High Complexity, Speed Tolerant

Low Complexity, Speed Critical

Use case: complex customer inquiries, detailed product recommendations, multi-step problem-solving

Use case: intent classification, sentiment analysis, basic routing, spam detection

Example scenario: A telecomms company's enterprise support team deals with complex network configuration issues. Customers might say: "Our VPN keeps dropping when employees connect from the London office, but only on Tuesdays between 2-4 PM." This requires understanding technical context, temporal patterns, and suggesting multi-step solutions.

Example scenario: A bank's mobile app needs to instantly categorise customer messages: Is this a transaction dispute, account question, or general inquiry? This happens thousands of times per hour and needs sub-second response times.

Model rationale: These scenarios need models that can reason through complex, multi-variable problems. The extra processing time (2-5 seconds) is acceptable because the alternative is a frustrated enterprise customer and a potentially lost account worth millions.

Model rationale: A fine-tuned classification model can achieve 95%+ accuracy on this task in under 100ms, versus 2-3 seconds for a large language model. When you're processing 50,000 messages daily, those seconds add up to poor user experience and higher infrastructure costs.

Model choice: GPT-4o, Claude 3.5 Sonnet, Gemini Pro

Model choice: Smaller specialised models, fine-tuned BERT variants or even traditional ML

High Stakes, Accuracy Critical

Lower Stakes, Cost Sensitive

Use case: financial advice, medical guidance, egal compliance, fraud detection

Use case: content generation, basic personalisation, routine comms

Example scenario: An insurance company's claims processing system needs to flag potentially fraudulent submissions. A false negative (missing fraud) could cost hundreds of thousands of dollars. A false positive (flagging legitimate claims) damages customer relationships.

Example scenario: An ecomm platform wants to generate personalised email subject lines for millions of customers daily. The impact of a mediocre subject line is low, but the volume is massive.

Model rationale: This calls for multiple models voting on decisions, with human oversight for edge cases. You might use a specialised fraud detection model for initial screening, a large language model to analyse claim narratives for inconsistencies, and computer vision models to verify damage photos.

Model rationale: A smaller, specialised model trained on your email performance data will outperform a general-purpose model at a fraction of the cost. Even if it's only 80% as "creative" as GPT-4, the cost difference makes it the right choice.

Model choice: Ensemble approach and human-in-the-loop workflows

Model choice: Smaller more efficient models, open-source alternatives

Highly Sensitive Data

Less Sensitive Data

Use case: personal financial information, health records, internal communications

Use case: public website interactions, general customer enquiries, marketing content

Example scenario: A healthcare provider wants to use AI to help doctors draft patient notes faster. The AI needs to understand medical terminology and suggest relevant information, but patient data cannot leave the organisation's infrastructure.

Example scenario: A retail company's website chatbot helps customers find products and answers basic questions about shipping and returns.

Model rationale: This requires either deploying open-source models on-premises (like Llama 2 or Code Llama) or using private cloud instances of commercial models with strict data handling agreements. The slight performance trade-off is worth the regulatory compliance and patient trust.

Model rationale: Standard API calls to commercial models work fine here. The data isn't particularly sensitive, and the convenience and performance of cloud-based models outweigh privacy concerns.

Model choice: On-premises deployment, private cloud instances, models with strong privacy guarantees

Model choice: Cloud-based APIs, shared infrastructure

Complex Integration needs

Simple integrations needs

Use case: enterprise systems, real-time processing, custom workflows

Use case: standalone application, proof-of-concepts, simple API calls

Example scenario: A logistics company wants AI that can process customer shipping requests, check inventory across multiple warehouses, coordinate with carrier APIs, and send automated updates, all in real-time.

Example scenario: A content marketing team wants to automatically generate social media captions from their blog posts. They need something that can take a blog URL, extract key points, and create 3-4 caption variations optimised for different platforms, eg. LinkedIn and Instagram

Model rationale: This requires custom integration work, possibly combining multiple models with traditional systems. You might use a language model to parse customer requests, a recommendation system to optimise routing, and rule-based systems for final validation.

Model rationale: This is a straightforward content task that can be handled with a simple API integration. A tool like GPT-3.5 through a no-code platform like Zapier can automate this workflow without any custom development. The team gets immediate value without technical complexity.

Model choice: Custom implementations, hybrid architectures, specialised tooling

Model choice: Ready-to-use APIs, plug-and-play solutions