How To Boost Revenue By Analyzing Sales Notes With AI (No ChatGPT / Token Limit)

Imagine putting all that effort into pitching hydrophobic surfaces, when clients actually just want a vacuum cleaner… 🤦

Wouldn’t it be great if you had a way to find out what clients really wanted to talk about on sales calls, before buying your solution? Here’s how!

Most businesses don’t use huge amounts of data they’ve amassed throughout years of selling, recorded in their meeting notes. Did you know you can use AI to find out what to talk about on sales calls to maximize revenue, personalized to your business?

Meeting notes contain valuable information 💡

The question is: how to extract that information.

If you’re a sales professional, you’ve probably tried using a bunch of different selling systems, changing your pitch and the questions you ask every time. Experience tells your gut what works and what doesn’t, but understanding what really drove your customer to buy is hardly ever backed by real-world statistics.

Pasting thousands of meeting notes into ChatGPT doesn’t exactly help you out here. We want to know: which types of conversations lead to high-revenue paying clients?

What if I told you there is a way to find out which sales strategy statistically works best for you? All you need is the history of your meeting notes and the revenue you made with each client.

What follows is an outline of how you might tackle this problem.

Step 1. Calculate the average revenue for each interaction 💰

When determining which sales strategy works best, it’s obviously important to look at which conversations convert to a follow-up conversation or deal, and which don’t. But what’s even more important is whether or not you’ve set yourself up for a good long-term relationship with that client.

Revenue is a decent measure of this: building sustainable repeat-business with happy clients is way more profitable (and fulfilling) in the long run than opportunistic deal-closing.

That being said, looking at final deal value is not the worst way to estimate the value of a sales pitch. Pragmatically speaking, it might sometimes be more appropriate to consider conversion rates instead of estimated revenue. Do what works for you.

The most simple way to get the revenue per call is to divide the total deal value by the number of calls.

If the conversation didn’t convert to a deal, then you count €0. If it did, and 5 calls ended up in a deal of €10k, then you count (deal value)/(number of calls) = €2k.

Although rocket scientists 🚀 might disagree with this approach, the point is: you want to put a consistent value estimate on each conversation. If you have a good CDP in place you can get a better estimate, and there are many other ways to model the value of a single call properly. Make sure you choose a distribution that doesn’t skew your results as a function of the number of calls you had with each client.

What we’re concerned with here is finding out the reasons customers buy from us or not.

Step 2. Clustering your notes with AI 🤖

The next step is to use AI to cluster your notes by what they mean in natural language.

The easiest way to do that is to create a vector embedding for each note by calling the OpenAI embeddings endpoint. A vector embedding is a numerical representation of your notes, which means you can measure how ‘close’ the meaning of two notes is by measuring their relative distance in the corresponding vector space.

→ If you don’t know what any of the above means, don’t worry. I’m happy to explain it to you in a video call 😎🤩 (availability permitting).

Once you have obtained a vector embedding for each of your notes, it’s time to analyze them. In this example we’ll use k-means clustering to group notes together, but know that there are hundreds of machine learning methods you could use for many different tasks, and there are many different techniques you could apply to get better insights. This is just a rough example, so you can get the gist of it.

K-Means clustering is a machine learning method that will divide your dataset into k clusters, where each data-point gets mapped to the cluster with the nearest mean. Practically speaking, this means you will obtain k different categories, where each category contains notes that somehow express a similar meaning.

To get useful insights, consider which is the best way to preprocess and clean your notes. You might want to analyze sentences or paragraphs instead of whole notes, if you want a more granular insight into which discussed topics lead to revenue over time, for example. Your final model will depend on your business model and the way your data is structured.

Step 3. Summarizing your findings 🦾

So far we’ve grouped notes with similar characteristics together and calculated a rough estimate of the average revenue for each group. Pretty useful, right!

Note that we do not need to tell our model beforehand which categories it has to sort everything into. This is very useful because you can get easily accessible insights into what works and what doesn’t, without having to assign individual categories to your notes by hand, which can be a tedious, time-consuming and error-prone undertaking.

In the context of sales, this approach allows you to find out which problems high-revenue clients are grappling with, so you can refocus your marketing efforts. If your notes are detailed enough, you can also find out which sales approach converts best. Pretty useful indeed.

And if you were really keen on a ChatGPT-based solution after doing all this, you can feed the clustered notes back into GPT with a framework like LangChain, so you can get a somewhat automatic analysis despite exceeding GPT-4's token limit.

Good luck on your data-driven journey!