Decision models like Jev handle what SQL rules cannot. Here's how data teams use them

Jessica Okay so the pitch for Jev is basically: you've been using a sledgehammer to drive a finishing nail. Every small judgment call in your pipeline — is this ticket urgent, do these two donor records describe the same person — you've been routing that to a full LLM and getting back a paragraph you then have to parse.

Cathy Right, and the format of the answer is actually the whole argument here. A decision model hands back a label and a probability — that's it. It drops straight into a column. You can filter on it, join on it, report on it. It's just data.

Jessica Which sounds almost too simple, but I think that's the point.

Cathy It is. TypeSafe AI shipped Jev on September fifteenth, and the outputs are defined in advance — fixed list, rubric score, or yes/no, each with a confidence value. The model literally cannot return something that doesn't fit the expected type. For high-volume classification that's not a limitation, that's the feature.

Jessica And then within like three weeks, Databricks, Snowflake, MotherDuck, Astronomer — all of them drop integration posts. That's fast. That's 'we were already thinking about this and someone gave us a name for it' fast.

Cathy Yeah, and the implementations are interesting because they're not all the same. Databricks went two routes: ai_decide, which is now in beta and compatible with TypeSafe's API, and also a community path using SemIf-OpenJev — that's an open-weight project that reads option probabilities off an existing model and you call it from SQL through ai_query.

Jessica Oh interesting.

Cathy Snowflake's version is in private preview inside AI_COMPLETE in Cortex. They built their own model specifically for decisions, it runs inside their security perimeter, and they're working on a faster variant. DuckDB and Postgres both have community extensions that just add jev functions directly to SQL — which, honestly, is the most data-engineer thing imaginable.

Jessica Of course there's a DuckDB extension. Of course. I love that ecosystem.

Cathy Look, they know their audience.

Jessica Okay but let's talk about the numbers because the benchmarks in the article are the part that actually made me sit up. Astronomer runs a production job sorting about 250,000 job titles by persona and function. They tested Jev on 4,000 of those titles — 66 seconds, sixteen cents. The same work on Snowflake's AI_CLASSIFY at list price: estimated eleven to twenty-one dollars.

Cathy That's a real cost gap. And MotherDuck's number is in the same territory — 100,000 news articles, 40 seconds, fifty cents, eighty-nine percent accuracy. An OpenAI model hit eighty-eight percent but took nearly 32 minutes and cost $37.58.

Jessica One percent accuracy difference.

Cathy Right. For most classification jobs that's noise. Now — vendor benchmarks, their own data set, four-topic news corpus, all the usual caveats apply. But the order-of-magnitude cost story isn't coming from one number, it's showing up across multiple integrations with different platforms. That's harder to dismiss.

Jessica The Southbridge AI case is the one I find most compelling, actually. They used Jev to match donors and committees across Ohio campaign finance records — which is a genuinely messy entity resolution problem — and they cut model costs by ninety-nine point five six percent against a frontier model by using Jev as the main workhorse and a second model only for the uncertain cases.

Cathy That routing pattern is the smart part. SQL handles deterministic filtering, Jev handles high-volume judgment calls, hard cases escalate to the big model. You're not replacing the frontier model, you're just not sending everything to it. Which — we've said variations of this before, right? The recipe is: use the right-sized tool, not the biggest available one.

Jessica Every time. And yet here we are in twenty twenty-six and data teams are still routing is-this-ticket-urgent to a model that can write a sonnet.

Cathy Okay, fair. So the confidence score — that's the part I want to sit on a second because it's both the most useful thing about this format AND the thing I'd be most careful about. Every Jev answer comes with a probability. Teams use that to decide which rows automate and which go to a human. Astronomer's Airflow integration will literally branch a workflow on the model's answer and hold low-confidence rows for review.

Jessica Which is exactly the kind of human-in-the-loop design we keep saying is the right move. The confidence score makes the handoff point explicit instead of invisible.

Cathy In principle. But Michael Leone's quote in the article is worth repeating — a model that looks confident on a vendor's test set can be wrong a lot more often on YOUR data. The threshold has to be validated against your own distribution before anyone acts on it. That's not a knock on Jev specifically, it's just… calibration is hard and confidence scores are famously overconfident.

Jessica Yeah, I mean — this is the same thing we kept saying about retrieval confidence in RAG setups. The number is a signal, not a guarantee.

Cathy Mm-hm.

Jessica And then there's the governance question that Leone raises, which nobody seems to have figured out yet. Someone has to define the answer list — which teams a ticket can go to, what counts as a suspicious claim. That's a business decision, and right now it's getting buried inside a data pipeline where nobody's looking at it.

Cathy This is the part that actually interests me more than the cost story. Because Leone's prediction is that data platforms compete next on building that ownership and audit layer — the same way they competed on data governance. Which, if you squint, is the open question we've had sitting around since the agentic orchestration report: who controls the policy layer.

Jessica Round… what are we on now, five of that same fight?

Cathy At least. And honestly I don't have a strong read on whether Snowflake or Databricks wins that governance race — my recent track record on platform calls is not something I'm going to brag about.

Jessica Stop it, you called the Cortex expansion.

Cathy I called the direction, I was wrong about the timing. Different thing. Anyway — I'll say this: the actual bet I'd make is that ai_decide hits general availability on Databricks before the end of November. It's in public beta now, the integration story is already there, and Databricks does NOT let things sit in beta quietly. Sixty-five, thirty-five.

Jessica I'll take that. And I think the more interesting near-term question is whether Snowflake Decision stays in private preview through the end of the year or gets pushed to public. They mentioned a faster variant without giving a timeline — that usually means Q4.

Cathy Yeah, 'no timeline' in an October launch post usually means 'ask us in January.'

Jessica The analyst line I keep coming back to is Pothukuchi's framing — decision models as traffic controllers inside AI pipelines. That's not just a use case, that's a product position. If Snowflake or Databricks can make this the default routing layer, that's a real foothold.

Cathy And it connects to something Stephanie Walter flags that I think is undersold in the article — the comparison shouldn't just be Jev versus expensive LLMs. It should include traditional ML classifiers and rules. A fine-tuned BERT-class model does classification too. The real advantage Jev has is that you don't need labeled training data upfront — you specify what to look for in natural language.

Jessica That's the adoption unlock, right there. You don't need to build and maintain a dedicated model for each task. You describe the judgment call, you get a typed answer. For teams that are already drowning in SQL and Airflow DAGs, that's actually a meaningful change.

Cathy Agreed. I came into this a little skeptical — 'System One models' felt like it might be harness-cubed energy, a new name for a thing we already had. But the output format constraint is genuinely the differentiator. It's not a chat model with instructions to be brief. It's a different contract.

Jessica Okay, Cathy finding something to be sold on — that's how you know we're done. See you next time.