Thinking Fast and Slow: AI Enabled Decisions

A new type of AI model has emerged that makes decisions instead of writing text. A little over two months ago, TypeSafe AI launched Jev, the first "System One” model. After Jev several others have followed, including open-source alternatives such as Laya and Strands Decider, and big tech players including OpenAI, Google and most recently Microsoft.  

These models take an input, for example a ticket, an email or a document, plus a number of questions, and return the answers in one pass, together with a calibrated confidence score. They do this at a significantly higher speed and lower cost, compared to asking a traditional LLM, even a smaller one. Depending on the model, some also handle other input formats, such as images or PDFs.

The benefits are lower cost, more reliability, since the answers are typed, and higher throughput. 

History of AI Enabled Decisions

Traditional machine learning has long been capable of handling classification problems, for example predicting if a customer is likely to churn or not, or an automated detection of defects on a manufacturing assembly line (and other variants of the "Cat vs. Dog" image classifier).

When generative AI and modern LLMs entered the scene, training and maintaining specific models for each task became too cumbersome compared to zero-shot prompting. For the past couple of years, we’ve therefore leaned towards using LLMs as the single, blunt instrument for every problem. You send a prompt to a generative model and hope that you can parse the reply cleanly enough to make the decision on what to do next: whether routing a support ticket, screening a document for personal data, or authorizing an ERP transaction. 

At proof-of-concept scale, this approach usually works. However, in production, it starts to break down: unpredictable latency, increased token costs, unreliable workflows with unexpected outcomes, and unnecessary data exposure are just some of the drawbacks.

System one models are therefore a welcome arrival that fill an important gap. Inspired by how the human brain makes fast, gut decisions, these models combine an LLM’s broad language understanding with instant decision-making. Instead of generating text token by token, they evaluate business rules in milliseconds and return a reply, together with a calibrated probability. This makes it possible to implement reliable and deterministic workflows around them, in a fast and cheap way. 

Under the hood, these models are trained via so called Reinforcement Learning for Calibrated Decisions (RLCD). Rather than rewarding the model for the most plausible-sounding text (which leads to rewarding it for sounding overly-confident), this approach rewards it for knowing when it is right. A 0.9 confidence score means the decision is correct 90% of the time (or at least very close to 90%), turning it into a predictable tool you can safely rely on at scale, across many use-cases.

Strategic Use Cases

System one models do not speak to your end users. Instead you should think of them as a fast and reliable decision layer connecting your business data, software, and other AI models. You hand the model a document or message along with a checklist of business questions, and it answers all questions simultaneously. The questions can be of three types: 

  • Yes / No Checks: Flags compliance rules, risks, or clear triggers, for example: "Does this document contain personal data?"

  • Multiple-Choice Routing: Sorts incoming requests into your predefined categories, showing how confident the system is in each choice.

  • Severity & Priority Ratings: Scores items along a clear scale (such as 1 to 5) to measure urgency, risk level, or quality.

A few of the use-cases where we see the instant value of these models include:

  • Compliance & security checks: Let a system one model verify your compliance, privacy, and security rules before sending the data on to an LLM or other tool. This protects the business and data, without making users wait.

  • Reliable guardrails for AI agents: Autonomous AI agents often break in production when left to figure out basic workflow steps on their own. System one models can be used as a guardrail, deciding when the agent should look up company data, wrap up a task, or hand off to a human.

  • Routing traffic to cut AI compute costs: Over 70% of incoming business requests don’t need an expensive frontier reasoning model. System one can handle routing so that routine tasks are handled automatically, moderate tasks are sent to smaller models, saving expensive AI compute strictly for the complex exceptions.

  • High-volume back-office processing: Rapidly sorting and scoring millions of invoices, insurance claims, or system alerts every day at a tiny fraction of the cost of using a frontier LLM.

A note on their reliability: Vendors frequently market System one models as "hallucination-free”. It is correct that they eliminate what we normally refer to as AI hallucinations, since they don’t generate free text. However, they do not completely eliminate classification errors. Safe deployment still requires implementing confidence thresholds and routing low-certainty cases to a more advanced model, or to human review. With some of the models, for example Laya, is also possible to fine-tune a system one model to further increase its accuracy.

How Decision Models Fit into the AI Landscape 

Adopting a system one decision model does not mean replacing all your LLMs; it means complementing LLMs and agents with a more deterministic model that lowers cost, increases speed, and makes the overall architecture more reliable. In a recent example, Amazon cut costs by 98% after replacing Claude Sonnet with the Strands Decider model in its internal ticketing system.

Picking the right decision model depends on your organization’s and use case’s specific requirements around privacy, speed, and convenience. If you handle sensitive data that must stay inside your own secure European cloud, or want to fine-tune the model to your business, a self-hosted option like Laya is the best choice. If you prefer a ready-made service that requires minimal set-up and maintenance, choose a managed option like Jev, or the vendor you’re already running.

Curious where a system one decision layer could cut compute costs or add deterministic guardrails to your current AI stack? Get in touch with us at Trice to discuss your architecture.

 

Next
Next

Dashboards are Dead: How AI-Native Data Changes the Executive Loop