Agent Copilot for Email

The goal sounded simple: make email communication smarter and more efficient for customer service teams. On closer look, however, you start to realize just how many variables there are:

- How do you create an AI tool that agents are comfortable using?
- How do you build in knowledge-base materials?
- What functionality is going to be the most helpful in streamlining the email process?
- How do you create something that makes sense in multiple languages?

Building this AI tool required us to carefully consider all of these factors. It required an understanding of the back-end hardware as well as the limitations we had on the front-end, which included use of Microsoft Adaptive Cards.

Through our exploratory user testing, we learned that agents would prefer to have the AI in their control, rather than coming into a pre-generated email, they wanted to create the prompt themselves. More sales-oriented agents performing shift-based work also wanted the ability to schedule their emails. Another factor that came into play was what type of third party integrations could be permissible here? Would we be able to suggest tasks based on the completed email?
‍
Last thing - in order to be able to create a good AI tool, it required some knowledge of LLMs and how they work. I took an awesome set of classes at Oracle that brought me up to speed on AI and made me feel even more empowered to help (🤩 I love hard tech and any excuse to learn it!).
‍
The flow up top is just a quick brain dump of what was on my mind and what needed to be researched before really diving into feature specific areas.

Based on our takeaways from user testing, we were able to create a multi-faceted email tool that allowed for regeneration, simplify and custom prompting. We are currently doing further research into what expand could be and we're learning that configuration of a lot of these features depends on how strong your prompting is.

We learned in the process that third party integrations beyond knowledge-base would require some further investigation which is planned for this coming year. Planned ideas for the following year include auto-complete and scheduling.