How We Run a GHL Agency on 2-4 Hours Per Week With Claude AI

Josh runs Redeemed Analytics. He gets on 2-3 sales calls per week, reviews Claude's work for 15-20 minutes per day, and sends agreements when prospects say yes.

Claude does everything else.

Prospect research. Outreach drafts. Client onboarding. Workflow building. Monthly reports. SEO audits. GHL sub-account setup. Retention monitoring. Content writing.

That's not a pitch. That's our actual operating model. Here's how it works, including the parts that are ugly.

The Hybrid Automation Model

GoHighLevel has an API. It covers contacts, conversations, opportunities, payments, calendars, and reporting. Claude uses it directly for those operations.

But GHL also has builders. The workflow builder, funnel editor, website builder, and snippet manager have no write API. The Workflows API is read-only. You can pull workflow data out, but you can't create or edit workflows through the API.

So we built a hybrid model. API calls for data operations. Browser automation for builder operations.

Claude opens GHL in a browser, navigates the UI, clicks buttons, drags actions into workflows, and configures triggers. It's slower than API calls. It breaks when GHL updates their UI. But it works, and it's the only option for builder tasks.

This hybrid approach handles about 90% of agency operations without Josh touching the keyboard.

What Claude Actually Does Each Week

Here's a real week, broken down by task.

Monday: Review Brief (15 minutes for Josh)

Claude generates a Monday morning brief. It includes: new leads from the past week, pipeline status across all client sub-accounts, any client websites with uptime issues, SEO ranking changes, and a prioritized action list.

Josh reads it over coffee. He responds with approvals, changes, or "skip this one." That takes 15 minutes.

Tuesday through Thursday: Autonomous Work

Claude runs through the approved action list. Real tasks from recent weeks:

• Researched 40 prospects in the OKC home services market. Scored each one based on their online presence, review count, and estimated revenue. Drafted personalized outreach emails for the top 12.
• Built a 7-step workflow in GHL for a new HVAC client. Trigger on form submission, send confirmation email, wait 2 hours, send SMS follow-up, create opportunity in pipeline, notify the client, wait 3 days and send review request.
• Generated monthly performance reports for 4 clients. Pulled data from GHL reporting, Google Search Console, and Google Business Profile. Formatted into branded PDF reports. Drafted the delivery email for each one.
• Ran an SEO audit on a client's website. Found 3 pages with missing meta descriptions, 2 broken internal links, and a sitemap that wasn't including new blog posts.

Josh reviews the outputs. He edits outreach emails if the tone is off. He checks workflows before publishing. He reads reports before they go to clients. Total review time: 15-20 minutes per day.

Friday: Sales Calls

When prospects respond to outreach, Josh gets on the call. Claude preps a one-page brief on each prospect before the call: their current tech stack, online reviews, competitors in their market, and specific pain points Claude identified during research.

After the call, Josh tells Claude the outcome. If it's a yes, Claude starts the onboarding sequence: creates the GHL sub-account, loads the appropriate snapshot (carefully, on a blank account), sets up the domain, and drafts the welcome email.

The Error-Learning System

Here's the part most AI automation articles skip: mistakes.

Claude makes mistakes. It loaded a snapshot onto a live client account and took down their website. It sent an email from the wrong GHL sub-account. It built a workflow that triggered on every page visit instead of form submissions, which burned through the client's SMS credits in 4 hours.

So we built an error-learning system. Every mistake gets logged with a severity rating, root cause, and fix. Every success gets logged too. Before Claude starts any task, it searches this database for relevant lessons.

The database has over 2,300 entries now. When Claude goes to load a snapshot, it finds the lesson about the Sooner State Haulers incident and checks for existing websites first. When it sends an email from GHL, it finds the lesson about verifying the FROM address and sub-account before hitting send.

The mistake rate dropped by about 70% in the first 3 months of using this system. It still makes errors. But it rarely makes the same error twice.

The Multi-Model Stack

Claude is the orchestrator, but it's not the only AI in the system.

Codex CLI handles filesystem audits and code reviews. When we deploy a client website, Codex scans the build output for missing meta tags, broken image references, and accessibility issues. It catches things Claude misses because it processes the actual file tree differently.

Gemini handles image generation and large-file analysis. OG images for blog posts, social media cards with client branding, and processing large CSV exports from GHL reporting.

ChatGPT Deep Research handles research that needs more depth than a web search. Competitive analysis reports, industry trend summaries, and technical documentation deep-dives. It runs in a dedicated subagent, downloads the research as markdown, and saves it to our knowledge base.

Claude orchestrates all of them. It decides which model to use for each subtask, dispatches the work, collects the outputs, and assembles the final deliverable.

What This Actually Costs

Time: Josh spends 2-4 hours per week on the business. That includes Monday review, daily check-ins, and sales calls.

Money: Claude API costs run about $200-400 per month depending on volume. Codex and Gemini add another $50-100. ChatGPT Plus is $20. Total AI cost: roughly $300-500 per month.

For context, a single full-time virtual assistant doing the same work would cost $2,000-4,000 per month and wouldn't have the error-learning system, the multi-model capabilities, or the ability to work 24 hours a day.

What Claude Cannot Do

Sales calls. Claude can research the prospect, prep the brief, and draft the follow-up email. But the actual conversation where a business owner decides to trust you with their marketing? That's human work.

Strategic decisions. Claude can present options with data behind each one. But deciding which clients to fire, which market to expand into, or when to raise prices? That requires judgment that AI doesn't have.

Relationship management. When a client is frustrated, they need to talk to Josh. Claude can flag the risk signals (declining engagement, missed meetings, negative feedback), but it can't rebuild trust.

Client approvals on deliverables. Claude generates the work. Josh reviews it. The client approves it. That chain of human review is non-negotiable.

Start Tomorrow

You don't need our full system to start. Here's the minimum viable version:

1. Get Claude (Anthropic API or Claude Pro). Connect it to your GHL account via API key.
2. Start with one task: prospect research. Give Claude a market and criteria. Let it find and score prospects.
3. Review the output. Edit it. Send it. Track what happens.
4. Add a second task: monthly client reports. Give Claude access to your GHL reporting data.
5. Log every mistake. Search the log before every new task.

The system grows from there. Workflow automation, content writing, SEO audits, onboarding sequences. Each one gets added after the previous one is stable.

We didn't build this in a week. It took 8 months of daily iteration. But the result is an agency that runs on 2-4 hours of human time per week.

If you want to see how we'd set this up for your agency, book a call. If you want to understand our AI automation approach, read what we mean by real AI vs. buzzword AI. And if you want the tools we built for this, check out the Cowork Brain Kit.

We also offer AI automation as a service and full GHL agency management if you'd rather skip the build phase.

FAQ

Do I need coding skills to set up Claude with GoHighLevel?

Basic API knowledge helps. You need to generate a GHL API key, configure authentication, and understand JSON responses. But you don't need to be a developer. Claude can write the integration code. You just need to know enough to verify it's connecting to the right accounts.

How reliable is browser automation for GHL workflow building?

About 85-90% reliable. GHL updates their UI periodically, which breaks automation scripts. When that happens, Claude detects the failure, logs it, and adjusts. Most breaks get fixed within one retry. The error-learning system means each UI change only causes problems once.

What happens when Claude makes a mistake with a client's account?

The error gets logged immediately with a severity rating. Critical errors (like the snapshot incident) trigger an alert to Josh. Claude writes the root cause and fix before retrying. The lesson goes into the database so it doesn't happen again. In 8 months, we've had 3 severity-red incidents. All were fixed within an hour.

Can this work for agencies with 20+ clients?

Yes, but you need more review time. Our 15-20 minutes per day covers 5-8 active clients. At 20+ clients, plan for 45-60 minutes of daily review. The AI work scales linearly. The human review time scales too, just more slowly.