Cloudflare has previewed Clairic, an AI model that extracts specified values from documents, on its AI inference platform, Workers AI. Users specify fields such as order numbers or invoice amounts, and the model returns values in a specified format along with confidence scores. The model’s open weights are also expected to be released soon, and Cloudflare has shared examples of workflows that combine it with an LLM.
An AI that extracts fields, not one that writes text
When people think of generative AI, many probably picture something that fluently writes text. Clairic, however, has almost the opposite character. Its role is limited to extracting specific fields, such as “pull the total from an invoice” or “find an order number in an email.”
Its design is distinctive, too. General-purpose LLMs generate text one word at a time, or token by token, but Clairic does not generate output that way. It returns answers for all the specified fields in a single pass. If a document contains no matching value, it returns “null,” rather than trying to make up an answer—a feature that makes it well suited to practical use.
In addition to extraction, Clairic is said to support simple calculations, estimates, and predictions. It may be easiest to think of Clairic not as a writer, but as a “prep worker” that organizes scattered information into a usable format.
Models divide up customer inquiries
One of the use cases Cloudflare presented is particularly suggestive: a customer complaint email. First, Clairic extracts the order number, amount, and details of the damage from the email, then estimates the repair cost.
Next comes Clef, a model for decision-making. It chooses whether to issue a refund or arrange a repair based on company policies. Finally, the LLM GLM-5.3 writes a response to the customer.
What’s interesting is that the entire task isn’t handed over to one giant model. In a workplace analogy, it’s more like a team in which an intake person organizes the information, a decision-maker sets the policy, and someone skilled at writing drafts the response. Cloudflare calls the small models that handle these fast tasks “System One” models and envisions chaining them together with LLMs.
Personally, I think this idea is the heart of the announcement. When adopting AI, it’s easy to focus on choosing a single model that’s as smart as possible. But when you break down the work involved, not every task requires full brainpower. Designing systems that use different tools for different roles could become standard practice for business AI.
Speed figures show the benefits of dividing up the work
Let’s look at the numbers. In Cloudflare’s comparison, the three-stage workflow described above took 19.3 seconds. Using only GLM-5.3 for the same three steps took 61.7 seconds and cost nearly twice as much, the company says.
These figures, however, come from a specific workflow. The same gap won’t necessarily appear in every business process; it’s best to treat them as an example of how much dividing up the work can change things. Still, the fact that processing took more than three times as long is a reason to consider the value of splitting up tasks.
The limitations are also clear. Clairic cannot yet reliably handle long chains of calculations. For now, developers will need to run such tasks using their own code.
Even so, a world populated by specialized models that do the tasks they’re good at quickly and cheaply is appealing for businesses. The benefits of dividing up the work should be greatest for standardized tasks that arrive in high volumes every day, such as invoice processing and customer support. Let LLMs focus on the work only they can do. Clairic offers a glimpse of how that future might begin.
Comments