Business decision makers responsible for the success of their companies’ AI initiatives are leaving no stone unturned in their pursuit of positive ROI. For many, that quest begins and ends with AI models and the overlooked business value they create.
AI models impact numerous business workflows, including customer interactions, payment processing, warehouse and supply chain automation, and more. These power generative AI (GenAI) technologies that help enterprises manage, search, and summarize content and power agent AI applications and services.
AI models optimized to reach their full potential enable faster, more accurate, and more cost-effective processes. If these models are not fully optimized to address a company’s unique business needs, they can create illusions, produce inaccurate information, and make context-aware decision-making difficult. This can disrupt business operations and expose companies to security and compliance risks. To optimize AI technology and achieve positive ROI, companies apply different optimization techniques depending on the application.
Fully optimized AI models improve customer and employee interactions
Some chatbots use “generic” third-party AI models that can search product databases and provide recommendations in response to customer inquiries. However, it is unlikely that an out-of-the-box chatbot would be able to access a customer’s purchase history and other data to generate more effective recommendations. For example, search augmented generation (RAG) allows your model to access additional customer data.
Employees can benefit from AI models trained to guide a company’s internal business processes. The model is trained on process documentation within your organization to understand processes across all departments, from finance to research and development. But consider a situation where an employee has a unique question for HR. To provide the best answers, chatbots can spend excessive time on parsing, consuming more computing resources. all Rather than focusing on data directly related to HR, focus on enterprise data. In this case, the following optimization techniques are used: model pruning Tell your chatbot to ignore certain data that it deems irrelevant. The result is a trimmer model that is smaller, faster, and requires fewer computing resources to operate.
How to optimize AI models
Whether companies use home-built models, third-party models, or both, optimization techniques exist to align AI with business needs and improve model effectiveness.
1. Search extension generation
RAG is a technique that provides a trained model with access to additional data that was not present during training. This additional data can help improve model accuracy, especially in use cases that require more context awareness than the model’s training data provides.
2. Compression
Model compression reduces the size of a trained model by reducing the total amount of data that the model parses or the effort spent when parsing the data. Pruning is a compression method that reduces model operating costs while improving accuracy when accessing relatively narrow datasets. Quantization is another compression method that speeds up processing and reduces cost, but it also reduces accuracy.
3. Retraining
Similar to RAG, retraining allows the model to access data that was not initially available during training. However, retraining digs deeper. This allows the model to discover new relationships and patterns in the data. This feature differs from RAGs, which allow the model to interpret additional data based on patterns already recognized within the training data. Although more costly and complex than RAG, retraining is more flexible when business processes fundamentally change and the model requires updated data. This is also useful when upgrading a model that was initially trained on low-quality data with higher-quality data.
4. Rehost and redeploy
In some cases, the root cause of suboptimal model behavior is not the model itself or the data it accesses. Rather, the model may lack the computational resources necessary to run effectively. Rehosting or redeploying your model on new infrastructure can improve model performance. For example, AI accelerators speed up the model’s inference capabilities and respond to prompts faster.
5. Input and output filtering
Users can optimize the model without changing the model itself. Filtering techniques can intercept and modify user or AI agent prompts and model responses. If your AI is expensive to operate because users send long prompts, filtering removes irrelevant parts of the prompts and reduces the amount of data your model needs to process, thereby reducing processing costs.
Best practices for AI model optimization
Improving AI models can be complex and difficult, but the following best practices can reduce risk and reduce the resources needed to achieve model optimization.
- Choose the right optimization method. Optimization techniques such as retraining can be complex and expensive, while RAG and input filtering are simpler and less expensive. If you have limited time or resources, filtering may be the best option.
- Ensure sufficient technical resources. Model optimization requires specialized knowledge. Companies without AI engineers on staff should consider working with a vendor to meet their optimization needs.
- Experimental support. The first attempt to optimize an AI model does not always yield the desired results. You may need to retrain your model multiple times to reach your target accuracy rate.
- Define a process for periodic model evaluation. Make continuous model improvement a systematic part of your overall AI project strategy. Establish guidelines for regularly reviewing AI models to determine if optimizations are needed.
Chris Tozzi is a freelance writer, research consultant, and professor of IT and social studies. He previously worked as a journalist and Linux system administrator.
