AI Patterns and Model Strategy

Guide · Enterprise AI Solution EngineeringPage 6 of 18Overview → … → AI patterns

Executive summary

Choose the solution pattern before the model. Match classification, extraction, RAG, forecasting or agents to the job, route large vs small models deliberately, and wrap generative steps in deterministic controls so the model does not control everything.

When to use this page

• Use when the problem is clear but the AI approach is not.
• Use before RAG or agentic design deep-dives.
• Skip to RAG or Agentic AI once the primary pattern is agreed.

Executive view

Insist on pattern clarity and “rules vs AI” before model shopping or vendor demos.

Decision required: Which pattern creates value—and is a simpler non-AI approach better?

Technical view

Score models on quality, latency, cost, security and ops maturity; draft large vs small routing rules.

List deterministic controls (schema, permissions, thresholds, approvals) beside every generative step.

Owners

Owner · AI Solution Engineer

Contributors · AI engineering, architecture, product, domain SME, security

Select the pattern before the model

The first question should not be:

Which model should we use?

The first question should be:

What solution pattern is appropriate?

Common AI patterns

Pattern	Typical uses
Classification	Ticket routing, risk categorisation, intent detection, document classification
Extraction	Invoices, contracts, forms, entity recognition
Summarisation	Case notes, meetings, research, interactions, reports
Retrieval / QA	Answers grounded in enterprise knowledge
Recommendation	Next-best action, products, content, decision support
Forecasting	Demand, revenue, capacity, failure, markets
Optimisation	Scheduling, routing, allocation, pricing, networks
Agentic workflow	Plan, select tools, execute, evaluate, multi-step tasks

Model selection criteria

Evaluate models across quality, latency, cost, context length, tool use, structured output, multimodal capability, region availability, security, data-use policy, deployment model, fine-tuning support and operational maturity.

Large models versus small models

Use larger models for complex reasoning, ambiguous instructions, multi-document synthesis and advanced tool use.

Use smaller models for classification, routing, extraction, repetitive structured tasks, high-volume workloads and edge deployment.

Deterministic and generative components

Enterprise AI should combine deterministic controls with probabilistic models:

• Schema validation
• Business rules
• Permission checks
• Confidence thresholds
• Approved tool lists
• Workflow states
• Human approval
• Transaction validation

The model should not control everything.

Decision required

Justify the chosen pattern against alternatives, and do not proceed to model selection until the pattern statement is agreed with the sponsor.

Case study: financial-services assistant

Primary pattern: retrieval and question answering, with classification for intent and risk routing, and deterministic policy checks before any write action.

Model strategy: small model for intent and routing; stronger model for grounded synthesis; never the largest model for every greeting.

Output

Pattern statement with justification, model criteria scores for shortlisted options, large vs small routing rules, and a list of deterministic controls beside generative steps.

Common failure modes

• Model shopping before pattern clarity
• Generative answers where extraction or search would suffice
• No deterministic envelope around the model
• One model for routing, reasoning and high-volume classification

Solution Engineer checklist

Solution Engineer checklist

• Pattern chosen and justified against alternatives
• Model criteria scored for shortlisted options
• Large vs small routing rules drafted
• Deterministic controls listed beside generative steps

Practical exercise

Rewrite a “chatbot” request as a pattern statement: “This is retrieval-augmented QA with classification-based escalation, not an open-ended agent.” Share it with the sponsor and confirm they agree.

Key takeaways

• Pattern before model; AI is a means, not a goal.
• Route large and small models by task economics and risk.
• Wrap generative steps in deterministic controls.
• Prefer simpler patterns until autonomy is justified.

What to do next

1. Continue with RAG engineering when retrieval/QA is the pattern.
2. Continue with Agentic AI systems only when autonomy is justified.
3. Revisit Architecture for multi-model routing fit.