How to choose the right AI model for your use case

Introduction

"Should we use GPT-4, or should we train our own model?" "Is transformer architecture the right choice?" "What about fine-tuning vs. prompt engineering?"

At Hyperbolic, we are often confronted with these questions from clients who want to get started with AI but feel overwhelmed by the options. The landscape of AI models is growing exponentially, and it can be difficult to navigate.

The good news is that the choice doesn’t have to be complicated if you ask the right questions. In this article, we share our framework for choosing the right AI model for your specific use case.

The five fundamental questions

Before we even look at specific models, we always answer these five questions:

1. What is the specific task?

AI is not one thing; it encompasses many different techniques for different tasks. Do you need to:

• Classify something? (e.g., is this email spam or not)

• Generate text? (e.g., product descriptions)

• Predict numbers? (e.g., next month's sales)

• Find patterns? (e.g., customer segmentation)

• Process images? (e.g., quality control)

• Understand speech? (e.g., transcription)

Different tasks require different types of models.

2. How much data do you have?

This is often the most critical factor. A rule I use:

• Fewer than 1000 examples: Use pre-trained models or simple ML methods

• 1000-100,000 examples: Fine-tuning pre-trained models or classical ML

• More than 100,000 examples: Consider training from scratch or extensive fine-tuning

3. What are your performance requirements?

Think about:

• Latency: Should the response come in milliseconds, or are seconds okay?

• Throughput: How many predictions should you make per second?

• Accuracy: How critical is it for the model to be correct?

A chatbot can tolerate a 1-2 second response time. Fraud detection in a payment gateway needs to happen in under 100ms.

4. What is your budget?

AI can be expensive. Consider:

• Training costs: How much will it cost to train the model?

• Inference costs: How much does each prediction cost?

• Maintenance: What does it cost to update and maintain the model?

Large language models like GPT-4 can incur significant costs with high usage. Sometimes a simpler model that costs 1% of the price and delivers 95% of the value is the right choice.

5. How critical is explainability?

In regulated industries, like pharma where we have extensive experience, explainability is often non-negotiable. In other contexts, it may be less critical.

Deep learning models: Powerful but hard to explain Classical ML models: Often more explainable

Categories of AI models and their use cases

Let’s go through the primary categories and when to use each:

Classical Machine Learning

Models such as logistic regression, decision trees, random forests, and gradient boosting.

When to use them:

• You have structured, tabular data

• You have limited data (under 100,000 examples)

• Explainability is important

• You need fast inference

• Budget is constrained

Examples from our projects:

• Credit scoring in the financial sector

• Predictive maintenance in manufacturing

• Customer segmentation in e-commerce

• Fraud detection

Advantages:

• Fast to train

• Requires less data

• Highly explainable

• Cheap to run

Disadvantages:

• Requires feature engineering

• Weaker on unstructured data (images, text)

• Poorly scales to very complex patterns

Deep Learning (Neural Networks)

Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Transformers, etc.

When to use them:

• You are working with unstructured data (images, text, audio)

• You have a lot of data (100,000+ examples)

• The task is complex

• You have computational resources

Examples:

• Image recognition in quality control

• Natural language processing

• Speech-to-text systems

• Time series predictions with complex patterns

Advantages:

• Extremely powerful for complex tasks

• Can automatically learn features

• State-of-the-art performance on many tasks

Disadvantages:

• Requires a lot of data

• Expensive to train

• "Black box" - hard to explain

• Overkill for simple tasks

Large Language Models (LLM)

GPT-4, Claude, Llama, etc.

When to use them:

• Text generation or understanding

• You have little or no training data

• The task requires broad knowledge or reasoning

• You can tolerate certain costs

Examples:

• Customer service chatbots

• Content production

• Document summarization

• Code generation

Advantages:

• Requires no or minimal training data

• Extremely versatile

• Constant improvement from providers

Disadvantages:

• Can be expensive at scale

• Potential hallucinations

• Dependency on third-party services

• Less control over model behavior

The decision tree: Which model should I choose?

Here is the framework we use at Hyperbolic:

Step 1: Determine the domain

Text tasks:

• Simple classification (sentiment, categories): Start with classical ML or fine-tuned BERT

• Complex understanding or generation: LLM (GPT-4, Claude)

• Domain-specific text with data: Fine-tune a transformer model

Image tasks:

• Classification with limited data: Transfer learning from pre-trained CNN

• Complex image analysis with data: Train custom CNN

• Simple pattern detection: Traditional computer vision

Structured data (tables):

• Almost always: Start with gradient boosting (XGBoost, LightGBM)

• Only if very complex interactions: Consider neural networks

Time series:

• Simple patterns: ARIMA or classical ML

• Complex patterns with external data: LSTM or Transformer-based models

Step 2: Assess data amount

Minimal data (<1000 examples):

• Text: Use LLM with prompt engineering or few-shot learning

• Images: Transfer learning from pre-trained models

• Structured: Classical ML with careful feature engineering

Moderate data (1000-100,000):

• Fine-tuning pre-trained models

• Classical ML with good feature engineering

• Small neural networks

Large data (>100,000):

• All options are open

• Consider return on investment of training large models

Step 3: Balance performance vs. costs

At Hyperbolic, we always conduct a cost-benefit analysis:

Option A: GPT-4 via API

• Training: 0 kr

• Per prediction: 0.10 kr

• With 10,000 predictions/day: 1000 kr/day = 30,000 kr/month

Option B: Fine-tuned open source model

• Training: 20,000 kr (one-time cost)

• Hosting: 5,000 kr/month

• Per prediction: ~0.001 kr

• With 10,000 predictions/day: 10 kr/day = 300 kr/month

• Total first month: 25,300 kr

• Total subsequent months: 5,300 kr/month

Over a year, Option B is significantly cheaper, but requires a larger upfront investment.

Practical cases from our work

Case 1: Document classification in pharma

Task: Classify regulatory documents into 20 categories

Our choice: Fine-tuned BERT model

Why:

• We had 15,000 labeled documents (good data amount)

• High accuracy requirements in a regulated industry

• Need to be able to explain decisions

• Sensitive data could not be sent to external APIs

Result: 96% accuracy, full control and compliance

Case 2: Product descriptions for e-commerce

Task: Generate product descriptions based on product data

Our choice: GPT-4 via API with structured prompting

Why:

• No existing training data

• The task requires creativity and linguistic finesse

• Moderate volumes (100 products/day)

• Time-to-market was critical

Result: Launched in two weeks, high quality, acceptable costs

Case 3: Predictive maintenance

Task: Predict machine failures based on sensor data

Our choice: XGBoost (gradient boosting)

Why:

• Structured time series data from sensors

• Need to explain predictions to technicians

• Fast inference necessary (real-time monitoring)

• 50,000 historical data points available

Result: 89% accuracy in failure prediction, 40% reduction in unplanned downtime

Common mistakes to avoid

Mistake 1: Choosing the most hyped technology

Just because everyone is talking about transformers or GPT-4 does not mean it's the right choice for your task. Sometimes a simple logistic regression is all you need.

Mistake 2: Underestimating classical ML

Many jump straight to deep learning without trying classical methods first. In our experience, gradient boosting solves 70% of structured data problems excellently.

Mistake 3: Ignoring inference costs

A model that performs 2% better but costs 10x as much to run is rarely worth it.

Mistake 4: Forgetting maintenance

The most advanced custom model requires specialized expertise to maintain. Are you ready for the long-term commitment?

Future considerations

Start simple, complicate later

At Hyperbolic, we almost always recommend starting with the simplest model that can solve the problem. You can then gradually increase complexity if necessary.

It’s easier to upgrade from a simple to a complex model than to downgrade from an overly complicated solution.

Build with flexibility

Design your system so that you can switch models without having to rewrite the entire application. Use clear API boundaries between model and business logic.

Measure and iterate

No one picks the perfect model the first time. Measure performance in production and be ready to adjust.

Tools to assist choice

We often use these tools to evaluate different models:

AutoML platforms:

• Google AutoML

• H2O.ai

• Auto-sklearn

These can automatically test different models and help find the best one for your data.

Benchmark tools:

• MLflow to track experiments

• Weights & Biases for visualization

• TensorBoard for deep learning

Conclusion

Choosing the right AI model is not about finding the most advanced technology. It’s about matching the problem with the right solution based on data, requirements, budget, and expertise.

At Hyperbolic, we always start by deeply understanding the business problem before looking at technical solutions. We almost always recommend starting simply and moving toward complexity only when necessary.

Remember: The best model is the one that delivers the necessary value at the lowest total cost, and one that you can actually maintain over time.

Are you unsure about which AI model fits your use case? Contact us at Hyperbolic for a technical workshop where we identify the optimal solution together.