How I replaced a $10K/month agency with an AI stack

The agency sent beautiful reports every month. Charts that looked great and meant nothing in the bank account. The stack that replaced them costs a few hundred dollars a month and runs on my own laptop. Here's the whole teardown, including the part nobody tells you.

The before

A full-service content agency at the $5K to $15K-a-month tier delivers a recognizable package. Call it four to eight long-form articles, a dozen or two social posts, a few emails, a landing page, monthly SEO, and an analytics report at the end. The real thing you're paying for isn't the deliverables. It's the pod: a strategist, a writer or two, an SEO person, an editor, and an account manager who keeps the whole thing moving. Most contracts lock you into a three to twelve month minimum.

The numbers in this piece are representative of a swap like this, not a forensic audit. But the architecture is exactly the one I run.

The agency does competent work. That was never the problem. The problem is the gap between the report and the bank account. I'd get a polished monthly deck with charts trending up and to the right, and none of it moved revenue. Turnaround was slow, a week or two for a single article. Every change went through an approval chain. An operator I follow described a founder who "was paying an agency that sent him beautiful reports every month, charts that looked great and meant nothing in his bank account." That's the trap. You stop measuring outcomes and start admiring the deck.

At some point I stopped asking "is this good work?" and started asking "could I build this myself, faster, cheaper, and with more control?" The answer turned out to be yes, with one large catch I'll get to.

The architecture we shipped

The stack mirrors how a content agency actually works, except the pod is software and the account manager is me. Here's the pipeline, tool by tool.

Capture. Every meeting, call, and customer conversation runs through Granola, a meeting tool that records and transcribes in the background. No notetaker bot squatting on the screen. The transcripts sync to a folder on my laptop.

Mine the angles. Those notes pipe straight into Claude Code, no copy-paste. I point it at the folder and ask what's there: the customer truth that came up twice this week, the objection that keeps surfacing, the story I told on a call that would land as an article. This is the strategist's job, and it runs on raw material I generated anyway just by doing my work. I wrote up the exact version of this here: content from your meetings and, for the bigger swings, book ideas from meetings.

Draft. Claude writes the long-form drafts. It's the best tool I've used for long pieces and for holding a consistent voice across them, as long as you feed it the voice. I load a style guide so the output sounds like me and not like a press release. The drafting runs as a repeatable Claude Code workflow, not a fresh prompt every time.

Render and publish. Scripts handle the formatting, the cross-links, and the push to the site. The mechanical tail end that used to eat an afternoon now runs on its own.

The cleanest way to think about the split is the AI agent versus assistant lens, and the honest version is which one for which job. The agent does the volume and the mechanical work: transcribing, drafting at scale, rendering, publishing. The assistant, which is me sitting in the loop, does the judgment work: the angle, the voice, the final edit, the call on whether a piece ships at all.

This article is the pipeline in action. It came out of the same workflow I'm describing.

The economics are the loud part. The agency retainer sits around $10K a month. The stack costs low hundreds. A serious solo setup runs roughly $100 to $400 a month: Claude Max in the $100 to $200 range, ChatGPT Plus at about $20, a meeting-capture tool, and some odds and ends. Call it 2 to 4% of the retainer. A lighter version on Pro plans only runs $40 to $60 a month.

The dollar cost collapses. The honest catch is that a different cost moves onto you. Your time. The setup. The learning curve. The QA on every single output. The retainer bought you a team that absorbed all of that. Run the stack yourself and it doesn't vanish, it relocates to your desk.

What we measured

Start with the most credible number I can point to, because it isn't mine. A controlled study out of Harvard Business School (the "Navigating the Jagged Technological Frontier" working paper, with BCG) found that consultants using AI finished tasks about 25% faster and produced work rated roughly 40% higher in quality. That's a real study with a control group, not a vendor case study, and it matches what I feel running the stack: faster, and the floor on quality goes up, not down.

On the content economics specifically, vendor-reported figures put per-article production cost at an agency in the $500 to $800 range, dropping to somewhere around $50 to $270 on an AI workflow depending on how much human polish you put in. That's a large reduction. I won't dress it up as one clean percentage, because the real number depends entirely on how much editing each piece needs, and a piece that needs heavy editing isn't really cheaper. Turnaround is the change I feel most: a week or two for one article compressed to a day or two. Output volume rises too; one benchmark I've seen reported is roughly 40% more articles a month.

My own results I'll keep qualitative, because that's the honest register for them. It's faster. It's cheaper. And I control it. I can change the angle at 11pm without waiting on an approval thread, and the voice stays mine because the voice guide is mine.

What we'd do differently

Here's the credibility section, the one the agency-killing posts skip.

AI did not replace everything the pod did. It replaced the production line. The judgment work is still on me, and some of it AI can't touch at all:

• Brand strategy. The point of view, the positioning, the thing that makes the writing worth reading. AI executes a strategy; it doesn't have one.
• Taste and editorial judgment. Knowing which draft is actually good, which line to cut, when a piece is trying too hard. The model can't feel the room.
• Original reporting and relationships. The customer who tells you the real reason they churned. The off-the-record context from a partner. None of that is in the training data because it hasn't happened yet.
• The forcing function. This is the sneaky one. An external retainer is an accountability machine. You're paying, so the work happens. Remove the retainer and the work can quietly stop, because the only thing making you ship is now you.

There's a cautionary case worth sitting with. A founder on X fired his agency, took over SEO himself with AI, and three months later ranked worse than he had in a year, despite pouring hours into it. The tools didn't fail him. The point of view and the discipline did. It is not a clean one-to-one swap. The human doesn't disappear from the stack; the human moves up it, from doing the work to directing it. The wins go to whoever has the clearest point of view and the best judgment about where human creativity has to lead and automation follows.

If I were starting over, I'd build the voice guide and the QA loop first, before I touched the cost savings. The savings are real, but they're the reward for getting the judgment layer right, not a thing you get for free by canceling a contract.

Want to maximize your AI leverage? Upgrade to Pro.

What to do next

Don't fire your agency on Monday.

Pick the single most repetitive thing they do for you. The weekly blog post is the usual one. Rebuild just that one function as an AI workflow: capture in Granola if it draws on your meetings, mine and draft in Claude Code against your voice guide, and if you want to feel the build yourself, start with build your first software tool. Run your new workflow and the agency in parallel for a month. Compare the output, the turnaround, and the cost, honestly. Expand only what wins.

That's the whole move. One function, one month, side by side. The stack earns the next swap or it doesn't, and you'll know either way instead of guessing from a deck.

If you want the wider menu of where to start, here's 10 ways a CEO can put AI to work this week. Tell me in a month what you measured. I'd love to hear what actually moved.

Andrew

Related reading

• 10 ways a CEO can put AI to work this week · the broader menu of first moves
• AI agents vs AI assistants: which one for which job? · the lens behind the whole architecture
• What is Granola? · the capture layer the pipeline starts with
• Content ideas from meetings · the exact mining-to-draft workflow