For years, most AI automation has treated workflows as coarse blocks. You hand a process chunk to an LLM or agent, wait for text or structured output, parse it, and decide what happens next. That works for occasional, high-value tasks. It scales poorly for the thousands of small judgments that actually move work through an organization.
System 1 models (also called decision models) flip the granularity. They return typed answers with calibrated probabilities—choices among options you define, scores on a scale, or yes/no probabilities—rather than generated text. The first widely discussed example, TypeSafe AI’s Jev, arrived in mid-September 2026; others followed quickly. Latency is typically tens to a few hundred milliseconds. Input is priced around $0.042 per million tokens, with output effectively free because there is none. On comparable classification-style tasks they have been measured at large multiples faster and cheaper than frontier LLMs, while matching accuracy on many of the narrow judgments software actually needs.
The name comes from Kahneman: System 1 is the fast, associative mode that recognizes a pattern or makes a snap judgment; System 2 is the slow, deliberate mode that reasons step by step. System 1 models are built for the former. They do not write replies, produce code, or explain themselves. You give them state plus typed questions; they return the selected option (or probability distribution) and a confidence signal. Your code sets the threshold, applies policy, escalates, or continues. They are closer to a high-quality, probabilistic if-statement than to an autonomous agent.
From Process Blocks to Micro-Decision Pipelines
Once decisions of this shape are cheap and fast, the natural unit of automation shrinks. A support workflow is no longer “run an LLM on the ticket.” It becomes a sequence of micro-decisions: which queue, how urgent, does it match a known exception pattern, is the required evidence present, should a human review before any action is taken, which specialized skill or model should handle the next step. Each is a single call that returns numbers your control flow can branch on immediately.
The same pattern appears in underwriting, claims triage, content moderation, agent tool selection, retrieval gating, and many operational loops. Instead of one expensive generative step that tries to do everything, you get a pipeline of cheap, inspectable gates. High-confidence paths proceed deterministically. Low-confidence or high-stakes paths escalate to a slower reasoning model, a human, or a rules engine. The architecture is dual-process by design: System 1 for volume and speed, System 2 only where the economics and risk justify it.
This is not merely technical decomposition. It changes what “automation” means. You stop asking “can an agent complete this process end-to-end?” and start asking “which micro-decisions can be made reliably and cheaply enough that the residual human or expensive-model work becomes smaller and more valuable?”
The Economics Shift
The cost and latency difference is large enough to alter what is worth automating. A single generative call that costs several cents and takes seconds is economical only for infrequent or high-value work. A decision call measured in fractions of a cent and a few hundred milliseconds can sit in front of every ticket, every document, every tool invocation, every routing choice. At that price, the fixed cost of instrumentation, threshold tuning, and monitoring can be amortized across high volumes.
Several consequences follow:
Volume becomes viable. Processes that were previously “too small to automate” now clear the economic hurdle because the marginal cost per decision is low.
Cascades improve the frontier. A fast System 1 stage can handle the bulk of cases and escalate only the uncertain fraction to a more expensive model. Independent benchmarks have shown this can recover much of the accuracy of the larger model at a cost closer to the cheap layer.
ROI calculations change character. You stop measuring “hours saved on a project” and start measuring decision quality × volume × cost delta. Small improvements in routing accuracy, exception detection, or false-positive rates compound across thousands or millions of daily decisions. The economic unit shifts from the workflow to the micro-decision.
Observability and control become cheaper to implement. Because every gate returns a probability and confidence, you can log, threshold, audit, and A/B test decisions without parsing free text. Escalation rates, calibration drift, and error concentration become measurable operational metrics rather than qualitative impressions.
None of this eliminates the need for System 2. Complex multi-hop reasoning, novel problem solving, generation, and high-stakes policy judgment still belong to slower models or humans. The point is that the expensive layer is no longer the default for every step.
What This Requires in Practice
Decomposing a workflow into micro-decisions is itself work. You must identify which judgments are stable enough, high-volume enough, and low-stakes enough (or gateable enough) to hand to a System 1 model. You need labeled outcomes or at least consistent human judgments to tune thresholds and monitor calibration. You keep the control flow, the business rules, and the side effects in ordinary code; the model only supplies the probabilistic judgment about the state you assembled.
There are also failure modes. Models can read the facts correctly and still choose the wrong action. Calibration is statistical, not a guarantee on any single prediction. Some decisions require combining facts in ways these models do not reliably perform in one pass; those cases often improve when you ask narrower factual questions and combine the answers in code. High-stakes actions still need explicit human or rules gates regardless of model confidence.
For organizations that already run decision management or rules engines, System 1 models fit as a perception layer: the model answers “what is this?” while policy answers “what do we do about it?” For teams building agents, they act as cheap, typed gates before tool use or generation.
Implications for Services and Operations
The practical result is that AI automation moves from sporadic high-intensity interventions toward continuous decision infrastructure. The economic leverage comes less from replacing large blocks of work and more from making every small branching point faster, cheaper, and more consistent. Service organizations, operations teams, and vertical process owners can now productize decision layers—routing logic, exception detection, prioritization, eligibility checks—that were previously either fully manual or too expensive to instrument with generative models.
The shift is recent and the models are still early. Accuracy, calibration, and multimodal coverage will continue to improve, and more open implementations are already appearing. The conceptual change, however, is available now: once micro-decisions are economically cheap, workflows stop looking like pipelines of coarse steps and start looking like dense graphs of small, measurable judgments. The organizations that redesign around that grain will capture the volume advantage; those that keep treating every decision as a generative task will continue to pay System 2 prices for System 1 work.