Generative AI (GenAI), built on powerful Large Language Models (LLMs), is changing the way businesses make decisions. Unlike traditional AI, which needs a lot of data to make predictions, GenAI can understand and create natural language. This makes it easy for business leaders to make quick, informed decisions without needing a lot of data.
At TAZI, we use both traditional AI and GenAI to help business users make better decisions more easily. Unlike traditional AI models that need clear steps and a lot of data, GenAI can create useful content right within a business process, making things run smoother and more efficiently.
While traditional AI models focus on specific results using complex methods, GenAI uses LLMs to produce text. This means GenAI can translate languages, turn speech into text, or create reply drafts, offering a flexible tool for many business needs.
Generative AI Methods Supported by TAZI Platform
The TAZI platform uses GenAI methods like Retrieval Augmented Generation (RAG), zero or few-shot learning, and fine-tuning of models. These techniques allow GenAI to work with data sets for tasks usually handled by traditional AI, but often with much less data. This means GenAI can achieve similar results to traditional AI by using just a small portion of the data.
To make these concepts clearer, let’s look at how they’re applied in a real life example such as classifying customer complaints:
Example: Traditional AI vs GenAI approach to Customer Complaints Classification
Approach 1: Traditional AI
Using AI for classifying customer complaints requires a lot of training data. This data includes the complaint texts as inputs and their correct labels as outputs. Before being used, data often needs preprocessing, where complaints are broken down into words or sentences. These are then fed into an AI model, like a neural network or decision tree, trained to match these inputs with the correct labels as closely as possible. When a new complaint comes in, the AI model classifies its severity.
The AI model’s performance is checked by how well its predictions match the actual labels from the training or testing data. If it’s not doing well, solutions might include changing the preprocessing or model, tweaking model parameters, or getting more data for parts where the model struggles.
The TAZI platform makes it easy for anyone to create and adjust AI models for this purpose, thanks to its no-code interfaces. You can track how well the AI and the business are doing and update your model with new data and feedback from experts in customer outreach. Usually, teams like customer outreach or business analysts handle creating, deploying, and maintaining these models, learning from the expertise captured in the training data.
Let’s see how similar tasks can be approached with Large Language Model (LLM) techniques.
Approach 2: GenAI
Zero or few-shot learning simplifies the process of classifying customer complaints by minimizing the need for extensive training data. Here’s how it works:
- Model Setup: Choose a Large Language Model (LLM) and set parameters like temperature (to control randomness), context window size (how much text the model considers), and response length. The task is defined in a prompt. Zero-shot learning doesn’t require examples, while few-shot learning uses a few example inputs and outputs to guide the model.
- Classifying New Complaints: To classify a new complaint, combine the prompt (with examples for few-shot learning) and the complaint text as context for the LLM. You can either use a fixed set of examples for all inputs or adaptively select examples based on each new complaint, using methods that identify similarities with previously labeled complaints.
- Adaptive Learning: Zero or few-shot learning models can adapt based on the descriptions and data labeled by domain experts, reflecting their past actions and decisions.
This approach leverages the LLM’s ability to understand and generate responses based on a small number of examples, or even just a task description, making it a flexible and efficient method for handling customer complaints without the need for a large dataset.
RAG (Retrieval Augmented Generation)
Retrieval Augmented Generation (RAG) is a sophisticated technique that combines retrieving information and generating responses based on that information. Although not the prime choice for customer complaints classification, understanding its application provides insight into its versatility.
A more typical use of RAG is in document analysis for answering questions. For instance, in handling an insurance demand letter, RAG can extract specific information (like provider names or diagnosis codes) and present it in an organized format. It enables users to ask complex questions about documents, as if consulting an expert. Another application could be analyzing educational material or entertainment content to answer related questions.
Application in Customer Complaints:
With TAZI, RAG can be used by leveraging internal process documents instead of traditional training data. Documents are broken down into sections and transformed into numerical vectors using an LLM. When a new record is received, it’s also converted into a vector. The system then finds the most relevant document sections. Another LLM generates a response using the input record (such as complaint in our example), relevant document sections, and guidance on formulating the reply.
While RAG excels at extracting and organizing information, traditional AI models might be better suited for identifying anomalies in data, such as detecting unusual patterns in insurance claims, by learning from examples of what’s considered normal versus anomalous.
Fine-Tuning of the LLMs
TAZI supports fine-tuning an LLM that involves making specific adjustments to the model, so it can learn from a particular dataset, enhancing its performance for specialized tasks. This process is distinct from zero/few-shot learning or RAG, where the LLM remains unchanged and is applied as is.
Fine-tuning updates the LLM’s parameters, usually the weights of the output layers, to better match the new data. This requires substantial computational resources and expertise to ensure the model adapts without losing its general capabilities.
Fine-tuning is useful for improving the accuracy of an LLM on unique datasets or developing a smaller, more efficient version that performs similarly to a larger model. For customer complaints classification, fine-tuning would involve using past classifications from the customer outreach team as training data.
Beyond complaints classification, fine-tuning can help create more accurate responses for customer support by training on specific interactions. Another application is in project management, where an LLM could be trained on past project requirements and test cases to generate new test cases that align with organizational standards.
GenAI, AI, and Analytics
Both AI and Generative AI (GenAI) offer solutions to business problems like in the example of customer complaints classification. However, GenAI excels in content creation, such as drafting replies to complaints or claim responses, with the ability to tailor the tone, sentiment, and various other criteria.
At TAZI we believe that GenAI can work alongside analytics to identify trends in customer sentiment, topics, or case types from messages. This insight can inform call center automation or guide product development teams by highlighting trending issues.
Traditional AI models are adept at spotting sudden changes in data trends, such as shifts in sentiment or topic frequency, offering a powerful tool for monitoring and responding to customer feedback dynamics.
At TAZI, this combination is key to the success of AI solutions. AI can handle tasks like classifying customer complaints and predicting churn, while GenAI can generate nuanced responses for customer outreach based on AI insights. This collaborative approach enhances customer service and operational efficiency.
For structured, tabular data or specific organizational datasets, AI is often the go-to method. GenAI shines in generating content. Strategically combining AI and GenAI can enhance problem-solving, potentially lowering costs and improving accuracy by optimizing the use of each for different data types or tasks.
Trustworthy AI and GenAI
Ensuring the trustworthiness of AI and GenAI systems is crucial, especially when these systems are deployed in areas impacting human lives. The European Union has established guidelines to assess the trustworthiness of AI systems, focusing on seven key criteria: human agency and oversight, technical robustness and safety, privacy and data governance, transparency, diversity, non-discrimination and fairness, societal and environmental well-being, and accountability. While implementing these criteria, the following guidelines are helpful:
It’s vital that domain experts are actively involved in the AI system’s life-cycle, providing essential oversight and decision-making capabilities to ensure humans remain in control.
AI systems should only access the data necessary for their tasks, strictly avoiding the use of personally identifiable information (PII) without proper safeguards, particularly when working with external LLMs.
The deployment of AI systems must include input from legal and IT departments to ensure continuous compliance and guidance throughout the system’s design and operation.
Quality assurance (QA) teams should regularly review the outputs and effectiveness of AI systems to make timely decisions on necessary updates or adjustments.
Feedback from human experts and the iterative design changes are essential for keeping AI and GenAI systems current and responsive to emerging needs and challenges.
For more details on Trustworthy AI, please see the Presidential Order and EU Trustworthy AI standards.
MLOps and LLMOps
In addition to the extended set of ML and business performance metrics available for AI models and their ensembles with GenAI models, TAZI also offers metrics on LLM models’ performance. These additional metrics include context relevance, answer relevance and groundedness. Due to its easy model creation and maintenance, TAZI offers creation of key customized business indicators that are computed based on the outcomes of the GenAI models. These metrics can be computed either through analytics, or through the utilization of easily created and deployed AI models. TAZI also offers monitoring for changes in data and models. All of the monitoring interfaces are easy to use and designed to be utilized by business and technical teams.
