We do not need a large LLM for every decision. With Jev, TypeSafe AI brings a practical question back to the centre of enterprise software design: how much intelligence does choosing the next action actually require?
Released in early access on September 15, 2026, Jev is TypeSafe’s first « System One Model ». The company describes it as a model designed to produce structured decisions with probabilities rather than free-form text. The official Jev announcement explains the approach.
For companies adopting AI-native software with CraftOS, this is about more than a new model. It raises questions about operational speed, control over AI usage costs and how different models work together.
What changes with Jev as a decision model?
Classification is not new. Machine learning has long answered well-defined questions: spam or not spam? Support, sales or billing? Low, medium or high risk?
Consider an incoming request: “I would like to order 50 units, but I need to know the delivery time.” The software must recognise purchase intent, identify missing information and route the request to the right place. It does not necessarily need a lengthy written reasoning process for each step.
With Jev’s approach, the developer defines the possible answers and their structure. The model receives the context and returns values that software can use directly, along with probabilities. A decision becomes a step in a business process.
Jev and LLMs play complementary roles
A general-purpose model such as those used in ChatGPT, Claude or Gemini can write, code, analyse, reason and classify. LLMs can also produce structured outputs. The distinction is therefore more nuanced than “one writes, the other decides”.
A specialised architecture focuses its resources on a particular task. A generative model remains useful for drafting a personalised sales proposal. Routing a request between a few known categories may only require a more specialised solution. And if a deterministic rule solves the problem reliably, no model is needed.
This is a central question in AI engineering and model integration: choose the right combination for the task, acceptable response time, cost of an error and processing volume.
Jev latency and pricing: TypeSafe’s launch claims
In its launch announcement, TypeSafe reports the following figures for Jev:
- 70 to 500 ms end-to-end response time;
- 40 to 200 times faster than the LLMs compared on the decision tasks tested;
- $0.042 per million input tokens, with no charge for output tokens.
These are the vendor’s figures at launch, not a guarantee for every application. Test conditions, service location and the models compared affect the results. TypeSafe discusses these limitations in its performance and methodology overview.
A business should measure its own complete process: receiving the request, accessing data, making a decision, taking action and, where needed, obtaining human approval. A fast model response is not enough if the rest of the process takes several minutes.
The same applies to pricing: one million tokens does not mean one million decisions. Actual cost depends on the amount of context sent, the number of calls, retries and the other services involved.
Why AI-native software multiplies small decisions
In an AI-native system, a single request can trigger several decisions in the background. Which information should be retrieved? Which process should run? What action comes next? Can the system continue automatically, or does someone need to approve it?
Across a business, those choices can add up to thousands or millions. Always using the most powerful model may introduce delays and additional AI usage without improving every outcome.
This directly affects AI agent orchestration: an AI agent may need substantial reasoning to prepare a response while relying on simpler steps to classify, route or check information. Three business examples illustrate the principle.
CRM: qualify and route sales enquiries
A prospect reaches out on WhatsApp, by email or through a social network. Before responding, the system needs to distinguish a pricing question from purchase intent or a support request. It also needs to identify missing information and recognise when to hand the conversation over to the sales team.
This is the kind of process an AI-native CRM such as CraftOS CRM supports: bringing customer context together and enabling a WhatsApp AI agent to qualify enquiries, respond and move the sale forward. Each step calls for a different level of reasoning.
Invoicing: distinguish a payment follow-up from a dispute
“I will pay on Friday” and “The amount does not match the quote” should not trigger the same action. Before sending another reminder, the system needs to understand the reply and check the account. A dispute should be routed to a person when necessary.
In a quoting, invoicing and payment tracking process with CraftOS Invoicing, value comes from continuity: an understood message, an up-to-date record and an appropriate next action.
Cloud: choose the model to suit the task
Preparing a presentation, analysing a tender document and classifying incoming requests are three different jobs. A multi-model architecture allows you to assess which model fits each one, based on the required quality, the data involved and AI usage.
Explore this approach through CraftOS Cloud and its built-in AI applications. These examples illustrate potential uses for specialised models; they are not an announcement of a Jev integration in CraftOS.
Before automating a decision, test what matters
A valid output format does not prove that a decision is correct. Probability estimates also need to be assessed against situations that reflect your business.
To move from a demonstration to production, we recommend starting with a limited scope:
- Choose a recurring decision. For example, routing incoming requests to the right department.
- Build a set of real cases. Include incomplete messages, ambiguous situations and the languages your customers use.
- Compare the options. Business rules, a classifier, an LLM or a specialised model: measure quality, response time and the cost of the complete process.
- Plan the human handover. Define when the system must stop or request approval.
- Monitor errors in production. Record decisions, collect corrections and reassess thresholds.
These choices involve both technology and enterprise AI governance. Access to information and its processing must also follow your data security and AI compliance requirements.
The right level of intelligence, in the right place
Jev does not replace general-purpose models, and it did not invent classification. It brings a useful question into focus: does every step of a process need a large model?
With AI-native systems, the goal is to build a coherent whole: rules when the problem is deterministic, specialised models when the task is specific, and general-purpose models when broader reasoning is needed.
Use just the level of intelligence needed, in the right place, at the right time and at the right cost. That is how AI can become part of a company’s everyday operations.
Let’s discuss the processes you want to automate with AI →
← All articles