Data platforms make decisions all day. An SQL query keeps the rows over a threshold, matches a keyword, and joins two tables on a shared ID. A pipeline sends a failed job down a retry path. Rules like these work when the answer sits in the data in a clean form.
They break when the answer depends on language or fuzzy judgment. Is this support ticket urgent? Do two donor records describe the same person? Does a job title belong to sales or engineering? Data teams have handled those cases with machine learning (ML) models trained for one task or a large language model (LLM) that answers in text that code has to parse.
Decision models, also known as System One models, provide another option. They make one specific judgment call, such as whether a ticket is urgent, and return a structured answer. Backers say the case for them is fit. Much of this work is small decisions repeated across millions of rows, and a paragraph of text or a model trained for each task is more than those decisions need.
The market for decision models has exploded in recent weeks following TypeSafe AI's debut of Jev on Sept. 15. It returns a choice, a score, or a yes or no, each with a confidence value, so the answer can sit in a column and drive the next step in a pipeline. Data platform vendors including Databricks, Snowflake, MotherDuck, and Astronomer each published support or a walkthrough of how to use a decision model approach. New model options from AWS, OpenAI, and Cloudflare have also arrived on the scene, expanding the number of decision model options.
Michael Leone, VP and principal analyst at Moor Insights & Strategy, said the format of the answer is what matters.
"A chat model hands you a paragraph you have to parse, while a decision model hands back a label and a probability that drops straight into a column you can filter, join, and report on like anything else in the warehouse," Leone told VentureBeat.
What a decision model is and where it fits in the data stack
A decision model answers a question with a typed value instead of text.
Given an input and a set of questions, a decision model like Jev returns a choice from a fixed list, a score on a scale, or a yes or no probability. TypeSafe's launch post says the possible outputs are defined in advance, so every answer fits the expected type, and each answer carries a confidence value.
Pavan Pothukuchi, head of AI platform at Snowflake, said the choice between SQL, a decision model, and a general-purpose LLM depends on what the answer looks like. A decision model works well when the answer comes from a known set of options, a rubric score, or a yes or no, and the job runs across thousands or millions of rows. Summaries, free-form extraction and real multistep reasoning still call for a general-purpose LLM.
"If plain SQL can settle it, like a value over a threshold or a keyword match, just use SQL," Pothukuchi told VentureBeat. "It's cheaper, and you know exactly what it's doing."
Pothukuchi said the best pipelines mix all three. SQL handles the deterministic filtering, the decision model makes the high-volume judgment calls and routes the work, and only the hard cases go to a larger model.
How data platforms are integrating decision models
In recent weeks, data platforms have added decision models in several ways.
Databricks took two routes. One option is using SemIf-OpenJev, an open-weight community project that reads option probabilities from an existing open model, on serverless GPUs and calls it from SQL through ai_query. Databricks also introduced ai_decide, a SQL function now in beta that is powered by a decision model that the company says is compatible with TypeSafe's API.
Snowflake Decision, now in private preview, is available through the AI_COMPLETE function in Cortex AI Functions. Pothukuchi said the model was built specifically for decisions and runs inside the Snowflake security perimeter under existing governance. Snowflake is also adding a lower-latency version called Snowflake Decision Fast and making decision models available through the Cortex REST API, Pothukuchi said, without giving a timeline.
Community extensions put Jev inside SQL engines directly. A DuckDB extension adds jev and jev_choice functions, and a Postgres extension does the same for PostgreSQL.
What data teams are building
It's still very early for the Jev-type decision models, although there are a number of use cases that are already emerging.
"We're seeing five major patterns in how customers want to use decision models: understanding user sentiment; classifying documents, emails or support tickets; routing AI agents; matching data across systems; and making operational and risk decisions," Pothukuchi said.
Classification and scoring
Astronomer, which sells the Astro managed service for Apache Airflow, said it runs a production job that sorts about 250,000 job titles by persona and function. The job uses Snowflake's AI_CLASSIFY function today. In a test of 4,000 titles, Astronomer reported that Jev finished in 66 seconds for 16 cents, against an estimated $11 to $21 for the same work on AI_CLASSIFY at list price.
MotherDuck, which sells a cloud service built on DuckDB, added a prompt_jev() SQL function to its paid plans. One example in its announcement pulls the main complaint from a table of support transcripts. On a company benchmark of 100,000 news articles from a public four-topic data set, MotherDuck reported that Jev labeled the rows in 40 seconds for 50 cents at 89% accuracy. An OpenAI model scored 88% in nearly 32 minutes at $37.58, according to MotherDuck.
Matching
Southbridge AI, a startup working on data and AI problems, used Jev to help resolve donors and committees across Ohio campaign finance records. In the startup's report, a pipeline that used Jev as the main workhorse and a second model to review uncertain cases cut model costs by 99.56% against a frontier model.
Pipeline control
Astronomer said Jev support ships in an upcoming version of the Airflow common AI provider package. An operator can branch a workflow on the model's answer and hold rows under a minimum confidence for human review. A separate retry policy asks the model what kind of failure occurred, so a permissions error fails at once while a network error retries after a short wait.
What to weigh before adopting a decision model
Whether a decision model is the right tool for a given data pipeline is a question that requires some inspection and consideration.
Stephanie Walter, practice leader for AI stack at HyperFRAME Research, noted that data platforms have supported ML models for classification, scoring, and prediction for years. Models like Jev let developers specify what to look for, such as a cancellation threat in a support ticket, without training a dedicated model for that task.
"That could make AI easier to apply across enterprise data, but the comparison should include traditional ML and rules, not just expensive general-purpose LLMs," Walter told VentureBeat. "Customers should choose based on accuracy, maintenance effort, and total cost."
It's not just about cost either. Walter cautioned that outperforming an unnecessarily large model on price does not automatically make a decision model the best tool for the job.
The other piece is the confidence score. Every Jev answer arrives with a probability, and teams use it to decide which rows to automate and which to send to a person.
"The probability that comes back with each answer is what people like most, because they can decide for themselves where to draw the line," Pothukuchi said.
That line still has to be tested, though.
"Each answer also comes with a score for how confident the model is, and those scores need checking against your own data before anyone acts on them, because a model that looks confident on a vendor's test data can be wrong a lot more often on yours," Leone said.
What this means for enterprises
For data professionals and AI teams, the rise of decision models has several implications.
"Where this goes is decision models becoming the traffic controllers inside AI pipelines and agent workflows," Pothukuchi said. "They'll figure out what each request needs and send it to the right place."
That raises the question of who is responsible for the decisions those models make.
"Nobody's really settled who owns those decisions yet," Leone said. "Someone has to define the answers the model is allowed to choose from, like which teams a ticket can go to or what counts as a suspicious claim, and that's a business call that often ends up buried somewhere in a data pipeline." Decision models could well be another great opportunity for data teams to exert influence.
"I expect the data platforms to compete on building that ownership and checking into the platform, the same way they did with data governance," Leone said.