Introducing JSC Labs

First things first: JSC Labs is LIVE. It's where our AI experiments go once they produce something you can actually use. Every issue, we take one idea, build it properly, break it, fix it, and hand you the artifact — a skill, a script, a template. The write-up stays here. The thing itself lives at JSC Labs.

Workflow to Watch is now that experiment.

Now, to the news...

In April, we predicted in these pages that AI's role going forward would be curation, not creation, as it is clear that different AI models have different strengths and weaknesses, and that pricing would inevitably change and would become impossible to keep up with.

Turns out, we were half right.

OpenRouter looked at over 100 trillion tokens of real routing traffic and found demand is basically price-inelastic — cut a model's price by 10% and usage goes up about half a percent (OpenRouter, State of AI). Nobody is switching models to save money. But that hasn't slowed their creation.

Ramp shipped a model router on August 20 — free through 2026, routing across OpenAI, Anthropic, DeepSeek and xAI (TechCrunch, August 2026). Cursor shipped one in July. Microsoft Foundry's now spans 30+ models. Everybody has one.

So if price isn't the reason, why is everyone building routers?

Remember that part we were right about? Turns out it's bigger than we knew. That same OpenRouter data shows how lopsided model specialization has gotten. Claude takes over 80% of programming and technical tokens. DeepSeek takes 66% of roleplay. Qwen splits the difference. Only way you get this is by finding out what each model is good at, one job at a time.

So we built a tool to help you do that. And with JSC Labs, you can download it and get started right now. Read on!

Point it at a workflow export from HubSpot, Zapier, n8n, Make or Power Automate. Get back a graded health report — with a fix written for every finding.

✅ Find the steps you pay for that don't do anything

✅ Catch workflows that've been half-finished since July

✅ Get a named cheaper model for every AI step — or an honest no

I pointed it at my own HubSpot portal first, which I'd recommend to anyone building a tool like this. Humbling, but instructive.

The tool spotted a test value left over from a previous workflow I copied. Sloppy, yes, but the tool quickly earned its place in the rotation.

Model routing is real and it's worth understanding, but it's a second-order optimization. The first-order one is that few people revisit their automations after the day they build them, and the failures are silent. A workflow that sends one email instead of four doesn't throw an error. It just quietly underdelivers until somebody looks.

One more caveat, and it's the reason I built a tool instead of quoting one. Most workflow builders have onboard AI tools you can use to check your work. But if you are going to check the AI model's work, there's no mechanism to check routing, efficiency and sloppy leftover values in the same go.

Go open the three automations you built most recently and check each one. Let me know how it turns out!

Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Of the thousands of vendors selling agentic AI, the firm estimates only around 130 are genuine, with the rest relabeling existing products. Learn More...

The transcription model shut down alongside gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize, and the Assistants API closed the same day with no automated migration tool. Calls to the retired endpoints return hard errors with no grace period; migrating a production system was commonly quoted at two to six weeks. Learn More...

A group of French daily newspapers filed a complaint with France's competition authority alleging that Google's AI Overviews have cut referral traffic to their sites by as much as 38%, and that Google generates those summaries from their reporting without compensation. Learn More...