Adoption of generative AI seems to have few reservations, and hopes are high about positive impact on empowerment, efficiency improvements in knowledge work, and productivity.

When considering strategic AI adoption, it is an unrealistic expectation that general-purpose AI tools, such as ‘chatbots’, are likely to be suitable for specific contexts or problems. Indeed, using the wrong tools can do harm, and there are now many documented cases of how the use of AI chatbots has caused failures in healthcare, legal, retail and other sectors. In this context, our focus is AI systems that use conversation as a powerful means of communication, especially in organisations.

The seminar reviews these problems and proposes an AI systems approach, discussing an example of how such systems that provide business context, focus, and clear objectives can be designed and built. To elucidate these ideas of a context-specific AI system, we explore the concept of AI conversations that are made better by aligning to stakeholder needs and values, and that can be reliably delivered at scale and low cost. Elements of an appropriate systems architecture will be discussed.

This is not about AI ‘chat’ and is about how design decisions are needed when dealing with non-deterministic interactions with language models. The example use case is the Ipseita Labs ‘CAIA’ system – meaning Conversational AI Alignment – which provides for the explicit design of coordinated, naturalistic interactions with generative AI and other resources.

This allows the power of LLMs to be harnessed and focussed to support:

(i) the design of purposeful AI conversations, particularly those which require the dynamic sensing and adjustment of behaviours to meet objectives, the ability to include specialist knowledge and private information, and the ability to create or to respond to events; and

(ii) the ability to integrate with organisational systems in the cloud or physically on premises, within a suitable orchestrated architecture.

Human conversations are complex, but we hardly notice that because we are trained in conversations from infancy.

What is important when designing AI conversations is that: stakeholder context and purpose is present and overarching, people’s goals and values are involved, conversations have memory, specialist or private knowledge may be involved, and information that is externally determined can be used but not changed by a conversation.

Conversations are personally and economically important: billions of conversations take place daily across the world. An uncountable number of conversations are started, continued or finished in organisations of every size. Inside organisations, conversations are crucial for the activities that make up processes. Conversations are needed to prepare for processes, are needed in carrying out processes, and are needed to complete, review and follow up processes.

Many of those conversations could be better: often workers don't have the information required or circumstances change during a process and such change is transmitted through conversations. AI can help.

The primary goal of the seminar is to show how AI systems which are to be quality assured and aligned to organisational or social values can be built to avoid inappropriate reliance on general purpose tools. In doing so, we expose and demonstrate the usefulness of properly designed and executed AI-driven conversations that represent stakeholder context, purpose, memory, and embedded values.

With an appropriate systems architecture, a rigorous conversational design method and quality-assured integration with organisational systems, these conversations can revolutionise organisations with streamlined operations & communication.

We hope that our argument in favour of tailored AI systems will persuade participants to rethink their AI adoption strategy and consider the well-established benefits of software and platform engineering.