JEV at the heart of InsightNarrator: the fast judgment Voice of the Customer was missing

Every day, a company receives thousands of sentences from its customers: reviews, tickets, survey answers, chats. Understanding what they say is the easy part. The hard questions are different: what do they actually want? How do they feel? How urgent is it? And above all: what should the company do right now?

The first three are questions of judgment; the last is a question of decision. For years, Customer Experience teams have tackled them in two unsatisfying ways: rigid rules that miss nuance, or general-purpose LLMs too slow and too expensive to apply to every message. With JEV, part of that problem opens up.

1. A model that doesn't talk, it decides

On September 15, TypeSafe, an AI lab that spent two years in stealth, released JEV, a model that classifies text and structured data instead of generating it. It doesn't converse and it doesn't write: it receives a context and typed questions, and answers each one with an option from those you defined, along with the associated probabilities. The name comes from the Jevons paradox: when an operation becomes cheap enough, you end up doing far more of it.

"System One" echoes Kahneman: fast, intuitive thinking, as opposed to slow, deliberate reasoning. Pricing explains part of the reception: at $0.042 per million input tokens with free output, a judgment that used to be "an LLM call you had to justify" becomes something you can run on every request. According to TypeSafe, end-to-end response times range from 70 to 500 milliseconds. Within the first 24 hours after its release on Vercel, about 13% of the platform's paying teams tried it, the fastest adoption of any new model on the platform.

2. Why intent and emotion detection are its natural territory

Recognising intent ("wants to cancel", "requests a refund", "is comparing competitors") and emotion ("frustration", "disappointment", "enthusiasm") are closed-taxonomy problems. JEV is built for exactly this:

• Always-valid categories. The model never answers outside the options you've defined. For an analytics pipeline, that means no invented labels, and data that enters the rest of the system already structured.
• Probabilities, not just labels. With probability distributions you can set thresholds: above a given value the insight proceeds automatically, below it goes to review.
• Real-scale cost and latency. You analyse 100% of feedback, not a sample, and in real time.

In InsightNarrator, JEV powers the CORE intent and emotion detection engines. In our tests on 1.5 million feedback items in different languages, we observed 96.7% accuracy, at a cost of a few cents per 1,000 feedback items compared with the LLM solution.

3. Where it fits in the swarm

InsightNarrator is not a model with a chat window on top: it is a swarm of ten agents coordinated by NarratorGPT, with specialists such as SentimentAI, ThemeMapper, SegmentLens, TrendRadar, RiskAlert, Friction Economist, Data Scientist, Domain Expert, Synthesis Analyst, and Quality Auditor.

In this design, JEV does a precise job: it produces the elementary signals (intent, emotion, urgency) that the specialised agents then aggregate, cross-reference, and interpret. SentimentAI and ThemeMapper no longer have to "read" every sentence from scratch, because they start from judgments that are already structured and carry a declared probability. Slow reasoning (why is a segment's sentiment declining? which friction costs the most?) stays where it belongs, in the agents that can afford it.

On top of this judgment layer sit functions that go beyond classification:

• ABSA taxonomy: sentiment is not assigned to the whole message but to the individual aspects the customer talks about (price, delivery, support), which is what makes an analysis actionable.
• C-Risk analysis: reading feedback through a risk lens, where intent and emotion become leading indicators.
• Surveys: the collection of structured feedback that feeds the same pipeline.
• Scoty: the conversational interface where users build their own analyses by talking to the platform.
• ROI Builder: translates insights into economic value, because an analysis that doesn't end in a business number rarely ends in a budget.

4. Without replacing what you already have

Most CX teams don't need another place to collect feedback. They already have one, often several: a survey platform here, a text analytics tool there, a ticketing system on the side. Ripping that out to try a new model is a non-starter.

InsightNarrator was designed to work on top of 25 Voice of the Customer systems, including Medallia and Qualtrics. It reads the feedback where it already lives, adds the intelligence layer (intent, emotion, aspects, risk, decisions), and returns insights without a migration project. That is also why a new engine like JEV can be adopted quickly: it plugs into the analysis layer, and your data sources stay exactly where they are.

5. The honesty that's needed: a wrong judgment looks identical to a right one

JEV is not magic, and anyone adopting it in production should know that. An independent analysis notes that every answer fits the format you defined, so a wrong answer looks exactly like a correct one; in their tests, text planted in the input pushed the model toward the wrong option with confidence above 70%.

In Voice of the Customer this matters more than elsewhere, because the text being analysed is written by untrusted third parties: reviews, comments, emails. A fast model alone is not enough. You need something around it.

6. "Beyond the Model": the harness as a system

This is where our academic work meets the news cycle. In the paper Beyond the Model: Designing the Harness as a System, we argue that the real differentiator of an AI product is not the model but the harness around it, described as a system of seven organs: constitution, knowledge, context management, memory, guardrails, verification and feedback, observability.

JEV is an almost perfect case study:

• The untrusted-input problem is a guardrails issue: customer text is data to be analysed, never instructions to be followed.
• The risk of the "wrong answer that looks right" is a verification and feedback issue: in InsightNarrator's harness, the Quality Auditor and the Goal-Loop pattern check the agents' work before it becomes an insight.
• The question "what did the model read when it got it wrong?" is an observability issue.
• Confidence thresholds and the role of the Domain Expert concern knowledge and constitution: what counts as an acceptable decision in a given domain, and who is allowed to make it.

The consequence is our thesis: models get replaced, architecture endures. We integrated JEV in a matter of days because the harness was already ready to receive it, and we will be able to replace it or pair it with something better without rewriting the product. Our work on the A2A protocol points in the same direction: interchangeable agents and models inside a stable orchestration.

7. What changes for CX teams

• From sample to totality: every piece of feedback is classified by intent and emotion, not only the ones someone has time to read.
• From report to signal: urgency and risk surface while the customer is still writing.
• From data to decision: intents and emotions feed priorities, routing, and actions, with a human in the loop wherever confidence is low.
• From cost to measurable investment: the ROI Builder ties insights to an economic value you can defend in front of leadership.
• From replacement to augmentation: no migration and no new collection tool, because the intelligence layer sits on the systems you already run.

8. Try it on your own feedback

JEV is one of the engines that make InsightNarrator a Decision Orchestration platform for Voice of the Customer: fast judgment where it's needed, reasoning and control where it matters. If you want to see how intent and emotion detection work on the feedback you already collect in Medallia, Qualtrics, or any of 25 VoC systems, start a free 14-day trial or book a demo.