How can I use AI to analyze lots of sales calls?

The technology Convvo uses to solve the problem of analyzing sales calls at scale using AI

Large Language Models (LLMs) like ChatGPT, Claude and Gemini are trained on massive amounts of data, but when you use AI to answer a specific question, then the amount of context it can reference in one go is limited.

This context may be large enough to process the transcripts of a few calls (depending on how long those calls are), but a single salesperson can eclipse the context limit in a single day. And that's just one person. 

What if I wanted AI to help me understand one year's worth of calls made by every salesperson at my company? This could mean 1,000s or 10,000s or 100,000s of calls. 

It's simply not possible to process that scale using traditional LLM chatbots like ChatGPT.

At Convvo, we are attempting to solve this problem of scale, via a model which tries to understand your portfolio of calls as a whole. This occurs in three main stages.

Stage One: Categorisation

First, when we pull in your inventory of calls, we mark each call up with tags that help the system understand what calls might be useful to reference for a particular question. 

Calls may be broken down into excerpts, as not every part of every call may be useful. Some parts of the conversation, like small talk, are rarely useful, so adding a tag of "small-talk" allows us to easily filter out these sections.

The tagging process is informed by our understanding of what marketing and sales teams find most useful. We store a unique list of the tags for your portfolio, which helps us in the following stage...

Stage Two: Building a Context Window

When you ask a question, our AI identifies which tags are most relevant and then uses that to select which excerpts should be used to build up a context window small enough to be understood by an LLM.

Since there may not be space for all relevant excerpts, they are ranked for relevance based on a range of criteria, I including recency, how many relevant tags that particular excerpt has, and user-submitted markers like "has the user manually interacted with or categorised this transcript?"

We also need room in the context for the last few questions and answers, so the LLM remains aware of the conversation.

Stage Three: Answering Your Question 

Once the context window is built, we simply send your question to the LLM with that refined context. This makes it possible to answer questions on your entire portfolio of calls, quickly. 

Usually, answers are generated within a minute, which is much faster than alternative solutions like fine-tuning. This also means new calls can be added or removed at any time.

(We are working on ways to incorporate fine-tuning into our product offering, but that is a topic for another time!)

For best results, we recommend a mixture of human and automated categorisation: you can manually group calls, and then start a chat using just that group. This allows our technology to hone in on what's important with greater accuracy and speed.