Yes or no.
Is this a refund request?
Probability of yes92%
A binary question. A probability from 0 to 1.
← Back to LearnDecision
Decision
Models.
What is the right
next action?
A ticket to route. An invoice to check. An agent to guide. Discover models designed to choose between defined alternatives.
Generate an answer.
Choose an action.
An LLM generates text, code and reasoning. A Decision Model evaluates defined alternatives and returns probabilities or scores. Software uses them to move work forward.
Request + policies + alternativesDecision Model
Like a multiple-choice test and an essay: the output differs. An LLM can classify too; we choose the approach that suits the task.
A CUSTOMER REQUESTS A REFUNDSimulated data
Refund8%
Credit67%
Reject3%
Escalation22%
Confidence c56%
An allowed choice can still be wrong. The process checks rules and permissions before acting.
Three ways
to decide.
From language to software: structured answers, ready for the next step.
One choice.
Which team should handle this ticket?
Billing80%
Sales5%
Support15%
A distribution across alternatives: probabilities add up to 100%.
A score.
How urgent is action?
Expected urgency1.8 / 2
0 · 5%1 · 10%2 · 85%
An ordered scale. The score is the probability-weighted average.
Simulated outputs. These names describe three primitives; interfaces vary by model.
Probability
How plausible the model considers an answer. p = 0.80 means an estimated probability of 80%.
Confidence
How concentrated the output is on one answer. A clear choice has higher c; close alternatives, lower c.
c is not the probability of being right. Its formula depends on the model; Noul returns p directly.
Decide. Reason.
Act together.
The Decision Model evaluates the case. Orchestration selects a path using probabilities, confidence and the cost of an error.
DATA · CONTEXT · POLICIES · ALLOWED ACTIONS↓Decision ModelSystem 1 · probabilities and scores↓ORCHESTRATION · THRESHOLDS AND RISK
HIGH CONFIDENCE · ACCEPTABLE RISK
Execute
APIs and tools perform the authorised action.REASONING NEEDED · SYSTEM 2
Involve an LLM
Examines the case and proposes a next step, which returns to the checks.UNCERTAINTY OR HIGH IMPACT
Hand over to a person
An operator reviews, decides or approves. Even with high confidence.Outcome → monitoring → process improvement
The cost of an error matters: a €10 refund and a €100,000 payment need different checks.
Structured
workflows.
Break a process into focused questions. Models interpret, code applies rules and people handle exceptions.
The next action for the customer.
Explore the project01 · Context
Conversation, history and policies
02 · Decision Model
Intent · urgency · consent
03 · Rules
Support rules and completed transactions
A CONCRETE CASE
A customer reports a duplicate charge.
Refund is the leading option.
The workflow checks the transaction and amount limits before acting. The LLM prepares the message.
CHOICE · pSimulated data
Refund82%
Credit4%
Ask for information10%
Escalation4%
Confidence c76%
Probabilities across alternatives · total 100%.
Possible actions
- Handle the request
- Ask for information
- Involve an operator
An alert. The appropriate response.
Explore the project01 · Context
Alert, asset and authorised activity
02 · Decision Model
Expected activity? Sufficient evidence?
03 · Rules
Asset criticality and authorised playbook
A CONCRETE CASE
An unusual login affects a critical server.
The case moves to investigation.
The answers alone do not authorise containment: the playbook sets the required checks and approvals.
NOUL · pSimulated data
Authorised activity?8%
Alert explained by context?12%
Signs of compromise?89%
Three independent questions: each bar is p(yes). They do not add up to 100%.
Possible actions
- Close
- Investigate
- Contain or escalate
Each invoice follows the right path.
Explore the project01 · Context
Invoice, order, delivery and supplier
02 · Decision Model
Consistency · anomalies · missing documents
03 · Rules
Calculations, due dates and permissions in code
A CONCRETE CASE
A €12,400 invoice is missing proof of delivery.
The invoice is put on hold.
The system requests the missing document. Amounts, due dates and authorisations remain deterministic checks.
CHOICE · pSimulated data
Pay7%
Hold86%
Manual review7%
Confidence c79%
Probabilities across alternatives · total 100%.
Possible actions
- Pay or schedule
- Hold for information
- Send for review
Did the agent actually complete the task?
Explore the project01 · Context
Execution trace and feedback
02 · Decision Model
Permissions respected? Goal achieved?
03 · Rules
Compare task outcome and satisfaction
A CONCRETE CASE
The customer says thanks. But did the agent finish the job?
A possible silent failure.
The workflow checks the trace against the expected outcome and opens a review: satisfaction and success are different signals.
NOUL · pSimulated data
Task completed?14%
Customer satisfied?92%
Unauthorised action?1%
Three independent questions: each bar is p(yes). They do not add up to 100%.
Possible actions
- Record success
- Investigate the issue
- Trigger review
Four workflows, four dedicated slides in the guide. Values shown are simulated; c follows the Choice example and is not comparable across models.
Less waiting.
More work moving forward.
When every process contains thousands of microdecisions, latency and cost per decision become drivers of scale.
Simpler automation
Software receives choices and scores, without having to interpret a free-form answer.
Teams that scale
The workflow handles repetitive work. People focus on cases that require judgement.
Measurable decisions
Compare accuracy, time and cost per case. Then choose where to automate.
For one decision.
Median reported for Strands Decider v18 on NVIDIA RTX 3090. Varies with context length.
Provider benchmarkInput tokens. No output tokens.
Price published by Liquid AI on 5 October 2026. Each question is billed as a separate prompt.
Provider price and conditionsThese are two separate reference points, not a direct comparison. In your process we measure cost per correct decision, latency and quality: the benefit depends on the task.
Which model suits your workflow?
Explore metrics and comparisons in the guideA new space.
More models to choose from.
A selection of available models. Access, inputs and performance guide the choice.
TypeSafe AIJev
Three primitives for structured decisions.
- Input
- Text
- Access
- API
Liquid AId1
Decisions on text and visual content.
- Input
- Text and images
- Access
- API
CloudflareClef / Clef-flash
Variants to balance quality and speed.
- Input
- Text and images
- Access
- API · open weights
Strands AgentsStrands Decider 2B
A compact model that can run locally.
- Input
- Text
- Access
- Open weights
Logos belong to the respective providers. The guide covers architectures and criteria for evaluating models on your use case.
Go deeper
with the guide.
From primitives to real processes. Understand where Decision Models can make a difference.
- LLMs and Decision Models: how they work together
- Four workflows, explained step by step
- Models, performance and selection criteria
Get the free guide
Enter your details to access the PDF.
The guide is in Italian.
Which decision do you want to improve?
Start with your process. Together, we assess models, data, risks and expected results.
Or email us directlyinfo@icarex.ai