I want the answer while I’m still on the call.
Someone asks whether a customer agreed to a revised delivery date. You remember the exchange, but the answer could be in an email, a meeting note or the latest project document. The conversation moves on while you search.
That is the kind of everyday interruption I have in mind when I look at Jev. For a senior executive, the interesting question is what models like this could make possible inside the tools you already use.
What’s Next
TypeSafe introduced Jev on September 15. It is a System One model which are designed to make fast, structured business decisions. Examples such as routing a customer case, scoring a lead, or flagging a risky transaction, rather than generating prose one word at a time.
For a business, that can mean AI outputs that are faster, cheaper, easier to plug into workflows, and accompanied by a confidence score so low-confidence cases can be escalated to a person or a more deliberative LLM.
That smaller job creates room for speed and lower cost. TypeSafe reports responses of roughly 70 to 500 milliseconds for its decision model. Its published input price is $0.042 per million tokens, with no output-token charge. Something like Anthropic’s Opus with reasoning could take 5-15 seconds and be orders of magnitude more expensive. Tokens are the units used to meter text. Those figures describe the model call. Finding documents, checking access and preparing a useful answer still take work.
The applications arriving around Jev help show the direction. Jevmail uses it to sort Gmail messages into trays, including those needing a reply. Developers are also exploring independent alternatives. SemIf uses existing open models to produce this kind of decision. Jevlike provides a starting point for training a small model to choose among text options. GLiClass is an established open classification project whose repository dates to 2024. JevChat uses retrieved documents to assemble an answer and is the closest I’ve seen so far to almost instant guidance when you ask a question.
These projects take different approaches. Open application code around Jev still calls a hosted model. Independent models and tools have their own capabilities. Their existence shows people trying new ways to assemble AI systems. A company might use one model to interpret a request, a faster specialist to screen material, and another to explain the result. That could make frequent, small checks affordable enough to put throughout a workflow.
Consider a possible version of the call I described. You ask, “Did the customer accept the revised delivery date?” The application searches the company records you can access. It finds an older proposal, a recent email and the meeting note. A specialist checks which passages contain an actual acceptance and which only discuss the change. The application then shows a short answer beside the relevant passage, its sender and date. If the records conflict, it shows the conflict.
Jev could perform part of that checking. TypeSafe’s passage-review example separates retrieval, classification and answer writing. Its citation example checks whether a quoted source supports the claim. This is a possible design for a meeting tool, not a claim that Jev already searches your company’s systems.
A probability can help the application decide which results need another look. The percentage needs a clear meaning, though. Jev’s confidence measure describes certainty within a defined decision; it does not certify the truth of the final answer. Suppose a tool displayed 97% support for the statement about the delivery date. That would be an illustrative model estimate. To rely on it, you would want to know how often similar judgments held up on your company’s records, and see the message supporting this one.
The opportunity is to get that evidence quickly enough to use during the conversation. A high, tested support estimate and a clear source could give you more confidence about what to say. An uncertain result could tell you exactly what still needs checking.
I see this as an early sign of the connected AI services people have been describing. New products can combine several kinds of intelligence, with each handling a part of the work. You may never build with Jev. You could still benefit when the software your company buys uses a specialist to find relevant evidence, route a request or check an answer at a cost that permits regular use.
Useful now
Use AI to find weaknesses in your writing
Thomas Ptacek argues for using a model as a demanding copyeditor while retaining the words yourself. His method separates finding a problem from choosing the replacement. For an executive memo, that offers a useful division of work. The model can point out repetition, missing connections or a buried point; you decide how to express the judgment. The aim is a clearer memo that still sounds like its author.
Make company knowledge useful in daily work
Theory describes connecting agents to its team’s documents, data and applications. Its practical lessons include retrieving several items together, showing when a long task has started, and tracing how an answer was reached. For a team considering an internal research assistant, this gives you concrete behavior to request in a demonstration. A useful answer should arrive with enough context to understand where it came from.
Reduce the reading burden of AI-generated updates
Sean Goedecke describes a familiar imbalance. Someone can generate a long update in seconds while colleagues spend much longer reading it. His responses include asking for the salient points and moving an unclear exchange into a conversation. For an executive team, the practical question is whether AI saves time across the whole exchange. A fast draft that leaves the reader to discover the decision may transfer the work.
Thanks for reading What AI makes PossibNow! Subscribe for free to receive new posts and support my work.
What’s possible
The same model can behave differently when asked to choose
Goedecke built an experimental decision interface around an existing open model without retraining it. His demonstrations explore separating longer-term goals from fast choices and narrowing a large field of options in stages. For executives, the significance is that useful advances can come from how developers organize the work as well as from a new model release. His game demonstrations show an experiment, not a tested corporate assistant.
The software around a model affects the economics
Tomasz Tunguz argues that coordinating an agent’s work can determine whether a service is economical to provide. His business examples illustrate why two products using similar models could have different cost structures. For buyers, this suggests evaluating the cost of a completed, useful assignment. A model’s token price alone leaves out the calls, retries and review needed to finish the work.
AI changes how teams establish trust
Domen Kožar examines software teams adapting their contribution and review practices to AI. He argues for evidence of outcomes while noticing that new requirements can make participation easier for some people and harder for others. The wider management question is worth carrying into corporate workflows. As producing a first draft becomes easier, teams may need to reconsider how they review it and what evidence makes a handoff useful.
Expert corrections can become shared company knowledge
Meta describes a domain-specific agent that separates organizational knowledge from the procedures used to apply it. Expert corrections feed a process for proposing, testing and reviewing updates to those materials, without retraining the underlying model. The possibility for a large company is that a specialist’s correction can help the next person facing the same question. Meta’s account describes its own implementation, not a ready-made system available to every organization.
The loop this week
Choose one question that regularly interrupts a meeting because the answer is scattered across email and documents. Give your team the question and a few representative records. Ask them to show what an assistant would need to return for you to use its answer in the conversation.
The useful result is a short answer, the passage that supports it, and any conflicting or missing information. That gives you a concrete way to discuss what you want from the next AI tool your company evaluates.
Thanks for reading What AI makes PossibNow! Subscribe for free to receive new posts and support my work.
Get the next one on Friday.
Applied AI for professionals, by Sal Carranza. One workflow a week, free.