The strongest AI system is rarely the one that sends every request to the biggest model. It is the one that gives each step the least expensive tool capable of completing it safely—and makes escalation explicit when that tool is not enough.
Start with the output contract
“Which model should we use?” is too early. First decide what the software needs back. An exact calculation belongs in code. A bounded probability or category can fit a decision model. A customer-facing explanation needs generation. A difficult, long-horizon investigation may justify a frontier reasoning model.
| Work shape | First choice | Escalate when |
|---|---|---|
| Known rule, formula, permission, or validation | Deterministic code | The rule depends on meaning that cannot be expressed reliably |
| Classify, score, route, gate, or detect | Jev, Clef, or another evaluated decision model | Confidence is low or the decision requires explanation and research |
| Summarize, draft, answer, translate, or extract flexibly | Fast general-purpose model | The task becomes long-horizon, highly technical, or multi-document |
| Complex migration, deep research, long multimodal analysis | Frontier model such as Gemini 4 Argon, when available | The action is high impact, irreversible, or outside authorization |
| Legal, financial, safety, security, or contractual approval | Qualified human owner | Never delegate final accountability to the model |
The four-layer route
Validate inputs, enforce permissions, calculate exact values, and reject impossible states.
Classify ambiguous language, score risk, select a workflow, and expose uncertainty.
Create the explanation, summary, content, or response from approved context.
Handle the small share of cases requiring sustained reasoning, many tools, or deep multimodal analysis.
Approve exceptions and own consequential outcomes.
Not every workflow needs every layer. The architecture is valuable because the boundaries are visible. A team can replace one model, change a threshold, or remove a layer without rebuilding the entire product.
Detailed route 01: a B2B lead journey
A prospect submits a form with company details, budget, timeline, requirements, and an open-text problem. The goal is not merely to generate a friendly email. It is to make a sequence of accountable decisions.
- Code: verify consent, normalize company fields, detect duplicates, and validate required inputs.
- Decision model: score fit, urgency, buying stage, request type, and missing information.
- Routing rule: send high-fit requests to sales, implementation questions to solutions, and incomplete requests to clarification.
- Fast generative model: draft a response using the selected route and approved service facts.
- Human review: approve pricing, commitments, unusual requirements, and strategic accounts.
- Frontier escalation: analyze a complex procurement package or multi-document technical brief only when the value justifies it.
The result is faster handling without allowing a model to invent qualification criteria or commercial promises.
Detailed route 02: a destination assistant
A visitor asks: “We have four hours, one wheelchair user, two children, and rain is expected. What can we do near the port?” The answer crosses live data, accessibility, geography, preferences, and time.
- Retrieval: fetch current opening hours, weather, accessibility records, travel times, and reviewed local information.
- Code: remove closed, unreachable, or time-incompatible options.
- Decision model: score suitability, accessibility confidence, family fit, and need for clarification.
- Generative model: compose a concise itinerary using only surviving options and cite the supporting sources.
- Human or partner confirmation: handle uncertain accessibility information or special bookings.
A larger model cannot repair a stale opening time. The knowledge and validation layers remain the product.
Detailed route 03: marketing operations
For campaign reporting, code should calculate spend, conversion rate, pacing, and period comparisons. A decision model can flag anomalous performance, classify probable issue types, and decide which campaigns need review. A fast language model can write the narrative. A frontier model is only useful when the job expands into a deep investigation across creative assets, customer research, attribution notes, and strategy documents.
This prevents the common error of asking a model to “analyze” numbers it should never have calculated in the first place.
Six questions before choosing the route
- What exact output must the software consume? A number, category, probability, paragraph, plan, or action?
- What is the cost of a wrong answer? Measure false positives and false negatives separately.
- Can correctness be checked? With a formula, schema, source citation, compiler, test, or expert review?
- What latency and volume are required? A live UI and a nightly research job have different economics.
- What data may the model see? Define region, retention, provider, access, and audit requirements.
- Who owns the exception? Every automated decision needs an appeal or correction path.
Keep the contract vendor-neutral
Define an internal response schema such as route, confidence, reasons, evidence, and requires_review. Write adapters for each provider. Store model name, version, prompt or question set, latency, cost, and outcome with every decision.
This creates the option to compare Jev with Clef, replace a generative model, or add Argon as an escalation path without changing downstream business logic.
Confidence is not permission
A 0.97 probability is a model estimate, not authorization to issue a refund, reject a candidate, publish a claim, or modify production. Permissions belong to the application and the organization. Confidence can inform the route; it cannot replace governance.
Thresholds should be tuned from observed errors. Track overrides, appeals, edge cases, and changes in traffic. When the input distribution moves, yesterday’s reliable threshold may become today’s silent failure.
What good routing achieves
- Lower cost because expensive reasoning is reserved for difficult cases.
- Lower latency because simple decisions do not wait for long generation.
- Clearer testing because every layer has a defined contract.
- Better safety because actions remain behind deterministic permissions.
- Less lock-in because providers sit behind adapters rather than inside business rules.
This framework combines my system-design analysis with vendor documentation checked on October 3, 2026. Availability and performance claims should be revalidated before implementation.
Route the work before buying the intelligence.
I design model-agnostic AI systems around real workflows, measurable gates, and safe escalation.
Map a workflow