Editorial illustration for Clairic AI Reaches 98% Accuracy on Customer Support, Receipt Data
Clairic AI Hits 98% Accuracy on Support Tickets
Cloudflare's Clairic system is sorting customer support tickets and receipt data with 98% accuracy, but only on the 72.8% of questions it feels confident about. The rest get kicked upstairs to a larger model. That split is the whole point of the design: a fast, cheap classifier named Jev handles the routine cases, while a bigger model called Luna picks up what's left over.
Jev is built as a classification tool. Feed it a set of possible answers and it returns probabilities for each one, then uses a confidence cutoff (set at 0.9 here) to decide whether it trusts its own call. Below that threshold, the question gets routed elsewhere instead of forcing a guess.
The approach borrows its name and logic from a well-known idea about how people think. Daniel Kahneman split human decision-making into two modes, one fast and automatic, the other slow and effortful. Clairic's architecture copies that split almost literally, treating most support and receipt queries as the kind of routine pattern-matching that doesn't need a full reasoning model to handle.
With a 0.9 confidence cutoff, Cloudflare's System One model handles the easy cases and routes the rest to a bigger, slower LLM like Luna. Jev was the first System One model, and I think it has already made quite an impact.
Why this matters
The 72.8% figure is the real story here, not the 98%. Clairic hits that accuracy on roughly three-quarters of queries, which means the other quarter still needs a bigger model like Luna to pick up the slack. That's a reasonable tradeoff for receipt parsing and routine support tickets, where the cost of running a frontier LLM on every single request adds up fast. For founders building on tight margins, a confidence-score router that offloads most traffic to a cheap, fast model is the kind of unglamorous infrastructure choice that actually shows up on a bill at the end of the month.
The Kahneman framing is cute, but don't let it distract from the harder question: what happens in that remaining 27%? If the fallback to Luna is slow or expensive, the savings from System One shrink quickly. Anyone evaluating Clairic should ask Cloudflare for the latency and cost numbers on the full pipeline, not just the headline accuracy on the easy cases. Routing architecture is only as good as its worst-case path.
Common Questions Answered
How does Cloudflare's Clairic system achieve 98% accuracy on customer support and receipt data?
Clairic uses a two-tier routing system where the fast classifier Jev handles routine cases with high confidence, achieving 98% accuracy on those queries. For the remaining 27.2% of questions where Jev lacks confidence, the system routes them to a larger model called Luna for more complex analysis. This hybrid approach balances speed and accuracy by only using expensive frontier models when necessary.
What is the significance of the 72.8% confidence cutoff in Clairic's design?
The 72.8% figure represents the portion of queries that Jev handles with sufficient confidence to achieve 98% accuracy, making it the critical metric for the system's efficiency. Rather than attempting to process all queries with a single model, Clairic intelligently routes only the confident cases through the cheap classifier while escalating uncertain cases to Luna. This selective routing is the core design principle that makes the system cost-effective for businesses operating on tight margins.
How does Jev differ from Luna in Cloudflare's System One model?
Jev is a fast, cheap classification tool that processes routine customer support tickets and receipt data by returning probabilities for each possible answer. Luna is a larger, slower language model that handles the more complex queries that Jev cannot confidently classify. The division of labor between these two models allows Clairic to minimize costs by reserving Luna's expensive processing power only for cases where Jev's confidence score falls below the 0.9 threshold.
Why is the 72.8% accuracy figure more important than the 98% accuracy metric for Clairic?
The 72.8% represents the actual percentage of queries that Clairic can handle efficiently with its fast, cheap classifier, while the 98% accuracy only applies to that subset of questions. For founders building on tight budgets, the real value lies in offloading three-quarters of traffic to an inexpensive model rather than running a frontier LLM on every request. The remaining quarter that requires Luna's processing still represents a significant cost savings compared to processing all queries with an expensive model.
What types of tasks is Clairic designed to handle efficiently?
Clairic is specifically optimized for receipt parsing and routine customer support ticket classification, where the cost of running expensive frontier models on every request becomes prohibitively expensive. The system's confidence-score routing mechanism makes it ideal for high-volume, repetitive tasks where most queries follow predictable patterns. By handling approximately three-quarters of these routine cases with the fast Jev classifier, Clairic significantly reduces operational costs while maintaining high accuracy on the queries it processes.
Further Reading
- Clairic Hit 98% Accuracy on 72.8% of Questions via Confidence Scores - Towards Data Science
- An updated Clairic and Clairic-flash are now open-weight, from the new System One class of extraction models - Cloudflare Blog
- Introducing Clairic, a new System One class of extraction models - Cloudflare Blog
- Cloudflare Open-Sources Clairic: Receipt Accuracy Jumps From 64% to 100% - Hamidun News
- Benchmarking Cloudflare Clairic Against My LLM Receipt Parser - Harshil Patel