AI that only decides: less talk, more work done

Imagine an automated workflow receives an email and has to decide who to send it to. You ask an assistant and it replies: “Based on the content of the message, this appears to be an enquiry about a lease agreement, so I would recommend…”. The program was only expecting a label: “leases”.

That gap between what the machine can write and what the job needs is starting to produce tools of its own. Jev, ScaleDown’s models and Decisions API focus on responses that another program can use directly. An option. A score. A piece of data retrieved. The operation ends and the process moves on.

I have already written about how to size AI when it becomes part of a workflow. Here I am interested in the next step: changing the kind of response we ask for. A great deal of professional work produces a result that fits in a single box.

From talk to decisions

Large language models got us used to talking to AI. We ask a question, get an explanation and keep asking. That format is excellent for exploring a problem, drafting a document or discussing a strategy. An automated workflow often needs to resolve something much more specific: which category applies, which condition is met or which path to follow.

TypeSafe introduced Jev in September as a “System One” model. It receives information and questions with defined response types; it returns decisions accompanied by probabilities, without generating prose. The options form part of the task the model is given. TypeSafe.

ScaleDown is now taking that movement further with specialised models for classification, extraction and summarisation. With Classify, its constrained decoding model (CDM), the response can be just a label and its probability. Extract finds passages in the document. Summarize does write summaries: each tool has its own function. Integration with LangChain Deep Agents already allows them to be used in agent workflows. LangChain.

I see a third wave here: after getting used to conversational models and seeing offerings such as Jev, we are beginning to see a range of specialised tools that handle individual parts of the job. The useful result becomes a decision that fits into the process.

What it returns matters as much as what it understands

OpenAI’s announcement this Thursday reinforces that direction. Decisions API, based on GPT-6 Luna, is now generally available. It returns probabilities, options or scores. According to OpenAI, it can handle these tasks up to ten times faster than the same model through Responses API, with no charge for output tokens. Official documentation. We covered it in the news for 23 October.

I find that example revealing: the same model can serve a conversation or a narrowly defined decision, using a specialised interface and processing path. The change concerns how we use intelligence as well as the size of the model.

The promise is lower costs and shorter waits, with more predictable responses for the program receiving them. These are recent announcements and the figures come from the providers; a response restricted to a set of options can still choose the wrong one, and its estimated probability should help organise the review process.

For anyone who wants to understand the bill, two articles offer further reading: what token consumption really involves and why automation can be billed separately from a chat plan. The new development here is that a task can finish without producing a paragraph.

Three jobs that fit into a short answer

Let us return to the law firm’s email. We can define categories such as contracts, billing, new matters and other, and assess urgency separately. The result would feed into the work queues. “Other” acknowledges that a message needs another look. A lengthy explanation is no longer a compulsory step in routing the email.

Now think of a contract accompanied by a three-hundred-page PDF. We want to locate the signing date, the expiry date and the notice clause. An extractor can return the passages and their position in the document so they can be checked in context. That is how ScaleDown describes its Extract tool.

A long input can result in a minimal output. ScaleDown advertises a context window of up to one million tokens for these models. The task is to find what was requested within a large volume of material. Copying the notice clause provides the information the lawyer needs to assess the applicable deadline.

The third example is deciding the next step for an enquiry: request documents, arrange an appointment or refer it to a lawyer for review. The possible responses are set in advance. The system could handle that initial routing and leave the assessment of the matter to the person responsible for it.

In all three cases, a component sits between receiving information and acting on it. That component can be very small and still eliminate a great deal of repetitive work.

The question is still ours to define

Choosing the right closed question is a human task. “Is it urgent?” requires agreement on what urgent means in that law firm. “Does it need review?” requires clarity about which situations must always reach a person. The model receives those criteria; someone has to formulate them and check how they work with examples.

I would start with a recurring decision that currently requires someone to read something and then tick a box. I would look at whether the options describe the work well and whether the answer allows the next step to be taken. That is where a concrete improvement can be measured.

I believe the future of AI in professional work will look like a toolbox. The large model for thinking and drafting. The small, specialised model for deciding. And the human for setting the criteria, reviewing and taking responsibility for what matters.