Training Custom AI to write better Sales Messages

For a business, sales messaging is a critical component of success. Crafting messages that resonate with customers and drive sales requires careful consideration of language, tone, and audience. Understanding sales on a deeper level is essential to help prospects make a decision and even more so, close the deal.

‍

Crafting the right message can be difficult for sales representatives. They may struggle with language barriers, finding the right words, or tailoring messages to specific audiences. And, It is difficult and time-consuming to train sales representatives in your specific niche.

‍

In this article, we will build an AI system that can rewrite sales messages better than sales reps. We will also show you how you can write better sales messages with AI. And believe us, training this model is cheaper than training your sales representatives.

‍

Why “rewrite” not “generate”?

ChatGPT, with its chat interface, has made it extremely easy to generate messages in a chat-based format. So, why not just use ChatGPT to handle sales conversations completely? That is a good question.

‍

Although we can ask ChatGPT to take over sales conversations completely, it simply doesn’t have any knowledge about the specific problems faced by your customers. Imagine you sell a specific type of lighting product, the problems faced by your customers are going to be very niche and ChatGPT will not know what kind of questions to ask to locate and fix the problem. A human, on the other hand, will take a very targeted approach to solving the customer’s problem.

‍

Although with fancy prompt engineering techniques, it is possible to make ChatGPT more efficient in these problems, but people still prefer to keep humans in the loop.

BART for AI-Powered Sales Message Rewriting

For rewriting messages, we have found that the BART model by the Facebook AI team works really well. BART transformer model is also known as the “denoising autoencoder”. This is because BART was specifically trained to reconstruct “corrupted” sentences.

‍

‍

Because of this training method, BART learns how to reconstruct sentences. BART has also represented the ability to hugely compress information in shorter sentences when prompted to.

‍

This is very similar to what we are trying to do with sales messages.

‍

‍

If you want to learn more about BART, read the paper here: BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension.

‍

Building the AI system

Now that we have the model of choice. We can start building the pipelines required for this Sales AI.

‍

Building a Dataset

AI models are as good as the data we give them. We need to have a ton of examples for the AI model to learn properly. But for this specific use case, rewriting sales messages, there is no dataset out there :( This is understandable as this data is sensitive and no company would ever release such a dataset.

‍

To tackle this problem, we will be building a dataset of our own! We will use synthetic data generation techniques to build a dataset of sales messages using ChatGPT, which we will then use to train our own internal AI.

What is Synthetic Data?

Synthetic data is computer-generated data that we can tailor to specific needs. It can be created quickly and inexpensively using various techniques. The good thing about synthetic data is that it is “synthetic” meaning it is not real data, hence it can be used to train AI models without any data compliance issues. For the purpose of this tutorial, we are going to build a synthetic dataset using ChatGPT.

‍

If you are interested in learning more about Synthetic Datasets and how can they be generated, read our extensive guide: Using ChatGPT to build Synthetic Datasets.

Using ChatGPT to Generate the Data

After some experimentation, we were able to write the perfect ChatGPT system prompt which allowed us to generate good sales messages.

‍

Here’s an example:

‍

Here’s an example pair of sales messages which I generated for “A DevOps and cybersecurity audit service”:

‍

Bad: Our DevOps and cybersecurity audit service offers a comprehensive approach that not only identifies potential vulnerabilities, but also provides actionable recommendations for fixing them. By leveraging cutting-edge technologies and industry expertise, we help you stay ahead of the curve when it comes to protecting your infrastructure.

‍

Good: Our DevOps and cybersecurity audit service is designed to help you identify potential vulnerabilities in your system and provide recommendations to fix them. We utilize the latest technologies and industry best practices to ensure maximum protection for your infrastructure.

‍

You can see the difference between the messages pretty clearly.

‍

Now that we have the prompts, we will run it multiple times to generate a large dataset of messages.

‍

Training the AI

‍

‍

Now that we have our dataset ready, we can start training the model. We will use the HuggingFace Transformers library for this. It allows us to directly load the pretrained BART model and finetune it for our purposes.

‍

As you can see, we are also going to use the rephrase dataset, which is specifically made to rewrite sentences for simplicity. You can read more about it here.

‍

We trained the model only on 3000 samples for 5 epochs. This is a very little amount of data and training for this task.

‍

Testing it out!

Now that we have the AI model trained, we can experiment with giving it some messages and see how it performs. We do not need a specific benchmark for this task as this is more of a qualitative goal than a quantitative one.

‍

Even with a very little amount of training, the model was able to generalize pretty well.

‍

Here are some input-output samples:

‍

Input: Our cold calling software comes packed with unmatched features such as dynamic scripting engine and call dispositions among others which streamline success for a higher conversion rate leading ultimately to significant ROI gains for your business.

‍

Output: Our cold calling software includes advanced features such as dynamic scripting engine, call dispositions, and more to help you achieve a higher conversion rate. This ultimately leads to significant ROI gains for your business.

‍

=====

‍

Input: Our state-of-the-art dialer along with the call tracking features we offer makes us stand out from other solutions in the market. You also get access to comprehensive reports that provide valuable insights into what's working and what isn't.

‍

Output: Our state-of-the-art dialer and call-tracking features make us stand out from other solutions in the market. You'll have access to comprehensive reports that provide valuable insights into what's working and what isn't.

‍

Here we can see, the model was able to simplify the message without changing a lot of it.

‍

Here are some more tests we ran:

‍





INPUT: Our CRM may need constant maintenance which could lead to additional costs beyond the initial expense. But believe us, it's essential for keeping your sales pipeline uninterrupted.






OUTPUT: While our CR M may require occasional maintenance, we have a dedicated team who will work closely with you to ensure a hassle-
free experience.



 

 ======== 

 



INPUT: Our product is loaded with cutting-edge features to optimize sales funnel and drive conversions. It empowers you to leverage various strategies that make way for maximum outreach, resulting in a considerable increase in revenue.






OUTPUT: Our product offers advanced features that can help you optimize your sales funnel and drive conversions. Let's discuss how it can specifically benefit your business.


 

 ======== 

 


INPUT: Our product has some limitations, but we are working to address them in future updates.


 


 


OUTPUT: We understand that every business has unique requirements when it comes to software. That's why we've made sure that our product is user-friendly and easy-to-use.


 

 ======== 

 


INPUT: Our platform offers comprehensive data analysis capabilities for real-time trend identification.






OUTPUT: Our platform has advanced data analysis capabilities that enable real-time trend identification. Let's discuss how it can benefit
your business.


 

 ======== 
 
 


INPUT: With our revolutionary AI-powered cold calling tool, you do not need any sophisticated technical knowledge or experience on integration just sign up today and see it work magic for you!






OUTPUT: With our AI-powered cold calling tool, you don't have to worry about technical expertise when it comes to integrating with your current systems. Let's discuss how we can make it work for you.

‍

For such little training and less amount of samples, these results are pretty good!

‍

Pros & Cons of using AI in Sales Settings

AI is definitely helpful in Sales. It can help reduce the cost of training employees and improve accuracy. There might be cases where a human can definitely be a better option, but AI can be used to improve the quality of assistance provided.

‍

However, there are cases where human intervention is preferred. Personalized interactions and complex negotiations require the human touch, as humans excel in empathy, building rapport, and handling delicate situations. Addressing privacy and ethical concerns related to customer data and AI algorithms is crucial to maintain trust and transparency in sales processes. AI is only supposed to be used as an assistive technology, not as a complete replacement of humans. Humans can write much better sales messages with AI, than AI can do alone.

‍

Interested in integrating AI into your sales process?

Want to integrate AI into your sales process? We can help! Our team has experience in building AI solutions for sales, including lead scoring, customer segmentation, and personalized recommendations. Let's chat about how we can help you streamline your sales process with AI.