Reduce AI costs by allocating tasks

Reading an amount from a receipt and explaining a complex situation are different jobs. Yet many applications send both to the same language model. Dividing up the work lets an application reserve its more expensive processing for the requests that need it.

System One and System Two

Cloudflare and TypeSafe use a distinction inspired by Daniel Kahneman's account of fast and deliberate thinking.

System One models handle a defined task. Jev and Clef can choose between options or assess a text's relevance. Clairic extracts values. The result arrives directly, without having to write an explanation to produce it.

System Two refers here to generative models that write, bring information together and can reason through several steps. They suit requests that need a developed answer.

System One

Choose or extractJev, Clef, Clairic

A value or a choiceDate, amount, relevance

System Two

Reason and writeGenerative model

A developed answerExplanation, summary

This distinction describes approaches. It does not measure a model's intelligence or guarantee that one method will always be faster or cheaper.

What Clairic changes for a receipt

To prepare an expense claim, an application needs a merchant, a date and amounts. Clairic accepts text or an image and returns the requested values with information about uncertainty. The software can then check the amounts and flag fields for review.

The benefit is to process the information the application needs directly. A generative model remains available when an anomaly needs explaining or someone asks a question about the document.

Clairic is still in preview. Complex calculations and long aggregations are among its limitations. A valid format does not guarantee a correct value, and a missing field should remain missing.

Where costs can fall

Start with calls that can be avoided. Reuse an approved answer. Let software perform a calculation. Give a repeated task to a specialised model, then send only the useful information to the model that writes.

We are implementing this approach in Askolia, using Jev to select useful sources before drafting.

Savings need to be checked across the whole workflow. Adding a filter while retaining the same calls can increase spending. Errors, retries and human corrections count too.

Choose a frequent task and compare both workflows on the same requests. Measure the cost per usable result, the response time and the corrections needed. You can then decide whether the new division of work earns its place.

Sources and references

• Cloudflare, Clairic and System One / System Two introduction, 19 October 2026. Vendor introduction.
• Cloudflare, extraction models, documentation updated 19 October 2026, accessed 24 October 2026.
• TypeSafe AI, System One and Jev introduction, 15 September 2026. Vendor description, accessed 24 October 2026.