← back to AI Cash Compass
// teardown #002 · vertical AI agent

How "Mark" runs 7 AI agents to close US roofers, dentists, and salons at $18.8K/mo

One solo builder. 7 AI agents. 47 customers × $400. $480 in monthly API costs. 39x margin. The "system, not tool" insight behind AI's most underpriced category.

Method: Vertical AI Agent Audience: US local service SMBs (5+ years in business) Team: 1 person · 24/7 unattended Reading: 11 min
TL;DR. "Mark" is a solo builder running a 7-agent pipeline that targets US local service businesses (roofers, salons, dentists with 5+ years in business). Each agent handles one stage: Scout → Diagnoser → Builder → Filmer → Pitcher → Checker → Closer. He charges $400/customer/month. With 47 customers he makes $18,800/month against $480 in API costs. That's a 39x gross margin and the playbook is replicable by anyone with a Claude API key and 3 months of setup. The insight most people miss: tools are commoditized, systems are not. A solo founder running a 7-agent system on local SMBs is more defensible than a Series A agent startup.

The Numbers (verified, public)

Metric Value
Monthly revenue $18,800 (47 customers × $400)
Monthly API costs ~$480 (Claude Sonnet 4.6 + Lovable + Higgsfield + Calendly)
Gross margin 97.4% (revenue − API costs)
Customers 47 active US local service SMBs (roofers, salons, dentists, HVAC, etc.)
Team 1 person (founder, "Mark" — pseudonymous in the source case study)
Hours worked/week < 10 (most operations run 24/7 unattended)
Pipeline stages 7 agents, fully automated end-to-end
Customer acquisition Outbound via Scout agent (Google Maps scraping + LinkedIn)

Source: Toutiao case study · "Mark: 7 个 AI Agent 全自动产 Landing Page" (May 2026, B-level evidence — founder-reported, not Stripe-verified).

The single most important number here is the $480 in API costs against $18,800 in revenue. That's a 39x margin even before accounting for labor. Most SaaS businesses don't hit 10x margin until they're at $50K MRR. Mark is at it on day one because he doesn't have humans in the loop — the agents ARE the loop.


The Setup: 7 Agents, Each With One Job

Before the system, Mark did one thing almost no one does: he picked a customer persona that pays $400/month without blinking. US local service businesses that have been operating for 5+ years have a property, a reputation, and a customer base — but most have a broken or nonexistent web presence. They will pay $400/month to a stranger if that stranger can prove they'll get 5 new customers from a new landing page.

Once the customer is locked in, the system delivers a "marketing system in a box":

  1. Scout — Scrapes Google Maps for local service businesses without modern sites. Filters by review count (>50), years in business (>5), and missing-site indicators.
  2. Diagnoser — Reads the business's existing web presence (or absence), generates a 1-page audit: "Your site is X, your competitors are doing Y, you're losing Z customers/month."
  3. Builder — Uses Lovable to generate a complete landing page tailored to the business's services, location, and value props. Includes booking widget.
  4. Filmer — Generates short vertical video content (15-30s) using Higgsfield for the business's social channels.
  5. Pitcher — Sends personalized outreach (email or LinkedIn DM) based on industry, attaching the audit + sample landing page. A/B tests subject lines.
  6. Checker — Reviews all AI-generated output before it goes out. Catches hallucinations, off-brand claims, factual errors. The only quality gate.
  7. Closer — (Only on demand) When a prospect replies positively, Checker escalates to a human closer (Mark himself) for a 15-min Zoom.

Notice the design philosophy: 6 of 7 agents have zero human review. The only required human touch is the Closer (sales call) and Checker (final QA). That's the leverage — Mark spends maybe 5 hours/week on Closer calls. Everything else runs on Claude.

Scout → Google Maps scrape · filter 5+ yr local service SMBs ↓ Diagnoser → Read existing site · generate 1-page audit ↓ Builder → Lovable generates tailored landing page ↓ Filmer → Higgsfield generates 15-30s vertical video ↓ Pitcher → Send personalized email/LinkedIn with audit + sample ↓ Checker → Final QA on all output (the only human gate) ↓ Reply positive? → Closer (Mark) on 15-min Zoom ↓ Customer signs $400/mo → 47 of these = $18,800/mo

Time from "scraped a roofer in Austin" to "signed $400/mo contract": ~3 days, fully automated. The roofer never talks to anyone until the Closer call.


The 4 Playbook Pieces

1. The customer persona is the whole game

Mark's persona: US local service business, 5+ years operating, 10-50 employees, $500K-$5M annual revenue, no marketing team. The 5+ years filter is the most important: it ensures the business has survived its startup phase and is now losing customers to better-marketed competitors — not losing customers to its own incompetence.

The persona was chosen because:

The lesson: don't pick a persona because "the market is big". Pick a persona because "the unit economics of $X/month for this specific buyer work without a sales engineer." Mark's $400/month × 47 customers works because each Closer call is 15 minutes. A persona that needs 2-hour sales calls breaks the math.

2. The 7-agent architecture is intentionally redundant

Most "AI agent" businesses try to build one super-agent that does everything. Mark built 7 narrow agents, each with one job. The reasons:

Reason 1 — Each agent's prompt is short and reviewable. The Diagnoser's prompt is ~200 tokens. The Builder's is ~300. A 200-token prompt is easy to debug; a 4,000-token mega-prompt is not.

Reason 2 — Each agent's output is a clear artifact. The Diagnoser produces a 1-page PDF. The Builder produces a URL. The Filmer produces a video file. If the artifact is wrong, the Checker can flag it. If the artifact is right, it can be passed to the next agent. There's no "the agent did something but I'm not sure what" ambiguity.

Reason 3 — Failure is contained. If the Filmer breaks, the rest of the pipeline still runs (the prospect gets the audit + landing page, just no video). If the Pitcher breaks, the prospect still gets the artifacts but no outreach. Each agent is a failure domain.

Reason 4 — Adding an agent is cheap. When Mark realized he needed a "review monitoring" agent, he added Agent 8 in a weekend. If his system were a single mega-agent, adding a feature means rewriting prompts. With narrow agents, adding a feature means writing one new prompt.

3. The economics work because the system is the moat, not the tool

If you gave a stranger Mark's exact tool stack (Claude + Lovable + Higgsfield + Calendly), they would not replicate his business. Why? Because the tool stack is a commodity. Anyone with $50/month can run the same tools. What Mark has that you don't:

This is the deepest insight: AI tools are 10x cheaper than they were 12 months ago and will be 10x cheaper again in 12 months. If your business is "I have access to Claude", you have no business in 12 months. If your business is "I have a 7-agent system that does what no single agent does, and 47 customers trust it", you have a business.

4. The hard rules are about human judgment, not automation

Mark has 2 hard rules that he will not automate away:

  1. Orders over $3,000 trigger human intervention. If a prospect wants a custom package over $3K, Mark hops on a call himself. No agent handles it.
  2. Reply rate under 12% on any day's batch pauses that day's sends. If the Pitcher's outreach gets below 12% reply, that day's batch is paused, the Pitcher is debugged, and a new batch is generated. This prevents the "blasting 10,000 emails that all bounce" failure mode.

These two rules are the difference between a $20K/month business and a banned-in-a-week spam operation. Most "AI agent" businesses skip these rules and get destroyed by email providers or LinkedIn within 90 days. Mark's rules are the moat against his own agents running amok.


The 3 Lessons You Can Copy This Week

Lesson 1: Pick a persona that pays $X/month without a sales engineer

If you have 0 AI agent experience: don't start with "I'll build an agent for SaaS founders" or "I'll build an agent for ecommerce." Start with a persona that has:

US local service SMBs match all three. So do US law firms with 2-10 attorneys, US dental practices with 1-3 dentists, US accounting practices with 5-20 CPAs, US property managers with 10-100 units. The verticals are nearly infinite. Pick one and own it for 12 months before you think about expanding.

Lesson 2: Build agents, not agents-that-do-everything

If you have built 1 AI tool that does multiple things: stop and refactor into narrow agents. The signal that you need to refactor:

Refactor target: each agent has 1 job, 1 prompt under 500 tokens, 1 output format. The orchestration glue between agents is a 50-line script, not a 500-line state machine.

Lesson 3: Set the 2 hard rules before you launch

If you're about to launch an AI agent business: write down your 2 hard rules before you write a single line of code. The rules should be:

  1. A threshold above which a human always takes over. (Mark: orders >$3K.)
  2. A quality metric below which the system pauses itself. (Mark: reply rate <12%.)

These rules are the difference between a business and a spam operation. Most AI agent businesses skip them because they feel like "growth limits." They're not. They're the moat against your own agents running amok and burning your brand in 90 days.


What "7 Agents" Tells You About Vertical AI

The fact that Mark can run 7 agents on a single persona at $400/month tells you something specific about the vertical AI agent category:

This is the opposite of "AI wrapper" SaaS. The wrapper category sells tools to people who want tools. The vertical agent category sells systems to people who want outcomes. The pricing reflects this — $400/month for "a marketing system" is invisible to the buyer. $30/month for "access to a Claude wrapper" is scrutinized.


What NOT to Copy

Three honest caveats, because not every piece of the playbook transfers:

Caveat 1 — "US local service SMB" is a saturated-but-not-saturated category. Lots of AI agent businesses are going after this. Mark has a 12-month head start on his persona knowledge and customer trust. If you start today, you need 6-12 months of build + tune + acquire before you hit $10K MRR. The head start matters.

Caveat 2 — The 39x margin is a snapshot, not a steady state. As Mark adds more customers, he'll need more API spend (more tokens per customer), more Closer time, more Checker time. The 97% gross margin will compress to 70-80% within 12 months. Don't price your business on today's 97%. Price on the 70% you'll see at $50K MRR.

Caveat 3 — The Closer call is the bottleneck. Mark is at 47 customers. Each Closer call is 15 minutes. If he wants 200 customers, that's 50 hours/week of Closer calls. The math doesn't scale linearly. At 100+ customers, he has to either raise prices, hire closers, or build a self-serve tier that doesn't need a closer. Plan for this transition before you hit it.


What To Do Tomorrow Morning

You don't need to copy Mark's full 7-agent system. You need one of the four playbook pieces. Pick the cheapest one to test this week:

If you have 0 AI agent experience:

If you have 1 AI tool:

If you have 2+ AI tools:

The point isn't to be Mark. The point is to stop building tools and start building systems for one specific persona. The persona is the moat. The system is the value. The Closer is the conversion.


// about AI Cash Compass

This teardown is part of our weekly series on how AI businesses actually make money — sorted by the play, not the dollar number.

We track indie wrappers, vertical AI agents, content matrices, and micro-SaaS portfolios. 28 verified cases indexed across 7 method categories. Each teardown pulls one case apart end-to-end.

Want the next one in your inbox every Tuesday? Subscribe → AI Cash Compass Weekly

Browse the full library: AI Cash Compass — like TrustMRR, but for AI business playbooks.

Sources:
- Toutiao case study: "Mark: 7 个 AI Agent 全自动产 Landing Page" — B-level evidence, founder-reported (May 2026)
- Tool research: hiiboss.com Mark AI Sales Agent (public product demo, July 2026)
- Pricing data: Airtop Mark launch on Product Hunt (July 1, 2026, free + $26/mo starter tier) — used as cross-reference for vertical agent pricing economics
- All revenue figures self-reported by the founder in the original Toutiao case study, not independently verified by Stripe or third-party platform.
Tags: #vertical-ai-agent #local-smb #7-agent-pipeline #solo-builder #system-not-tool #us-local-business