QUAICU
Most of the AI work inside a business is not writing. It is deciding: which team should take this ticket, does this customer need a person, is this message allowed, which model should answer this request. Large language models can make those decisions, but they are slow and costly for it, and their answers come back as text you then have to parse.
Jev, a model from TypeSafe AI, is built for exactly that kind of decision. This guide covers what it is, where it pays off, how to roll it out at enterprise scale, and what others have measured with this kind of system. We implement Jev for client teams, and you can read more about that on our Jev implementation page.
What Jev is, and what it is not
Jev reads text and answers structured questions about it. According to TypeSafe’s documentation, each call takes the content to evaluate and a list of questions, and answers them together in one pass. There are three kinds of question:
- Yes or no, returned as a probability between 0 and 1.
- Choice, returning the selected option with a probability for every option.
- Score, placing the text on a scale of 2 to 10 ordered levels.
It does not write. TypeSafe is explicit that Jev cannot produce prose or code or hold a conversation. Its published price is $0.042 per million input tokens with no charge for output, because the output is a handful of probabilities rather than paragraphs of text. It entered early access on 15 September 2026, and it is also listed on Cloudflare’s model catalogue.
TypeSafe also documents known weaknesses: it is not reliable at counting or exact numbers, its accuracy drops when unrelated material is included in the input, and score levels should not be treated as exact distances. A good implementation is designed around those limits rather than surprised by them.
The best uses
1. Support ticket classification and routing
This is the most direct fit. Each incoming ticket gets a category, a product, an urgency level and a language, and is routed to the right queue. TypeSafe’s own guide to building with Jev uses this example, with tickets below a confidence of 0.8 sent to a person.
Automated ticket routing is well studied. Microsoft’s DeepTriage, a peer-reviewed system for routing Azure incidents to the right team, reached 82.9% F1 and has run in production since 2017. Accenture reports that automated classification in ServiceNow assigns about 2,000 of its roughly 10,000 daily internal IT incidents without a person. Those systems used other models, but the problem and the measurement are the same.
2. AI cost control through routing
If you already use large language models, the cheapest call is the one you don’t make. A fast classifier in front of your models can decide which requests a small model can handle and which need a stronger one. Research on this pattern is encouraging: the RouteLLM team reported cutting costs by over 85% on one benchmark while keeping about 95% of GPT-4’s quality, and the FrugalGPT paper reported savings of up to 98% on its benchmarks. Treat those as best cases from research settings, not promises.
TypeSafe publishes a cascade example in which a small model extracts fields, Jev checks each field, and an expensive model is called only when Jev thinks a field is likely wrong.
3. Escalation to a person
In customer chat, the expensive mistake is keeping someone with a bot when they needed a human. A yes-or-no question such as “does this conversation need a person now?” with a threshold you control is a simple, auditable escalation rule.
4. Guardrails and moderation
Screen user input and AI output against your policies and record a probability for every decision. TypeSafe’s guardrails cookbook maps hazard probabilities to pass, review, block or support. Because each decision is a number, you can show an auditor exactly where the line was drawn.
5. Privacy and DPDP data mapping
India’s DPDP duties for most businesses start in May 2027, and the first step in every compliance guide is knowing where personal data lives. Classifying records, documents and messages by whether they contain personal data is a practical first pass. It should be backed by a person reviewing anything uncertain. No automated scan finds everything.
6. Inbox and alert triage
Shared inboxes and operations alerts are full of noise. Separating real requests from the rest, and grouping duplicates, frees people for the work that matters. In IT operations, PagerDuty’s customer IG Group reported an average 58% reduction in alert noise using this kind of grouping.
How to implement it at enterprise level
The model is the easy part. What makes a classifier safe to rely on is the process around it.
1. Agree the labels. Write down every category with a definition and real examples, agreed with the team that will live with the results. Decide what a wrong answer costs in each category. Misrouting a billing question is cheap; missing a security incident is not.
2. Build a test set. Label a sample of real, recent tickets or messages as your reference. This is the yardstick for everything that follows.
3. Measure per category, not on average. Report precision and recall for every category and look at the confusion matrix. A system that is 92% accurate overall can still be poor on the rare categories that matter most.
4. Run in shadow mode first. Let the classifier run alongside your current process without acting on its answers, and compare. Only switch it on when the numbers hold up on live traffic.
5. Set confidence thresholds with a person behind them. Apply high-confidence results automatically and send the rest to a person. Set thresholds per category, and watch the size of the review queue. If too much lands there, the labels or thresholds need work. Under data-protection law, a person reviewing a significant decision must make a real judgment, not just approve it.
6. Connect your tools. Helpdesks such as Zendesk, Freshdesk, Jira Service Management and ServiceNow can send a webhook when a ticket arrives and accept tags, fields or an assignment back through their APIs. Jev is also available through AI gateways such as Cloudflare AI Gateway, Vercel and LiteLLM.
7. Log and monitor. Record every decision with its probabilities and the input it was based on. Recheck accuracy on fresh samples regularly, and update the labels when your products, customers or language change.
8. Agree the data route. Decide what text is sent, how personal data is minimised first, and which route the data takes. TypeSafe’s privacy policy says it does not train models on customer inputs; zero data retention is offered to enterprise customers, and Cloudflare lists Jev with zero data retention.
A quick cost example
Using TypeSafe’s published price of $0.042 per million input tokens, 10,000 tickets of about 500 tokens each come to 5 million input tokens, or about $0.21. That is our arithmetic on the list price, before any gateway or engineering costs. The real cost of a classification project is the work of labelling, testing and integrating it well.
Public case study: Jev in candidate screening
The clearest public example of Jev in production comes from TypeSafe itself. In its Jack & Jill case study, replacing part of an LLM-based candidate screening step with Jev cut cost by 88% and halved median screening time per candidate, from 20.3 to 10.3 seconds. The change in quality was not statistically significant, which is the right way to read it: the same decisions, faster and cheaper. It is a vendor-published case, so test the pattern on your own data before relying on similar numbers.
When not to use it
Don’t use Jev where the job is to write something, to do exact arithmetic or counting, or where the input is long and full of unrelated material. Don’t use any classifier to make a significant decision about a person without meaningful human review.
Talk to us
We implement Jev for client teams, from labels and test sets to helpdesk integration, monitoring and human review. If you want to explore it, message us on WhatsApp or see our Jev implementation page.
Jev is a product of TypeSafe AI. QUAICU is not affiliated with TypeSafe AI. Product details come from TypeSafe AI’s public documentation and were last reviewed in October 2026.