Fast, intuitive and calibrated judgment
- ObjectiveInstant assessment and classification
- OutputConstrained and typed: booleans, enums, scores
- Latency10 – 50 ms
- NatureDeterministic, probability-calibrated
- ComputeMinimal: small SLMs and classifiers
Design · The future of AI
Scroll to explore
S1·S2 — The evolution of AI decision models: integrating fast judgment (System One) and slow reasoning (System Two)
Fast when it's enough, deep when it matters.
Inspired by Kahneman's two cognitive systems, we separate fast, calibrated judgment from slow, deliberate reasoning: the first resolves almost everything in milliseconds, the second steps in only for complex cases.
The principle
An architectural misconception: an LLM for every task.
For years, enterprise software engineering has deployed general-purpose generative LLMs for every task, including plain classification, triage and binary decisions.
Asking a 70-billion or 400-billion parameter model to generate free text just to evaluate whether a transaction is fraudulent or whether a ticket belongs to billing is the engineering equivalent of using a jet airliner to travel 100 meters.
Separate and orchestrate fast judgment and slow reasoning.
01 · Cognitive architecture
From the human mind to software engineering
The brain does not process every stimulus with deep analytical focus. On a familiar road or facing a known face we rely on System 1: immediate, energy-efficient, pattern-driven. For an equation or a chess move we engage System 2: slow, sequential, demanding.
Slow, analytical and deliberate reasoning
- ObjectiveComplex problems and multi-step planning
- OutputGenerative: Chain-of-Thought, reasoning graphs
- Latency2 – 60 s
- NatureAutoregressive, test-time compute
- ComputeHigh: frontier LLMs and reasoning models
p The mathematics of probability calibration
The problem with traditional generative LLMs is overconfidence: they can output a completely fabricated statement (hallucination) with apparent absolute certainty. System One models are trained or post-calibrated with dedicated techniques (for example reinforcement learning with calibration rewards, or temperature scaling) to align stated confidence with true empirical accuracy:
ℙ( Ŷ = Y | P̂ = p ) = p
If the model states p = 0.92, then across a large sample of cases with that score about 92 in 100 are correct. That makes confidence a reliable threshold on which software can act.
02 · Orchestration
The decision routing flow
In production, System One acts as a Decision Gateway and Security Router: it filters, classifies and resolves the vast majority of the workload, escalating to System Two only when genuine complexity or ambiguity is detected.
Incoming requestpayload · transaction · prompt
AI Firewall / Proxysanitization & PII masking
System Oneclassification + confidence
High confidence
score ≥ threshold
Fast path
- Dynamic forms generated
- Direct backend API call
- Automated execution
Low confidence
score < threshold
System Two
- Chain-of-Thought (CoT)
- Graph / tree search
- Deep audit & synthesis
response or action
A decision rule, not an opinion
azione = { execute se p ≥ τ ; escalate se p < τ }
τ = 85%System Two / humanFast path
With calibrated probabilities the threshold τ is a measurable business lever: the higher it is, the fewer automatic errors and the more cases go to slow analysis.
Execution lifecycle
- Ingress and sanitization the AI Firewall scrubs sensitive PII and blocks Prompt Injection vectors.
- System One evaluation a single forward tensor pass returns intent, action and confidence.
- Fast path above the threshold (e.g. ≥ 85%) the action runs directly in under 50 ms.
- Slow path below the threshold or for complex edge cases: System Two multi-step reasoning or a human supervisor.
03 · Comparative matrix
Three model families compared
Typical latency (log scale)
System One
10-50 ms
Generative LLM
1-3 s
System Two
5-60 s
5 ms50 ms0,5 s5 s50 s
Cost per request (log scale)
System One
$0,0001
Generative LLM
$0,005
System Two
$0,03-0,15
$0,0001$0,001$0,01$0,1
| System One | Standard generative LLMs (GPT-4o, Llama 3) | System Two (o1, o3, Reasoning) | |
|---|---|---|---|
| Primary objective | Instant decision, classification, routing | Free-form text generation and synthesis | Complex problem solving and logical planning |
| Output type | Strict data types (Enums, Bools, JSON) | Unstructured text, Markdown | Chain-of-Thought + final answer |
| Typical latency | 10 – 50 ms | 1,0 – 3,0 s | 5,0 – 60,0 s |
| Confidence calibration | High, to be verified on real data (ECE) | Uncalibrated / absent | Embedded in the search process |
| Hallucination risk | Low (constrained output space, but not zero) | Medium-high | Low on logic, medium on raw facts |
| Compute cost | Extremely low (~$0.0001 / req) | Medium (~$0.005 / req) | High ($0.03 – $0.15 / req) |
| Hardware target | Local edge, CPU or light SLM GPU | Data-center GPUs | High-VRAM multi-GPU clusters |
Orders of magnitude at full load. In the case study below, a €350/month local instance over 900,000 requests costs €0.00039 per request: real cost depends on volume.
04 · Enterprise use cases
Where fast judgment makes the difference
Real-time fraud detection
At high payment volumes per second, sub-second windows are required: a generative LLM is non-viable on latency and cost. System One inspects the transaction, evaluates behavioral graphs and outputs APPROVE / FLAG_FOR_REVIEW within tens of milliseconds with calibrated confidence.
AI guardrails and ingress firewalls
Before a query reaches downstream RAG or corporate databases, System One evaluates intent: bypass attempts, policy violations, out-of-domain requests. Threats are blocked at the perimeter.
Intelligent customer-service triage
Instead of a verbose dialogue, System One identifies the business process and renders a dynamic form for the missing parameters, routing the case to the right backend without wasting tokens.
05 · Economic sizing
Cost and financial sustainability
An enterprise processes 1,000,000 operational workflows per month for document classification and administrative routing.
Monthly technology cost comparison (€/month)
Infrastructure cost reduction: −93.5%
Everything on a cloud LLM
cmono = 4.000 × $2.50 / 1M + 500 × $10.00 / 1M = $0.015
1,000,000 × $0.015 = $15,000 / month (≈ €13,800)
Average perceived latency: 2,5 s
System One + escalation
90% (900,000) → System One local: €350 / month
c2 = 1.000 × $2.50 / 1M + 350 × $10.00 / 1M = $0.006
10% (100,000) × $0.006 = $600 ≈ €552
Chybrid = €350 + €552 = €902 (≈ $980)
Average latency: 0,9 × 30 ms + 0,1 × 2,5 s = 277 ms · for 90% of users: 30 ms
Cost as the share resolved by System One varies
Chybrid = F + N × (1 − a) × c2
F = local instance (€350 ≈ $380). Each extra percentage point resolved by System One removes $60 per month of escalation: router quality is worth money.
−93.5%on infrastructure costs
€12,898saved per month
€154,776saved per year
~277 msaverage latency, from 2.5 s
Assumptions: $1 = €0.92; GPT-4o list price ($2.50 / $10 per 1M tokens); monolith with 4,000 tokens in and 500 out per request, escalation with 1,000 in and 350 out thanks to the context trimmed by System One; estimated local instance cost. To be replaced with the client's actual data.
DigitalSolutions
The future is orchestrating the two systems.
Fast, calibrated judgment for almost everything, deep reasoning only where it matters: faster, cheaper, more reliable.