Cloudflare's Clairic pulls typed fields from documents in one pass. Its chain ran 19.3 seconds against 61.7 for an all-LLM version

Most extraction steps in business automations are still an LLM asked to "return JSON." Cloudflare is betting that job needs a different kind of model.

What happened

Cloudflare released Clairic in preview on Workers AI, with open weights planned. You name the fields you want, such as order number, delivery date or total paid, give each a type, and Clairic answers every field in a single pass without generating tokens. Each answer comes back typed, with a confidence score and a probability that the value is missing. Text values include character offsets showing where they came from, and missing fields return null. Cloudflare says it charges only for input.

In Cloudflare's example, Clairic reads a damage complaint email, a decision model called Clef applies the refund policy, and GLM-5.3 writes the reply. That chain took 19.3 seconds against 61.7 seconds for GLM-5.3 doing all three steps, and cost 49% less. Cloudflare also says Clairic's repair estimate was identical every run, while the LLM's changed.

Cloudflare is clear about limits: one or two step calculations seem reliable, longer chains of arithmetic are not, so let code do the math. Forecasting is basic.

My take

Extraction is where I see the most silent failures in client automations. Invoices, order emails, intake forms. The LLM usually gets it right, then one day it misreads a digit or invents a field, and nothing downstream notices.

Clairic's design gives you the things I normally build by hand: typed output, a null instead of a guess, and a confidence number you can route on. A practical pattern:

1. Extract with a typed model.
2. If confidence is high, continue automatically.
3. If it is low or null, send it to a person with the source span highlighted.
4. Do totals and date math in code, not in the model.

It is a preview, so I would test it on a client's real documents before trusting it. But the idea of splitting extract, decide and write into separate steps is worth applying today, whatever models you use.