Your AI agent just received a customer message. A checkout has been failing for every user for the past hour, and someone needs to decide whether this is a priority-one emergency, which team should handle it, and how quickly a response needs to go out.
A traditional large language model can answer all of that. It might take half a second, cost a few cents per query, and return a full paragraph when you needed three simple answers. That is not efficient at scale.
Decision models solve exactly this problem. They are a new category of AI model designed not to generate text but to make fast, structured classifications, returning typed answers with probabilities your code can act on directly. No explanation required. Think of them as a highly trained judge for your automated workflows.
This week, Cloudflare released Clef and Clef-flash, two open-source decision models now leading accuracy benchmarks. Clef responds in a median of 209 milliseconds, while the leading alternative takes over 524 milliseconds for the same classification tasks. That is not a marginal difference.
The timing matters. Cloudflare is not moving alone here. Amazon released Strands Decider 2B, Perplexity released pplx-decider-v1-27b, and a family called Kev covers everything from 0.8 billion to 27 billion parameters, all in the same week. A new category of AI infrastructure is forming, and multiple players are staking out positions at once.
What does this mean for a business that is not building AI models?
Think about any workflow where a human currently reads something and makes a quick judgment call. A support ticket arrives and someone decides its priority. A document comes in and gets routed to the right department. A transaction flags and needs to be classified as legitimate or suspicious. Each of those is a decision with a finite set of outcomes, and decision models are built to handle all of them at speed and scale without the overhead of a full language model running every single time.
Faster and cheaper, yes. Also more predictable. Cloudflare’s Clef classified a website domain, fetching and rendering the page and scoring it across multiple categories, in 2.2 seconds, while their fastest general-purpose model took 4.7 seconds and returned fewer results. That is a two-to-one edge on the same task.
There is a deeper opportunity for companies with historical data. A business with years of support tickets, routing decisions, or fraud reviews has exactly the labeled training data needed to fine-tune a decision model to its specific context. Cloudflare is already offering a fine-tuning service for this reason, and smaller, more specialized models consistently outperform general-purpose ones within their specific domain.
This is not a replacement for language models. When agents need to reason, write, or generate something, a full LLM is still the right tool. For routing, scoring, triage, and classification, a decision model is faster, cheaper, and more consistent.
Not every AI decision in your pipeline needs the full weight of a language model behind it. Start identifying the classification tasks in your existing workflows. They are probably cheaper to automate than you think.
Want to explore how AI automation could benefit your business? Let’s talk.