Most AI features in web applications do not need free text at all. They need a decision: does this request belong in sales or in support? Is this record complete enough for approval? Is this comment spam? Nevertheless, a large language model that was actually built for open-ended conversations is often connected for such tasks – and its response then has to be laboriously turned back into a usable form.

This is exactly where a new approach comes in. The company TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has presented a model called Jev that deliberately does not generate text. It delivers only classifications, i.e. the assignment of an input to one of several predefined categories. Response times are quoted as starting at 70 milliseconds, and token prices – i.e. billing by processed text units – are said by the provider to be very low.

What doing without text means technically

A generative model has a practically unlimited response space. It can name the desired category, paraphrase it, justify it, apologise, or invent a category that does not exist in the application at all. Development teams counter this with prompt rules, post-processing and validation – effort that arises anew in every project and never entirely disappears.

A pure classifier does not have this problem. It chooses strictly within the options specified for it. For the software, this means: the response always fits the expected format, without parsing pitfalls and without special handling for unusual outputs.

What matters here is the distinction the provider itself draws: this approach does not protect against errors in content. The model can decide on the wrong category, just like any other system. All that is guaranteed is that it stays within the permitted options. Formal reliability is therefore not the same as substantive correctness – this distinction should be clearly communicated in every project.

Where this pays off in practice

The benefit lies wherever a decision runs in the background and nobody is waiting for a formulated response:

  • Routing: automatically assign incoming forms, tickets or emails to the right department or queue.
  • Approvals: pre-sort editorial contributions, comments or uploads before a human reviews them.
  • Data checking: classify entries for plausibility, completeness or suspected duplicates.
  • Content enrichment: assign categories and keywords to articles, products or documents.

In all these cases, what counts above all is how quickly and how consistently the result comes back. Response times in the double-digit millisecond range change the architecture: a classification that takes 70 milliseconds can run synchronously in the request while the user is still waiting for the page. A call that takes several seconds has to be moved to a queue – with everything that entails: jobs, status displays, error handling, tracking.

Cost and testability

The second practical advantage is predictability. With very low prices per processed unit, a classification can be used even when it runs millions of times a year – for example on every page view, every import or every form submission. With generative models, this is often the point at which a proof of concept fails because of running costs.

But the most important thing for project work is testability. A classifier delivers discrete results, and discrete results can be checked with classic means: you create a test dataset with known correct assignments and measure how often the model gets it right. The result is a number that can be monitored over time and compared when models change.

With free text, this is far more complex. There you need either human evaluation or a second model as a checking instance – both expensive, slow and themselves prone to error. Anyone who has to demonstrate the quality of an AI feature to management has a much easier time with classifying systems.

The design question comes before the model question

The actual effort thus shifts to a different place: the categories have to be defined cleanly. Which options are there? Are they free of overlap? Is there a fallback option for unclear cases, and what happens then? Who decides when the model is uncertain?

At its core, this is classic process work and not an AI question. Experience shows, however, that it is what decides success. A poorly cut category tree delivers unusable results even with the best model – and a well-cut one often gets by with astonishingly simple technology.

Assessment

For web and software projects, this above all means a sober preliminary check: before a chat model is connected, it is worth asking whether the task is not in truth a decision between a few options. If it is, response time, cost and testability clearly favour a specialised model. Generative models keep their place where language is actually meant to emerge – for everything else, the smaller, narrower building block is usually the more stable one.

Sources