We recently decided to add an AI-enabled chatbot to Blanchard’s most popular leadership development program, SLII®.
We did this with a clear purpose in mind: to help our learners transition from a positive training experience back into the workforce. You see, far too many learners leave an SLII® class excited to use their new knowledge and skills, only to be overcome by workload and habit and realize a month later that their SLII® skills are fading fast. We believed an AI-powered chatbot could help them over this chasm through a combination of reminders, tools, and just-in-time support. So we set off to build the SLII® Chatbot™.
Together with our platform partner Mobile Coach, we got to work.
Preparing for two types of interactions
Chatbots can be structured or unstructured. A structured interaction is planned out ahead of time, with prompts written to guide users through it. An example is the “dial or say 1” experience we’ve all had calling an 800 number.
Unstructured dialogs are not prewritten. ChatGPT is a good example of this type of interaction, where an AI system is trained using a Large Language Model (LLM) to interact with humans using our natural language. We wanted both a structured and unstructured system for our SLII® Chatbot.
The structured part of the chatbot would be focused on helping learners through reminders and assistance for one to three months after their training. We wrote the dialog that the bot would say and the questions it would ask. We used preformatted responses; e.g., “type 1 for tools, type 2 for tips.” Our structured script ended up being about 50 pages in length. Our key learnings were to be clear about the goal, to allow plenty of time for development and testing, and to blend information with entertainment.
We also recognized there’s a whole different writing style and syntax used in text messaging that includes acronyms (e.g., lol, btw), use of emojis, lack of punctuation and capitalization, etc. We chose to adopt these in the SLII® Chatbot. It helps to be clear about your writing style guidelines up front.
Venturing into new territory
Creating the structured dialog was a lot of work but it was work we were familiar with, having done much scripting over the years. The unstructured, conversational part of the chatbot was new territory.
Our goal was to support learners through an SLII® Chatbot that could respond to any question asked. This required a smart AI engine based on an LLM.
Again we leaned on our partner, Mobile Coach, who offered a private LLM so that we could safely assist our clients without feeding our proprietary content to the public ChatGPT system.
We wanted our SLII® Chatbot LLM to be an expert in SLII®, so we started by feeding it existing documents from our SLII® program, facilitation notes, and supporting materials. The LLM was quite flexible in its ability to consume documents, having learned this by absorbing a wide variation of content across the web.
Designing prompts
The next step was designing prompts. To understand an LLM prompt, think of a sandwich. A user might type a question to the bot such as “I just had a disagreement with a team member—what should I do?” Think of that as the meat of the sandwich. Although the LLM would be able to respond to that question alone, the response would be much more useful and controllable if you embedded the question in a larger prompt. Think of adding bread, lettuce, tomato, and mayo to the sandwich.
Prompts are able to do many things, but we focused on context and writing instructions. Context is about providing the backstory around the question so that the LLM can more effectively understand what’s going on. This can take the form of descriptive text that explains what the user is doing or trying to accomplish. An example is “I asked an employee how their one-on-one conversation went with a team member and they commented: [user’s response].” It’s like how we provide a little context when a new person joins a conversation in progress: “Oh, we were just talking about the merger announcement.”
The writing instructions you provide to the LLM in these prompts can cover many aspects of writing. Whether you know it or not, you probably have an idea of what appropriate and inappropriate responses look like. I remember an early response that was way too long, so we added an instruction around length to limit the LLM responses on that dimension. You can also tell the LLM what tone you are looking for, and where to pull content from. Here’s an example: “Please write an upbeat response that validates their struggle, is less than 100 words, and pulls at least one concept from the following content: [source list]”
Care and feeding during these early days
The combination of providing lots of source content, offering context and backstory, and defining the tone and style of written responses will give you a good start. The last tip is to invest in care and feeding. LLM chatbots can be easily enhanced and enriched over time by reviewing the performance of the bot and working on interactions that aren’t serving users well.
It still feels like early days in the LLM chatbot space. We are excited to be experimenting with this technology for such an important purpose. I’d highly recommend you give it a try if you can stick with it for the long haul. Good luck!
_____________________________________________________________
Editor’s Note:
Would you like to learn more about AI applications for learning and development? Download our free eBook, An HR / L&D Primer on Generative AI. You’ll find this post, together with five other perspectives from Blanchard thought leaders exploring the challenges leadership, learning, and talent development professionals are facing in today’s changing environment. Use this link to download your complimentary copy now.
About the Author
More Content by Jay Campbell
