AI Agents Need to Make Decisions. Cloudflare Just Built the Right Tool.
Your AI agent receives a customer support ticket. Before it can do anything useful, it needs to decide: is this urgent? Which team should handle it? Is this a billing issue or a technical one? A large language model can answer all of that, but it costs you several seconds and real money every single time. Multiply that by ten thousand tickets a day and the math gets uncomfortable fast.
Cloudflare released a direct solution to that problem this week. The company launched Clef, a family of open-source decision models purpose-built for fast, structured choices inside automated workflows.
What a Decision Model Actually Does
Decision models are a distinct category of AI model that classifies rather than generates. A standard language model will write you an essay about whether a ticket is urgent. A decision model returns “urgent: yes, 91% confidence” in under a second, in typed output your code can act on directly. That narrow focus is what makes them faster and far cheaper to run than their general-purpose counterparts.
Think of it as the relay switch in an electrical circuit. It does not generate power. It decides which direction the current flows.
What Clef Brings to the Table
Clef and its faster sibling Clef-flash are now live on Cloudflare’s Workers AI platform, and both are fully open-source under an Apache 2.0 license. Your team can run them locally, embed them in existing pipelines, or fine-tune them on company-specific examples. Cloudflare also launched a reinforcement learning platform alongside Clef that lets businesses train the model on their own data without needing a machine learning team or a background in AI research.
The performance numbers are clear. Clef classified a website domain in 2.2 seconds, returning multiple categorized results. A large general-purpose language model took 4.7 seconds to do the same job and returned fewer results. On the Jev Decision Index, the emerging benchmark standard for this model category, Clef currently holds the top position.
The Detail Most People Will Miss
Clef has a vision encoder built in. It classifies images, not just text. Routing product photos by category, flagging documents by type, or moderating visual content are all possible within the same decision layer your workflow uses for text. Most competing decision models today handle text only.
Clef also ships with a 64k context window, double what the leading alternative model offers. That means more of your workflow state can be passed in when the decision needs additional context.
Why This Matters for Business
The most common bottleneck in production AI automation is the routing step. Something has to decide what to escalate, what to ignore, which service to call, and who to notify. Teams typically solve this by feeding every decision into an expensive reasoning model, which works but burns budget and adds latency. Decision models handle that routing cheaply and at speed, reserving your heavier model for situations that genuinely require reasoning. That is smart architecture and real cost savings in production.
Open-source, multimodal, fine-tunable from day one. Clef is infrastructure worth paying attention to.
Want to explore how AI decision models could benefit your business? Let’s talk.