The Real AI Cost of Automating 15 Videos a Week

Onmee currently produces 15 videos a week across 3 YouTube channels — DailyWonders, LostEras, and VaultedMyths. The monthly AI bill for the entire system lands around $4-7. That’s not because we use cheap models. It’s because most of the pipeline never calls AI at all.

The most common mistake businesses make when “adopting AI”

The typical approach is to bolt AI onto every step of a process — data intake, processing, publishing, reporting. The API bill balloons within a few months, and the conclusion becomes “AI is too expensive to be worth it.”

The problem isn’t the price of AI. It’s the architecture. Most steps in a repeatable process — fetching data, scheduling, reformatting files, pushing data between systems — don’t need creative judgment or reasoning at all. They just need plain automation, running on a script or an automation tool, at close to zero marginal cost.

The real architecture behind 15 videos a week

A finished video passes through 5 steps: topic selection, scriptwriting, voiceover generation, video assembly, and scheduled publishing. Only 2 of those 5 steps actually need a generative AI model — scriptwriting and voiceover, because those are the two steps that require human-like expression. The other three run entirely on AI-free automation.

Breaking the cost down step by step

Topic selection runs on a script that reads channel performance data against a prepared topic list — no AI call involved. Video assembly stitches footage, captions, and background audio using a fixed template through an automated rendering tool — also no AI. Publishing runs on a schedule through the platform’s API — still no AI.

The only two steps that cost anything are scriptwriting and voiceover, because that’s where a generative model actually needs to sound human. Even there, defaulting to a speed- and cost-optimized model instead of the flagship one keeps the per-video cost at $0.06-0.12.

Why the number stays this low even after counting everything

Multiply 3 channels by 5 videos a week by 4 weeks, and the system produces roughly 60 videos a month. At $0.06-0.12 in AI cost per video, the entire system’s monthly AI bill lands around $4-7. That number matters more than the hours saved — it proves that whether AI is expensive or cheap depends entirely on how you design the process, not on the technology itself.

How to apply this to your own process

1. List out the full process as discrete steps, in the order they actually run.
2. Mark which steps genuinely require creative judgment or contextual reasoning — only those need AI.
3. For every other step (fetching data, reformatting, pushing data into another system), replace it with plain automation using a script or a tool like n8n, Make, or Zapier.
4. On the steps that do require AI, default to the cost- and speed-optimized model first, and only upgrade if you have evidence quality isn’t good enough.
5. Measure cost per unit of output (per video, per email, per invoice processed) to know whether the architecture is actually optimized.

This principle applies to any repeatable process, not just video production — the same logic runs through a 3-step process for invoices and contracts and automating customer service response: AI should only show up at the step that needs judgment, and everything else is plain automation.

Frequently Asked Questions

Why not just use AI for the entire process for convenience?

Because every AI call costs money and takes longer to process than a plain script. Using AI for a step that requires no creativity, like pushing data into a Google Sheet, only slows the system down and inflates the bill without improving output quality.

Does this cost include the automation platforms (n8n, Zapier, Make…)?

No. The $0.06-0.12 per video figure covers only the AI model calls. Automation platform costs are separate, typically a flat monthly fee that doesn’t change whether you run 15 or 30 videos a week.

Is 15 videos a week a sustainable long-term pace?

That’s the current publishing cadence across 3 channels, 5 videos each. Because most of the process is plain automation, adding more channels or videos doesn’t scale cost linearly with headcount hours — it only scales with the number of AI calls at the two steps that need judgment.

Conclusion

$4-7 a month sounds almost too small to be real, but that’s exactly what the right architecture produces. Most businesses don’t fail because AI is expensive — they fail because they let AI do work a plain script already did better, and cheaper. The first question before picking a model should always be whether this step needs AI at all.