tools, you are writing your operational playbook in someone else's format. If they change pricing, shut down, or remove a feature, you lose your workflow. Our Python scripts are in a version-controlled repository. We can run them anywhere, hand them off to any developer, and extend them without asking permission.

Con: Not "set and forget"

SaaS tools are designed to be configured once and monitored occasionally. Claude Code requires active engagement. Someone on the team needs to understand how the system works, catch when outputs drift from expectations, and maintain the directive library that governs Claude's behavior. The system is more powerful than what we replaced, but it requires more intellectual involvement.

Pro: Can do things SaaS tools literally cannot

The biggest unreplaceable advantage: cross-tool orchestration with custom logic. Our enrichment workflow now pulls from three sources, cross-references against our historical campaign data to avoid re-enriching contacts we already have, applies a custom ICP scoring model, and updates four different tables in a single coherent workflow. No combination of SaaS tools could do this cleanly. Everything would require manual handoffs between systems.

Con: Learning curve is steep

Getting Claude Code to behave reliably requires building out a proper CLAUDE.md, a directive library, and a tested execution script library. This is weeks of work, not days. Teams that approach Claude Code as a drop-in replacement for their SaaS tools will be disappointed. It requires building a system, not just installing a tool.

---

A Real Example: Enrichment Before and After

To make this concrete, here is how a lead enrichment run worked before and after the migration.

Before (Clay-based):

1. Export new leads from Apollo as CSV
2. Import CSV into Clay table
3. Clay runs waterfall enrichment (Prospeo → Hunter → Clearbit)
4. Manually review Clay output for errors
5. Export enriched leads as CSV
6. Import CSV into our CRM/database
7. Manually flag low-confidence rows for follow-up

Total time: 45-60 minutes for a batch of 200 leads. Mostly manual handoffs and exports. After (Claude Code + Python scripts):

1. New leads imported directly to Supabase via execution/lead_sourcing/import_and_prepare_leads.py
2. Claude reads the enrichment directive, runs execution/enrichment/enrich.py against the Supabase records
3. Script calls Prospeo, Hunter, and Clearbit APIs in waterfall sequence
4. Results written back to Supabase automatically
5. Low-confidence rows flagged in a needs_review field
6. Claude posts summary to Discord: "Enrichment complete -- 200 leads, 163 verified, 18 flagged, 19 skipped"

Total time: 15-20 minutes for the same batch, with one person monitoring rather than executing. No CSV exports. No imports. No manual review of every row -- only the flagged ones.

---

Who Should Consider This Switch

Good candidates:

• Agencies doing $1M+ ARR with high-volume outbound
• Teams with at least one person who can write and maintain Python
• Operations where the SaaS tool ceiling is already being hit -- you need custom logic the tools do not support
• Companies paying per-row or per-seat fees that scale painfully with volume

Not a good candidate:

• Agencies under $500k ARR (the setup cost does not justify the savings)
• Teams with no technical capacity (no developer, no one who can debug scripts)
• Operations running standard workflows that existing SaaS tools handle well
• Anyone looking for a "plug it in and walk away" solution

The honest summary: this migration was the right call for us at our current scale and with our technical capacity. It would have been the wrong call eighteen months ago when we were smaller and the SaaS tools were sufficient.

---

What to Do If You Want to Explore This

If you want to see whether this approach makes sense for your operation, I would suggest:

1. Audit your actual SaaS bill -- not just the tool costs but the manual labor hours spent managing and working around tool limitations. That number is usually higher than people expect.

1. Identify the highest-friction workflows -- where are you spending the most time on manual handoffs, export-import cycles, or fighting tool limitations? Those are the best candidates for replacement.

1. Build one script and test it -- before committing to a full migration, pick one workflow and build the replacement. Run it in parallel with your existing tool for 2-3 weeks and compare output quality, reliability, and time investment.

1. Model the real ROI -- include setup time, ongoing maintenance, and the value of customization, not just the monthly cost delta.

For a rough calculation of what this might look like for your operation, see our ROI calculator. If you want to understand what a managed version of this infrastructure looks like as a service, the B2B lead generation page has more detail.

The switch was worth it for us. The $4,400/month in savings is real. So is the four weeks it took to build the system properly. Both things can be true.