In brief
Clairic is a new Workers AI preview for extracting and computing typed values. For SMEs, the practical opportunity is a reviewed document-to-record workflow, with uncertain fields held back instead of silently posted.
What changed on 19 October
Cloudflare announced Clairic on 19 October 2026 as a preview on Workers AI. Rather than producing a written answer, the model returns values for fields a developer defines, such as text, numbers and dates. Its release also includes estimation and forecasting types, which should be kept distinct from facts extracted from a document.
This is a developer building block, not a ready-made accounting application. A business still needs a workflow for receiving documents, checking results and deciding which system may be updated.
Useful structure, with limits
Cloudflare’s model page describes typed output with confidence information. That can help an integration route an extracted value for checking instead of treating every response alike. The page currently lists a 65,536-token context window and image-input support; these capabilities still need testing against the documents a team actually receives.
The announcement is explicit that well-formed output can still be wrong. It recommends doing complex arithmetic in code rather than relying on long calculation chains or aggregations across many rows. It also warns that dates without a year depend on context. These are important boundaries for any document-processing pilot.
A practical SME pilot
Our suggested starting point is one document type and a small, representative test set. For example, use approved sample supplier invoices to capture the invoice number, issue date, currency and stated total. This is an illustrative workflow, not a MIRA CODE customer case or a claim that Clairic has passed our own benchmark.
Agree on the correct values before testing. Include credit notes, duplicate files, unclear scans and dates that could be read more than one way. Keep the original document beside each result so a reviewer can check the source without searching through email again.
Keep extraction separate from action
We would give each record an explicit state: extracted, needs review, verified or rejected. A missing currency or ambiguous date should stop the record from moving forward, rather than invite a guess. Recalculate totals using defined rules and compare them with the source.
Before any write to a business system, check for an existing record using identifiers appropriate to that workflow. Save the outcome and the destination record reference so a retry cannot quietly create a second entry. Sending an invoice to an extraction model must never, by itself, authorize a payment.
Decide who can see the documents, what data the provider receives and how long copies are retained before using live business records. Start with approved, non-sensitive samples while those decisions are being made.
MIRA CODE’s perspective
The useful question is whether a team can complete a specific task with less correction work while retaining control. Measure field accuracy, exceptions and reviewer effort before expanding the pilot. Avoid promising savings from a model announcement alone.
This fits the work MIRA CODE describes in AI automation, document analysis and connected business processes: designing the checks and handoffs around the technology. Clairic is a new option worth evaluating, not a reason to replace a working process without evidence. Product details were checked against Cloudflare’s official sources on 21 October 2026.
