The right planning ensures that the business can accelerate its implementation because the risks are understood and can be managed when other risk events occur. AI tools must be used in combination to achieve the desired outcomes. In contact centres, the desired outcomes are often to reduce the cost-to-serve (the cost a business incurs to serve a customer) while striving to increase the customer’s delight factor.

Agentic Generative AI processes are complex and can produce imperfect answers, hallucinate, or “go off the rails” entirely through interaction with bad actors or through their inherent non-deterministic outputs. Integrating them into existing data and
business systems is also complex and involves more than just pressing a button.

To prevent inherent problems with AI and to simplify integration into business data and systems, a well-designed AI solution must, by necessity, be highly modularised with the ability to assemble various tools and functionality into workflows. These modularised  workflows enable the creation of solutions that can mirror complex human interactions and leverage the immense power of agentic GenAI safely using real-time safeguards,
auditing, and the ability to place a “human in the loop” at each stage.

Modularisation

The ideal solution should be implemented in a modular and open manner to ensure safety. The client should have access to the logic and sequences that generated a response. Within a contact centre, this ability to check and recheck each part of the steps used to create a response must be fully understood. GenAI solutions should be tightly regulated. This can be done by implementing them in a modular format with structured output requirements or secondary analysis of outputs. This creates simple modular blocks that can be interrogated individually and combined into highly sophisticated, auditable, testable, and controllable processes for complex tasks.

Governance Controls

AI controls must be implemented to prevent the generation of information that is not aligned with company policy. Implementing an AI system requires checks and balances to ensure that the outputs from an AI system that leverages GenAI are controlled with deterministic outputs. The controls must consider broader ethical, moral, and biased concerns and the practicality of the ideal outcome that will quickly solve a customer’s issue.  These controls can be applied as part of a secondary analysis to inputs and outputs from a given GenAI module.

Intelligence Layer

For a modularised AI system to be implemented effectively, an intelligence layer must combine and control the modules that chain workflows, data, and integrations across multiple platforms to deliver concise intent and sentiment-based AI responses in seconds. By separating an AI system into building blocks, businesses can integrate AI safely, systematically, and with confidence that the company and its customers are not being put at risk. The intelligence layer can be implemented in the following way:

• Categorisation Module: Customer communications are categorised from initial queries to replies.
• Intent Module: Determines the intents within the content and can be implemented as follows;
  • Simply identify the following: “What is the main intent?”
  • Or a more complex one – “What is the main intent, and what are the other intents? How do they relate to each other?”
  • Intent identification could be limited or directed to a fixed list of available actions or left open based on the actual intent expressed by the customer.
  • The intent module monitors the status of each intent in terms of the resolution process.
  • Finally, to ensure completeness, intent gathering occurs during every interaction, and as the conversation extends, it becomes an array of intents.
• Sentiment Module: Gathers multiple sentiment levels, including emotional state and customer-specific requirements such as vulnerability. This is critical when faster responses for those in need must be prioritised or when providing consistent, compassionate responses based on emotional state.
• Action Engine: This is a series of tools that can perform actions in the real world, such as getting/sending data to/from third parties and Interacting with external APIs.
• Retrieval-Augmented Generation (RAG): It incorporates a company’s proprietary intellectual property, not part of a large language model. This intellectual property is used to build a contextual response. It is not “shared” externally but is tightly controlled in a dynamic database ideal for complex, secure, enterprise-grade implementations.