In the early 1930's, Lev Vygotsky formulated the Zone of Proximal Development. In short, he theorised that a child could learn how to execute a range of tasks effectively, within a guidance framework, but not yet fully independent.
His work led to what we today call "scaffolding". It is the framework we create as parents, teachers and employers to allow people to learn, experiment and ultimately become independent of outside support. "That's a proper grounded laaitjie (young person)", the older person at the family BBQ would say.
The importance of grounding in a generative AI solution has a lot of analogy to Vygotsky's work. This week the First Digital AI Research team put this to the test. Using the First Technology web site as data, and CVs from some of our account managers, we configured a chatbot to respond to a client's questions.
The Ungrounded Chatbot:
Using one of the leading LLM models available, and limiting it to the specific knowledge only, a typical conversation went as such:
User: Do you do Azure migrations?
Dimwit Bot: Yes, First Technology provides Azure migration and cloud services, enabling end-to-end Microsoft cloud solutions, including migrations.
User: Can somebody contact me?
Dimwit Bot: Certainly! Please share the best contact information, or alternatively, you can reach out directly to First Technology through their official website firsttech.co.za or via email.
Nothing wrong with that. Or is there?
The problem is not the chatbot's ability to answer simple or challenging questions. The problem is that we assume a that a one-way conversation is going to get the user to the place they want to be. Just image going on a date with this chatbot. The conversation will start with a lot of excitement, and you will be blown away by the general knowledge of your date. But soon isolation will kick in, when you realise that he/she/it has no interest in you, your intentions or your ambitions. As soon as you stop asking questions, the conversation goes quiet.
The Grounded chat bot:
Our team then crafted a grounding statement to give the chatbot a specific narrative. The instructions included how to respond to a product or service question, how to show an interest and how to get more information from the user with a context specific question as part of the response. The instructions then got to work (a call to action); if the user shows an engagement intent, establish their geographic location and industry, so that we could best match them with the most knowledgeable and local account manager from our team:
User: Do you do Azure migrations?
Prodigy Bot: Yes, First Technology offers comprehensive Azure migration services. As part of First Technology Group, First Digital provides end-to-end Microsoft Cloud solutions. Are you looking for licencing or implementation assistance?
User: I am looking for implementation assistance.
Prodigy Bot: For hands-on implementation assistance, First Technology, a specialist partner, provides in-depth support for Azure and other cloud migrations. This includes services from initial planning and assessment to full-scale deployment and post-implementation support. To connect you with the best person for this, could you please share your location or industry? That way, I can direct you to the most suitable account manager or Azure specialist.
User: I am in Cape Town, in the Manufacturing industry.
Prodigy Bot: Tammy Nicholson in Cape Town would be an excellent contact. Tammy is an Account Manager at First Technology with expertise in software solutions, services, and digital transformation projects. If you leave me your contact details, I will ask her to contact you. Is there anything more I can assist with?
Not only does a well architected grounding statement give the system a narrative, but it is crucial to anchor your AI system in clear, responsible principles, gives it a defined call-to-action, and ensuring it delivers a consistent, trustworthy, and human-centered experience.
Comments