
Breakdown / AI models · Automation
Cloudflare’s Clairic tackles the paperwork behind AI automation
Cloudflare’s new model pulls amounts, dates and other details out of documents. The promise is less copying and lower costs, but correct formatting isn’t the same as a correct answer.
An invoice arrives by email. Someone has to read it, find the supplier, amount and due date, then put those details into the right system. Multiply that by hundreds of invoices and the copying becomes a job of its own.
Clairic, a new AI model from Cloudflare’s Workers AI team, is built for that kind of work. It takes text or images and returns specific pieces of information that software can use. Cloudflare launched a hosted preview on October 19.
Give it the document and tell it what to find
Imagine an invoice that says the total is $240 and payment is due November 1. You ask for two fields: an amount and a date. The intended result is those two values, ready to fill the corresponding fields in an app.
Developers specify what each field should contain, including whether it should be a number, date or piece of text. Clairic returns answers with confidence information. It can also work with images, which makes photographed receipts a potential use case. Model documentation
The distinction from a decision model such as Jev or Cloudflare’s Clef is the question being asked. A decision model chooses between options you supply, such as a refund request or a delivery complaint. Clairic supplies the amount paid or the delivery date. Those answers aren’t a list of categories you can prepare beforehand. Cloudflare’s explanation
Why use a separate model for this
A general language model can already extract information. Cloudflare’s proposition is to give that repetitive step to a specialist, then use a language model where writing or more involved reasoning is needed.
Clairic processes input without generating an answer token by token, according to Cloudflare. Its listed price is $0.10 per million input tokens, with no output-token charge. The company plans to release downloadable model weights later; the announced release is a hosted preview.
For a developer building accounting or customer-service software, the appeal is straightforward: less time and money spent turning incoming documents into usable records. Whether it delivers that saving depends on how often the results need correcting.
A neat answer can still be wrong
Cloudflare explicitly warns that valid formatting does not guarantee a correct value. It also acknowledges weaknesses in longer calculations and basic forecasting. Confidence estimates can be too optimistic on unfamiliar data. Preview limitations
That leaves a practical test for Clairic: can it read the crumpled receipt, handle the missing date and distinguish the subtotal from the amount actually owed? Those are the mistakes that determine whether an automation saves someone work or gives them another queue to check.
Sources
- Clairic launch announcementCloudflare
- Clairic model documentationCloudflare