Incident sheet · /enterprise-ai
- System
- Enterprise AI
- Filed
- Read
- 4 min · 830 words
- Severity
- Routine
Decision Models Land in the Enterprise AI Stack at 4 Cents Per Million Tokens
TypeSafe's Jev popularized decision models in September 2026; by November, OpenAI, Google, Microsoft, Snowflake and AWS had all launched rivals at fractions of a cent per decision.

Event log
- TypeSafe's Jev costs about 4 cents per million input tokens with no output-token charge
- Jev launched September 2026; Google's Gemini Decisions API followed in November
- Decision models respond in under a second and return probabilities for every allowed option
- At 1 million decisions per day, a 1% error rate means 10,000 mistakes
- The EU AI Act requires human oversight of certain high-risk AI decisions
A judgment call that used to cost a large language model several cents now costs 0.000004 dollars. TypeSafe's Jev, a hosted decision model priced at about 4 cents per million input tokens with free output tokens, is pulling an entire category into the enterprise AI stack — and by November 2026, OpenAI, Google, Microsoft, Databricks, Snowflake, Cloudflare and AWS had all shipped rivals.
Start with the scene where these models actually earn their keep: a supplier email about a delayed shipment lands in a procurement inbox. A decision model reads it and answers three narrow questions — does it specify a firm new delivery date, does the delay threaten a customer commitment, which escalation category applies. Conventional software then compares dates and stock levels, and procurement rules decide whether to expedite or find another supplier. Everything outside the judgment stays in code that teams can read, test and modify.
What exactly is a decision model?
It's a smart if-statement. You feed it input data — an email, customer record, document, sometimes an image or call recording — plus questions with predefined answer options. It returns an answer and a probability for every allowed option, in under a second. It generates no text.
That matters because most AI in business processes is small judgment, not content generation: is this ticket about billing or a defect, does this invoice match the purchase order, should the agent send that email, which model should handle this prompt. Until now the default was querying an LLM and parsing a one-word answer — slow, costly at scale, and silent on confidence.
Terminology varies. TypeSafe calls Jev a "system one model," after economist Daniel Kahneman's term for fast, intuitive thinking. Others say decision model or small judgment model (SJM).
Isn't this just an ML classifier?
In one sense, yes — spam filters and fraud models have classified inputs for decades. The difference is cost and ease. Traditional classifiers demand thousands of labeled examples, a data science team, a separate model per question and ongoing maintenance. Decision models take plain-language questions, accept documents in various formats, and one model serves scenarios from ticket routing to contract review. Jev's request format has become a de facto standard: users can often switch providers by changing an endpoint, not rewriting an application.
A trained classifier still wins when a question is stable and labeled data is plentiful. Decision models shine when questions change frequently or no labeled data exists.
Who's selling what?
The landscape splits along two axes.
By capability:
- General-purpose text and records: TypeSafe Jev, OpenAI Decisions API, Microsoft-Decision-1
- Multimodal (images, audio, video): Google Gemini Decisions API, Cloudflare Clef-omni, Liquid AI d1
- Longer reasoning: Fastino GLiDE, Respan span
- Source-consulting with citations: Parallel Decisions API, Perplexity Decider
- Industry-specific: Superagent Security-One, Nace AI Drex Max Finance
By delivery: major AI platforms (OpenAI, Google, Microsoft); data platforms reachable through SQL, such as Databricks ai_decide and Snowflake Decision; open-weight self-hosted models like Cloudflare Clef and AWS Strands Decider running via Ollama, vLLM or Hugging Face; and specialist APIs, some still on early-access terms with uneven enterprise support.
The timeline compressed fast: TypeSafe launched Jev in September 2026, OpenAI, Databricks, Cloudflare and AWS followed quickly, Microsoft and Snowflake announced in October, and Google shipped the Gemini Decisions API in November.
Where do they pay off?
Use cases span nearly every function — routing tickets and flagging at-risk customers, matching invoices to purchase orders, holding payments when a customer appears coached by a scammer, fast-tracking straightforward insurance claims, scoring security alerts, routing prompts to the cheapest capable model, checking agent actions against permissions before execution, and identifying claim denials worth appealing. A growing pattern: a fraction-of-a-cent check before each AI agent action, reinforcing spending limits and approval rules. Look for decisions that are frequent, narrow, data-heavy, governed and time-sensitive.
What are the risks?
The uncomfortable arithmetic: a well-calibrated model at 90% confidence is still wrong one time in ten. At 1 million decisions a day, a 1% error rate means 10,000 mistakes. A structured answer isn't necessarily a correct one — structure makes errors easier to detect, not to eliminate.
Other traps:
- Calibration varies by domain; test against your own labeled cases before setting thresholds
- The model sees only the inputs provided — missing records produce confident wrong answers
- High-stakes calls in credit, employment, housing, insurance, education, healthcare or legal rights need human oversight; the EU AI Act requires override capability for certain high-risk systems
Practical next steps are unglamorous but proven: start with existing data-platform contracts, integrate through a common interface to avoid lock-in, demand vendor calibration results, and treat thresholds as versioned policy — checking whether 80%-confidence predictions actually prove right 80% of the time.
A layer that answers narrow questions for fractions of a cent doesn't replace agents or LLMs — it decides when they should be trusted.
via techtarget.com (Source)
More from Olivia Hart
Show full bio
News editor covering money, funding and market shifts around enterprise information technology and the leaders who run at CIO Monitor.
215 articles
The newsroom digest
One email a day with the stories that matter. No noise.
Cross-references
- Microsoft Ships Decision-1, a Cheaper AI Model Built for Verdicts, Not Prose
- Microsoft Ships Decision-Decision Model as Category Hits 25% of Router Traffic
- Jev and the Arrival of Model Tetris: Decision Models Get a Standard API
- Agentic AI pricing splits three ways as buyers gain leverage
- Enterprise AI Bills Set to Outpace Falling Model Costs