How to Use Jev: Email Triage, Lead Scoring, and Fast Routing

Direct Answer

Jev is most useful when information arrives faster than a person can sort it, but the next step is a bounded decision. In this episode, Greg Isenberg asks Ryan Vogel to explain the model through a real inbox. Ryan feeds Jev 1,700 emails and asks it to classify category, priority, spam likelihood, and whether a reply is warranted. He reports a total model charge of 18 cents for that run.

The interesting business pattern is not “replace the whole team.” It is put a fast, inexpensive decision layer at the front of a costly queue, then route high-value or uncertain items to a person, low-risk routine items to an approved workflow, and the rest to further review. Ryan calls the model advisory; his failed Bitcoin signal experiment is a useful reminder that a fast answer is not necessarily a good one.

Watch Greg Isenberg and Ryan Vogel

Credit: this article is based on Greg Isenberg's conversation with Ryan Vogel and the supplied transcript. Follow Ryan on X or his YouTube channel. Product details are cross-checked against TypeSafe and Vercel's documentation; demo results remain Ryan's observations.

What Jev Actually Returns

Ryan describes Jev as a classifier, but TypeSafe's API gives that term a more useful shape: send a state, define typed questions, and receive structured answers. A Choice selects among supplied options and returns their probabilities. A Score evaluates an ordered rubric. TypeSafe calls its 0-to-1 true/false-style primitive Noul; Vercel's AI SDK presents a Boolean question type. Several questions can be evaluated against the same state in one call.

For an email, the state might include sender, subject, body, and known account context. The questions might ask which queue it belongs in, how urgent it is, and whether it looks like spam. Jev does not write the categories for you or compose a reply. The application supplies the schema and decides how to act on the returned values. This makes the model a component in a workflow, not a conversational replacement for ChatGPT or Claude.

Probabilities are not the same thing as verified truth. A result that assigns 0.9 to “spam” should be tested against labeled email examples before your system deletes, archives, or ignores messages. TypeSafe's confidence guidance recommends different thresholds for different levels of risk and an escalation path when the model is uncertain.

The 1,700-Email Demo: What the Numbers Mean

Ryan runs about 1,700 emails through four evaluations: category, priority, spam score, and reply likelihood. He reports roughly 4.2 million input tokens, 500,000 output tokens, and an 18-cent model charge. The visible result is a sortable inbox with decisions attached to every message. These are values reported in one creator demonstration, not an audited guarantee for every account or workload.

The number is plausible against Vercel's listed $0.04 per million input tokens at the time of writing: 4.2 million input tokens alone would be about $0.17 before rounding. But the full cost of an email system also includes access, storage, setup, any separate AI steps, human review, and the cost of a missed important message. The episode's “about 200 milliseconds per query” is Ryan's reported experience, not a latency promise independent of payload size, network, or load.

Put Jev at the Front of the Right Queue

The episode moves from email to a practical startup question: which business has an expensive stream of incoming requests that someone must repeatedly sort? Greg's “traffic cop” framing is helpful, provided the traffic cop does not silently become the final authority.

• Lead scoring: Ryan describes a design agency scoring contact-form submissions so promising leads get a quicker personal response. A score should prioritize attention, not decide that a lower-scored person is unworthy of service.
• Support routing: classify product area and urgency so a ticket reaches the correct team. Compare results with a reviewed set of historical tickets and make manual correction easy.
• Local-service matching: a request such as driveway cleaning could be matched to eligible nearby providers. Jev can help rank fit, but availability, distance, service terms, and a binding price still need verified data and business rules. An “instant match” is not automatically an “instant quote.”
• Video clipping: Ryan transcribes a video, then uses Jev to score 17 candidate moments in about three seconds. Transcription and clipping are separate steps; the decision model ranks moments rather than editing the video itself.
• Browser control: a shown flight-selection demo completes in 7.1 seconds. Treat this as a demonstration of fast routing inside a browser agent, not evidence that it will safely complete purchases or handle every travel site.

The Bitcoin Test Is the Most Useful Caveat

Ryan also asks Jev to produce a buy, hold, or sell signal for Bitcoin every minute. He says it performs poorly and advises against putting it in front of a portfolio. The contrast explains the product boundary: routing a support ticket among known queues is a narrower task than forecasting a market with incomplete, changing information. In the video, a frontier model does somewhat better because it can incorporate news, but that is not evidence that either model should control trades.

There is a second boundary: Jev does not make an unsafe workflow safe simply by returning structured output. A valid schema can contain the wrong classification, an overconfident score, or missing context. The workflow needs a labeled test set, thresholds, a review queue, logs, and a way to undo mistakes. For a deeper examination of the model and its reliability claims, see our Jev explainer.

How to Try Jev Today

1. Pick one reversible decision. Start with email priority or support queue assignment, not a financial trade or automatic outbound message.
2. Define the possible answers. Use TypeSafe's quick start to send a state and one or more Choice, Score, or Noul questions. Include an “unclear” route in your surrounding workflow.
3. Choose access. TypeSafe offers its own documentation and sign-up path. Vercel says Jev is available through AI Gateway as typesafe-ai/jev, even while direct access may be staged. Confirm current availability and pricing before relying on it.
4. Measure before automating. Compare Jev's labels with human-reviewed examples, record latency and total cost, and send low-confidence or consequential cases to a person. Then expand only the part of the queue where it performs reliably.

This is the opportunity Greg and Ryan identify: faster decisions can remove waiting from a business workflow. The useful product is the whole reviewed service around those decisions, not just a model call that looks fast on screen.

Video Chapters

Sources and Further Reading

• Greg Isenberg and Ryan Vogel: Jev is HERE. How to use it (video and supplied transcript)
• TypeSafe: Jev introduction; quick start; confidence guidance
• Vercel AI Gateway: Jev model and current pricing; availability announcement
• Ryan Vogel on X; Ryan Vogel's YouTube channel

YouTube lists the video's publication date as 18 September 2026. This article was reviewed on 18 September 2026. Prices, access routes, and latency can change.