When a customer explains a problem, a business first needs to know who can resolve it. Delivery, billing, refunds: even before producing an answer, the case needs routing. That largely invisible task is at the heart of Clef, a new model family announced by Cloudflare.

What happened: constrained choices

On 1 October 2026, Cloudflare presented Clef and Clef-flash. Its documentation describes tools designed to choose among predefined answers. Instead of free-form text, they return a probability for each permitted option. The official presentation cites customer request classification and deriving decisions from documents among their uses.

The models can process several kinds of information, including text and images. Their weights are published under the Apache 2.0 licence, according to Cloudflare's changelog and model card. This makes it possible to study or adapt their use under that licence's terms.

Why it matters: the organisation behind the chatbot

Our analysis: AI can transform a service without ever speaking to its customer. If it classifies a request correctly, it can help reduce unnecessary transfers. If it makes a mistake, the customer may instead have to repeat their story to several departments.

Constrained choices make outputs easier to compare. But they also restrict the system to the options provided. A new or ambiguous request does not become familiar merely because the model assigns it a category.

What it could change for you: less waiting, under conditions

For support teams, the promise is more time for complex cases. For the public, the issue is the quality of the journey: reaching the right person, understanding why a request was routed that way and being able to report an error.

A high probability expresses the model's preference among the proposed options. It proves neither a document's truth nor the validity of a decision affecting a person. Our editorial assessment is therefore that a classification system needs a path for uncertain cases and human recourse when consequences are significant.

Cloudflare publishes evaluations of its models, but these remain the supplier's own. They do not alone demonstrate that an entire customer service operation will become faster or more satisfying.

Conclusion: solving the problem remains the right measure

Clef illustrates a trend towards specialising AI in precise stages of work. Its real value will be judged by requests handled correctly, mistakes corrected and users' experience. Efficient classification is only a beginning; a useful answer still needs to follow.