The $0.0001 Decision

Healthcare organizations are overusing expensive reasoning models. Highly sophisticated AI models are being used to perform remarkably routine tasks such as classifying a document, routing a worklist item, or identifying a gap in submitted information.

These small decisions consume a great deal of time throughout payer operations. Some can be handled with explicit rules, while others may require interpreting messy documentation.

Lower-cost AI built for bounded decisions could bring continuous judgment into healthcare workflows without relying on expensive reasoning models.

System 1 and System 2

Daniel Kahneman popularized the concepts of System 1 and System 2 thinking in his book Thinking, Fast and Slow. System 1 is fast and intuitive — recognizing a face, for instance, or understanding a familiar phrase without consciously working through a series of steps. System 2 is deliberate — comparing evidence, performing calculations, and reasoning through unfamiliar problems.

The analogy maps surprisingly well to AI. Give a model like Anthropic’s Opus a difficult problem, and it can analyze evidence, plan, use tools, reconsider its approach, and work through multiple steps before reaching a conclusion.

In contrast, TypeSafe AI recently introduced Jev, which it describes as a “System One” model optimized for fast, probabilistic decisions among predefined options.

A model evaluating which specialty queue should receive a work item is making a focused judgment. A reasoning agent investigating missing evidence across several systems is working through a more complicated problem.

Both capabilities have a place in healthcare software. The main challenge is to match the level of intelligence to the task, using the simplest approach capable of handling each task reliably.

System 1 and the Rise of Bounded Judgment

An application using a judgment model (a.k.a., a decision model) supplies the information to evaluate, the questions to answer, and a set of permitted answers. In turn, the model produces typed results and probability estimates rather than generating an explanation.

For a prior authorization, the input might include procedure details and submitted clinical documentation. The application could ask if specialist review appears necessary and which specialty should receive the case. An illustrative response might look like this:

A conventional large language model (LLM) can also return structured answers. A model optimized for bounded decisions, however, is designed to deliver them quickly and at a lower cost while providing probability estimates that software can use.

The application then applies its rules and thresholds to determine next steps. It might route the case, request additional information, or send an uncertain result for review. The permitted actions remain defined in software.

Healthcare Runs on Micro-Judgments

Spend time inside payer operations, and you’ll find small judgments everywhere. Claims teams assess unusual activity and potential duplicates, care managers identify members who may need outreach, and appeals teams determine whether evidence is relevant or missing.

Traditional software has difficulties with decisions that require interpretation rather than explicit rules, leaving experienced people to make many of these judgments. As a result, we’ve built conventional systems, such as queues and dashboards, around them.

Cheap probabilistic judgment models could evaluate these questions continuously and flag cases that need human attention. Instead of searching queues for work that requires their expertise, people could spend more time addressing the cases that warrant it.

When the Economics Change, the Architecture Changes

Jev’s published pricing is based on input tokens, so the effective cost per judgment depends on how much information an application supplies and how many questions it evaluates together. The economics example below uses an illustrative unit cost to explore what inexpensive judgment could enable in healthcare workflows. Actual inference costs would need to be measured against the intended workload.

For the past few years, software architects have reserved model calls for decisions important enough to justify their cost. Lower-cost judgment models change that calculation, making it practical to evaluate cases more frequently as new information becomes available.

This creates a new architectural possibility: a continuous judgment layer between deterministic software and more expensive reasoning models. Alongside conventional rules, this layer could evaluate the meaning of new information as a payer case changes and determine whether further review is needed. The application could then escalate selected cases for further investigation, reserving more expensive reasoning for situations that warrant it.

Put a Judgment Layer Around Utilization Management

Consider a prior authorization for an advanced imaging procedure entering a payer’s utilization management (UM) platform. The case contains member and provider information, procedure details, clinical documentation, benefit data, timestamps, and workflow history.

The application first performs deterministic checks verifying eligibility and benefits and calculating the review deadline. A judgment model then evaluates bounded questions about the submitted material. The following values are hypothetical:

One question is deliberately absent: Should we approve the authorization? Medical-necessity determinations carry substantially greater risk than administrative triage. A more appropriate use of bounded judgment is supporting the workflow around that decision while leaving the determination itself to qualified reviewers. These tasks still require safeguards because administrative errors can delay care.

In this example, the estimated probability of missing evidence crosses a hypothetical investigation threshold of 0.80. A System 2 reasoning agent then investigates the missing evidence using the systems or APIs it is authorized to access.

If the agent finds relevant evidence, verifies that it belongs to the case, and adds it, the case state changes. The judgment model evaluates the updated information. The estimated probability that required documentation is complete rises to 0.98, while the probability of missing evidence falls to 0.06. The probability that clinical review is required remains high at 0.95.

A nurse receives the case in the radiology queue with the available evidence assembled. Less time is spent locating documents and resolving administrative gaps, leaving more attention for clinical interpretation and decisions within the nurse’s professional authority.

The $0.0001 Decision

Now consider the economics. Imagine a large payer processing 10 million UM case events per year, with each event requiring an average of 20 bounded judgments. At an illustrative effective cost of $0.0001 per judgment, 200 million judgments would result in an annual model-inference cost of $20,000.

That doesn’t mean a payer can operate UM for $20,000. A production system still requires data integration, infrastructure, security, monitoring, model evaluation, and human input. This calculation isolates one cost: model inference.

At this price point, inference itself may no longer be the limiting factor. Lower-cost judgment models can make it practical to evaluate cases throughout a workflow while reserving more expensive reasoning models for situations that require deeper analysis. Rather than waiting in a queue for someone to inspect it, a case can be evaluated as its state changes, utilizing human decisions when it is needed.

Confidence Becomes Part of the Architecture

In addition to accuracy of AI models, organizations must also establish an acceptable confidence threshold for each action based on the consequences of being wrong. A routing error may be recoverable, for example, but it can still delay review and affect access to care.

The system should therefore respond differently based on both the model’s confidence and the consequences of the action. A strong indication of missing evidence might trigger an investigation, while uncertainty about routing could send a case for review.

High-confidence, low-consequence decisions may be appropriate for automation, while greater uncertainty may warrant System 2 reasoning or human review. High-consequence decisions may require human input regardless of model reliability.

This approach to selective automation provides a more practical way to apply AI models in healthcare while preserving human input where it’s needed.

Keep the Rules Engine

Judgment models should supplement rather than replace deterministic healthcare logic. Eligibility dates and regulatory deadlines can be calculated, required fields can be validated, and established rules can be evaluated without probabilistic AI.

When deterministic logic can reliably provide the right answer, introducing a probabilistic model adds uncertainty without providing additional value.

A mature architecture should instead match each task to the appropriate approach. Deterministic software can handle known facts and explicit rules, while decision models can make bounded probabilistic judgments and System 2 models can address problems that require investigation. Human involvement remains essential for consequential decisions and situations where uncertainty requires accountable judgment.

The priority is determining which layer is appropriate for each task.

Cheap Decisions Can Still Be Expensive Mistakes

A $0.0001 inference can create a very expensive problem. Judgment systems need guardrails just as generative ones do.

Healthcare organizations must understand confidence score reliability, how model performance changes over time, and whether decisions can be traced and reproduced. They must also define which outputs can trigger automated actions and where human input is required, particularly when operational decisions begin to approach clinical decision-making. These considerations are fundamental to whether a system is safe and useful.

One advantage of bounded judgment models is that their decision space can be explicitly constrained. The application defines the questions a model can evaluate, the answers it can return, and how those outputs may affect the workflow. This creates clearer, more governable boundaries than giving an autonomous agent broad authority to determine what actions it can take.

What Happens When Judgment Is No Longer Scarce?

Cheap bounded judgment models could allow healthcare software to evaluate each meaningful case change, surface missing information, and flag cases that need deeper review before a person opens a work item.

Human attention would remain a limited resource, but less of it could be spent assembling information and sorting through routine work, leaving more time for decisions that require human expertise.

For organizations looking to apply this approach, a practical starting point is a single queue. Identify the routine judgments people make before the substantive work begins, then determine which could be handled reliably by a bounded judgment model.

The opportunity lies in matching each decision to the level of intelligence it actually requires.