What You’ll Learn

If you run a lot of AI work, you know the feeling: you’re plugging along and Claude tells you you’ve hit your usage limit, or you’re at 68% before the week’s even over. James pays for a Max plan and got tired of it — so he cut his costs instead of just paying more. In this session he shows how he did it with two moving parts: an AI HR Department that inventories every AI worker he owns, and a router that sends the cheap, no-thinking tasks to a free model running on an old Mac Mini in his closet.

By the end you’ll understand why the model you pick drives your bill, how to see every skill you’re actually running, and how to offload routine work to a free local model — even if, like James, you have no idea how to install anything technical.

The Problem: You’re Paying Premium Prices for Cheap Work

Every time you start a session in Claude you pick a model — Opus, Sonnet, Haiku, and the pricier tiers. The stronger the model, the more it costs. That’s fine for hard, complex work. The waste happens underneath: skills fire off inside your sessions on an expensive model even when the task needs almost no thinking at all.

“Why am I using 4.8? Man, that doesn’t need any thinking.”

The trouble is you can’t fix what you can’t see. Most people have no idea how many skills, agents, or scheduled tasks they’re running — or which ones are quietly burning premium tokens.

Step 1: Take Inventory With an AI HR Department

James built a plugin he calls the AI HR Department — a system of record for your AI workforce. It gives you one honest roster of every AI employee, team, playbook, and scheduled task you own, runs a weekly review that flags anything stale, duplicated, or off-spec, and includes a “light mode” for running routine steps on a cheaper model.

He asks his orchestrator (Dean) for the HR registry and gets a real headcount: 164 employees and skills across 19 agent teams, two playbooks, and 28 scheduled tasks — two of them stale. It builds an employee directory, just like an office would have.

💡 In Plain English: it’s a staff list for the robots in your business. Until you have one, you’re paying salaries to workers you can’t even name.

Step 2: Add a Router to Send Cheap Work to a Cheaper Model

Once James could see his workforce, he built a router — a routing policy that sends work that doesn’t need heavy thinking to a lighter model. Dean, the chief orchestrator, applies the rule automatically: important, judgment-heavy work goes to the strong model; routine steps go to something lighter.

What gets routed down: summarizing, scanning for news, searches, researching companies or content opportunities, crunching data, quoting, classifying, sorting, tagging, and rough drafts. What stays on the premium cloud model: the Content Flywheel brief (turning one idea into 15–20 pieces of content), final copy, voice/brand/consistency checks, and anything that publishes to the community. It even routes to the right Claude tier — if a job only needs Sonnet, it goes to Sonnet; if it only needs Haiku, it goes to Haiku.

Step 3: Stand Up a Free Local Model on Old Hardware

Here’s the zero-dollar employee. James had a six-year-old M1 Mac Mini (2020, 8GB RAM, 256GB drive) sitting in a closet doing nothing. He asked Dean what he could do with it — could it run an AI model? — and set up a free, open-source LLM called Qwen 2.5 on it. Now it sits on his home network and handles the grunt work for free.

The setup was simpler than it sounds, and James is upfront that he’s not technical:

“I asked Claude, ‘I want Ollama running on an old machine, what do I need to do?’ It walked me through it and gave me prompts I copied and pasted into the terminal.”

The steps: install a free piece of software called Ollama (it hosts language models locally), then tell Ollama to download the Qwen model. That’s it. The terminal is just the programmer window every Windows or Mac already has — you paste in what Claude gives you and let it run.

The Payoff

Now James has a whole workforce of employees he doesn’t pay for. Anything that meets the “no heavy thinking” bar routes to the old machine and the free Qwen model; the real work still goes to the cloud where it belongs. As your AI-native business grows and agents take on more, this is how you keep the bill from growing with it.

If you want the whole system, the campus AI operating system (where James talks to Dean) is at trainingsites.io/os. If you just want the HR department — the registry, the weekly review, and the model routing — it’s at trainingsites.io/hr.

Teach more, and let the agents do the rest — for less.