When using generative AI, there are a lot of applications where a business cannot afford hallucinations that lead to errors. What to do to ensure accurate results?

Basically, you have two options: put a human in the loop to validate results, or build technical guardrails. Not even the first option is fully risk-free. Think about a diagnosis bot service helping doctors. A senior doctor can probably use it in a safe way, but how about an intern trusting it in a critical case, when they should really ask a more experienced doctor – but they do not?

There are many methods to build guardrails to mitigate the risk of incorrect generative AI answers. You could use your company documents to augment large language models (LLMs) using RAG (Retrieval Augmented Generation). For example, technical manuals describe service guidance and provide the information needed to provide answers based on your products, not general information for similar devices in the market, that may have served to train the LLM. However, there is still a risk for hallucination because using documents does not change the fact that the conclusion made by generative AI is based on the stochastic nature of its algorithms. In addition, using RAG successfully relies on correctly identifying all relevant document chunks for a given query, which is difficult to achieve at scale. RAG can lower the risk but hardly eliminates it!

Another approach is to use two or more generative AI services together: The first one generates a proposed conclusion and a second evaluates and gives an improved version of it. You could even use the third and fourth service for specific tasks, but the basic problem remains. In addition, this approach is costly: it consumes ever more tokens and is not performant. And you still need to manage the issue of relevance with RAG, with its inherent limitations.

The only way to completely mitigate the risk is to use a technique that uses a different approach than a probabilistic generative AI for making decisions. Business rules is a one such approach. A rule defines the business logic of a specific task. For instance, if a customer is in a particular segment or has spent so much with us over the past period, give them a particular discount. Business rules have been for decades in business automation tools – but without generative AI and its revolutionary capabilities of understanding and generating language. If you combine generative AI and business rules, you get something really game changing: The generative AI can understand intention, extract data, and formulate elegant text while the business rules can make clear, explainable, precise decisions that comply with regulations and business policy. In addition to reliability, this approach will make your service accountable because the rules used to make decisions can be traced and audited.

Athena Decision Systems specializes in creating hybrid solutions combining generative AI and business rules in areas such as customer service, public service, insurance claim processing, PO management, fraud, and AML. I am looking for customers and use cases where Athena can help you boost your business. Read here for more information: Athena Decision Systems