// AI Lessons
How to Cut Your AI Bill by Moving Mundane Work to Free Open Source Tools
Keep one paid seat for the hard 20%. Run the boring rest on free open-source tools — for nothing.

My monthly AI bills are, frankly, atrocious. I added them up recently and winced — a seat here, a seat there, an image tool I'd half-forgotten, a couple of seats for the team. The standard paid seat for ChatGPT or Claude runs $20 a month, and once you stack a few — chat, writing, images — it's easy to land quietly at $50 to over $100 a month. And that's before the autonomous agents: when I turn something like OpenClaw or Hermes loose on a task, the meter spins in a way that's downright sickening — they'll happily chew through tokens for hours while I'm off getting coffee.
Then I look at what all that frontier horsepower is actually used for: summarizing a PDF, cleaning up a draft, upscaling a logo, and a quick chat to name a file—mundane stuff. I don't need a frontier model — or a subscription — to resize an image.
The move is to shift the boring 80% to free, open-source tools that run on your own machine, and keep the paid seat for the hard 20%. Here's how to do it, and how to make sure what you install is actually safe.
Sort your AI work by what truly needs a frontier model — then move the rest off your paid seats.
// The Takeaway: Most of what you pay an AI subscription for is routine: summarizing, drafting, rewriting, local chat, document Q&A, image cleanup. A short list of free open-source apps — Jan, Ollama, AnythingLLM, Upscayl, ComfyUI — handles that work on your own hardware for nothing, and your data never leaves the machine. Keep one paid seat for the genuinely hard 20%. Before you deploy any of these, run a 10-point vetting prompt to avoid mistaking "free but proprietary" for open source—net result: a smaller bill and more control.
Download Jan and run your first local chat → · Thirty minutes. Free. Runs on your laptop, online or off.
The trap isn't that paid AI is a rip-off. It's that you're paying premium rates for economy work — hiring a town car to get the mail.
// The real shift: Stop asking "which AI is best?" and start asking "which tasks actually need a frontier model?" Once you sort your work that way, most of it doesn't — and a big chunk of your AI bill turns out to be optional.
Step 1 — Audit where the money goes
Pull your last statement. The standard tier is $20/month per person; Microsoft Copilot for business runs about $42.50 per user per month once you add the Office license it requires. Now write down the five things you used AI for most this week. Be honest about each one: did it need deep reasoning, or was it a chore?
Step 2 — Map the chores to free open-source swaps
Routine chat, drafting, summarizing → Jan. A local chat app with a built-in model hub that runs models on your machine, 100% offline after download. Free.
The engine under everything → Ollama. The MIT-licensed model runner that downloads and serves open models with one command; most other tools plug into it.
Chat with your own documents → AnythingLLM. MIT-licensed, runs locally, your files never leave your machine.
Image upscaling → Upscayl. A free open source upscaler that runs locally — no cloud uploads.
Image and video generation → ComfyUI. Free and open source under GPL-3, with a one-click desktop app.
Coding help → OpenCode. A free, model-agnostic open source coding agent.
One to skip if you want actual open source: LM Studio. It's a polished local chat app and it's free, but the app itself is proprietary — only its command-line piece is open. Fine to use; don't file it under open source.
Step 3 — Stand one up in thirty minutes (Jan)
Download Jan from jan.ai — there's an installer for Windows, macOS, and Linux. No terminal required.
Open it and go to the model hub. Download one small model to start — a 7–8B model is a few gigabytes and runs on a normal laptop.
Start a chat. It runs on your hardware; turn off your wifi and it still answers.
Point it at chores: paste a long email and ask for a three-line summary, drop in a rough draft and ask for a cleanup, have it outline or rename a document.
When you hit something that genuinely needs frontier reasoning, that's what your one paid seat is for.
That's the skill. You now have a free, private assistant for the routine 80%.
Step 4 — Before you deploy, run the safety check
Free and open doesn't automatically mean safe. Before you let any of these touch real work, paste this into your AI and fill the four brackets:
You are an open source due-diligence analyst. I'm evaluating an AI tool for
enterprise deployment. Score it on the 10 checks below, flag any dealbreakers,
and give a one-line verdict: DEPLOY / DEPLOY WITH CONDITIONS / DO NOT DEPLOY.
Tool: [TOOL NAME]
Repo / license URL: [GITHUB OR LICENSE LINK]
Where it will run: [LOCAL / OUR CLOUD / VENDOR CLOUD]
What data it touches: [e.g., customer PII, source code, internal docs]
Check and report on:
1. License — OSI-approved, or "source-available"/proprietary-but-free? Name it.
2. Whether the whole app is open, or only its CLI/SDK (the common open-washing gap).
3. Maintainer health — sole maintainer, company-backed, or foundation-governed?
4. Release cadence and last commit date.
5. Data residency — does anything leave the machine by default? Telemetry?
6. Model provenance — what models ship, under what license, local or via API?
7. Known CVEs or advisories in the last 12 months.
8. Supply chain — dependency count, signed releases, SBOM available?
9. Commercial-use terms — does "free" change at work or at scale?
10. Exit cost — if it's abandoned, can we fork it and keep running?
It catches the "free but proprietary" mix-up, flags sole-maintainer projects, tells you whether anything phones home, and hands you a deploy / don't-deploy read in about ten minutes.
Bonus: the team math
This compounds. If five people each move routine work off one $20 seat, that's $1,200 a year back — and the work stays in-house instead of on someone else's server. Run the swap once, write it down, and it pays every month.
Why This Matters Now
Security teams have stopped treating "shadow AI" as someone else's problem. In late April, Cisco open-sourced its Model Provenance Kit and an AI-BOM scanner precisely so companies can inventory every model and tool running in their stack. Moving your mundane work to a short list of vetted open source tools does two things at once: it cuts the bill, and it gives you a real answer when someone asks what's actually running on your machines. Start with one tool this week. Run the prompt. Cancel one seat you don't need.

Your AI Sherpa,
Mark R. Hinkle
Founding Publisher, The AIE Network
Follow me on LinkedIn
If you want to get in contact or give me feedback, reply to this email. I read every single one of them.
