Cloudflare previews Clairic for turning business documents into typed values

Cloudflare announced Clairic on October 19 as a preview on Workers AI. Instead of asking for a written response, an application supplies input information and named fields with types, such as text, a number or a date. The model returns values for those fields.

The company's model reference includes an invoice example with separate fields for an identifier, a calculated due date and an estimated time to payment. These are different kinds of answers even though an application can receive them in one structured response.

Cloudflare's launch post describes the hosted endpoint as an early checkpoint and says an open-weights release is planned. That statement is a plan, not evidence that the weights have already been released.

Cloudflare's extraction-model guide distinguishes a value present in the input from a calculation derived from supplied facts. It separately defines estimation as predicting an unknown outcome and forecasting as predicting a sequence of future values.

Consider a hypothetical delivery note that states an order reference and two dates but says nothing about repair costs. Copying the reference is an extraction task. Calculating elapsed days needs an explicit definition of which dates count. A repair-cost estimate would require different evidence and should remain labeled as an estimate.

An operations record could retain those distinctions as separate fields: the value, how it was obtained and the source used. That is an editorial design suggestion, rather than a required Clairic integration or a report of a customer deployment.

The model reference shows text answers with source-character positions and includes an example where an absent purchase-order number produces a null value. Its estimated payment-time example instead includes a median and prediction interval.

The extraction-model guide warns that null can mean information was missing or that the model failed to identify it. Estimates and forecasts can still return predictions when the input supplies little information. Treating every populated field as equally well supported would erase that difference.

For a team reviewing an imported record, retaining the relevant source excerpt and the response metadata would make a disputed field easier to trace than saving only the final value.

Cloudflare's launch post cautions about longer chains of arithmetic and aggregation across many table rows. It recommends leaving the math to code where appropriate. It also says prediction intervals can be too narrow on unfamiliar data and describes forecasting as basic.

The October 19 changelog makes another limitation explicit: a correctly formed answer can still be wrong. It advises using confidence, missing-value probabilities and intervals when deciding which results need review.

These are the vendor's documented limitations, not OrangeHat test results. This article reports no benchmark, cost saving or accuracy improvement from using the preview.

A useful implementation discussion starts with the meaning of each field. The following questions translate the documentation into a business-record handoff without assuming that a model response has already been verified.