How I'm building an AI agent system for my agency

Most of my clients come to me paying five different vendors to do what should be one job. An SEO agency. A paid ads person. A content writer. Someone who runs reports. Someone who manages the website. Each vendor has their own dashboard, their own invoice, their own communication style, and their own definition of what success looks like.

The client is the one holding it all together. Which means the client is the bottleneck.

That’s the problem I’m building against.

What the AI agent system does

I’m building a system using Claude that consolidates four things into one managed workflow:

SEO monitoring

The agent pulls ranking data, identifies pages that are slipping, and surfaces what needs attention. Not a weekly report. Ongoing awareness.

Ad performance tracking

Instead of logging into three platforms and manually building a comparison view, the agent ingests data from Google and Meta and flags anomalies. CTR drops. Budget pacing issues. Campaigns that stopped spending.

Content generation

First drafts for blog posts, ad copy variants, meta descriptions. The agent works from a brief. I review and approve. The client never sees the draft phase.

Lead reporting

A weekly summary that actually says something. Not raw numbers. Context. “Your top converting page this week was X. The drop in form submissions on Y page is likely tied to Z.”

The architecture

Right now the system runs in three layers.

Inputs

Data comes from Google Search Console, Google Analytics, Meta Ads Manager, and the client’s Webflow CMS. Most of this is pulled via API. Some of it is still manual because I haven’t built every connector yet.

Processing

Claude handles the analysis and generation layer. I’m using the Claude API with structured prompts and tool use. The agent doesn’t decide strategy. I decide strategy. The agent executes tasks inside defined parameters.

Outputs

Slack messages, Google Doc reports, and eventually a simple dashboard. The client gets one update per week. I get flags in real time when something needs my attention.

This is not a no-code automation. It’s a proper system. Prompts are versioned. Outputs are validated. The agent is one component, not the whole thing.

What Claude handles vs what I decide

This distinction matters and I want to be clear about it.

Claude handles: data processing, pattern recognition, first-draft generation, formatting, and summarization.

I decide: strategy direction, what to prioritize, whether an output is ready to send, and how to interpret unusual data.

The agent is fast and consistent. I’m the one who knows the client’s context, the history, and what actually matters to their business. Those two things together are more valuable than either one alone.

Why I’m testing it on Carril first

I’m not selling something I haven’t used. That’s a line I won’t cross.

So before this becomes a client offer, I’m running it on Carril’s own setup. Our SEO. Our content. Our performance data. I need to know where it breaks, where it surprises me, and what still requires more human input than I expected.

So far: the reporting layer is working well. The SEO monitoring needs better signal filtering. The content generation is useful for structure but always needs editing. That’s expected.

I’ll keep documenting this as it progresses.

What this offer looks like when it’s ready

One system. One point of contact. Real-time flags when something needs attention. A weekly summary that reads like it was written by someone who actually understands the business.

Priced as setup plus monthly management. No long contracts. Clients can see exactly what the system is doing.

That’s the goal.