NLP and GPT-4 applications in Sales
The rapid advancements in AI and Large Language Models like OpenAI’s GPT-4 and Google’s Bard, have made it possible to extract valuable insights from large corpora of data. These capabilities can be used to make better customer interactions and optimize sales processes. By leveraging Language Models and custom NLP pipelines, businesses can analyze vast amounts of data, such as customer conversations and interactions, to gain valuable insights. This enables sales teams to understand customer needs, anticipate intent, and personalize the sales experience.
In this article, you will learn how LLMs like GPT and NLP can help you boost your sales.
What is NLP?
NLP, or Natural Language Processing, is a branch of artificial intelligence that enables machines to understand and interpret human language. It encompasses a set of algorithms and techniques that allows computer programs to comprehend, analyze, and generate natural language text or speech.
Large Language models are a part of NLP where we train massive neural networks to understand and perform tasks with human language. There has been a lot of development in LLMs recently.
By leveraging advanced algorithms and machine learning models, NLP enables businesses to extract meaningful insights from vast amounts of textual data, empowering them to make data-driven decisions and drive positive outcomes.
How can NLP help Sales processes?
As mentioned before, NLP pipelines allow computers to understand human language better. Computers can analyze a ton of information and extract key information from it. This allows sales teams to identify and understand customer needs more effectively and faster.
NLP algorithms and LLMs can help in understanding customer intent, going beyond the literal interpretation of their words. By analyzing language patterns, sentiment, and context, NLP algorithms can discern the underlying meaning and emotions behind customer messages. This understanding of customer intent empowers sales teams to provide more relevant, targeted, and personalized responses, enhancing the overall customer experience and increasing the likelihood of successful sales engagements.
Here are some ways NLP can help sales:
Identifying customer needs
NLP models are really good at understanding language and inferring meaning. Even early language models like BERT, and GPT-2 can be used to extract relevant information from online reviews, surveys, comments, and feedback forms. This data can help the sales and marketing teams to understand what customers want and what are the issues they are facing.
Larger models like GPT-4 and Claude can perform these tasks out of the box with zero-shot capabilities.
Understanding customer intent
Language models can decipher customer intent by analyzing their language patterns, sentiment, and context. This data can be paired with their browsing behavior and pattern to understand what they are looking for and why.
Understanding “why” or intent behind the purchase is very valuable information as it can help understand user behavior and needs. This can lead to selling more products to the same customer and giving better suggestions too.
Personalizing the sales experience
Sales outreach is often done in a “one-to-many” way. A single piece of text, like a cold email, is taken and sent to a list of ten thousand prospects. This is not very effective.
AI can help you personalize your sales message for every prospect. LLMs can take some information and generate or rephrase some text based on that information. You can use this to write a unique cold email for every prospect in no time.
There are SaaS tools out there that help you personalize your cold emails by generating “personalized first lines”. This approach has shown great success and seems to significantly increase open and reply rates. Studies have shown that better first lines can increase open rates by 50%.
Measuring the effectiveness of sales campaigns
Once a campaign is launched the data collection becomes crucial to make sure the sales and marketing teams learn what they did right and what they did wrong. Classic NLP techniques like sentiment analysis can be used to understand qualitative data better and learn from it.
You can collect all the campaign replies and feedback, and give it to a Langauge model like GPT4 and ask it to analyze and find patterns in it. This data can come from cold email replies, sales email threads, and even from social media campaigns.
How GPT and NLP can Help Sales
Now let's do a deep dive into use cases of NLP in sales and how it can help improve sales processes and overall performance.
Chatbot for Sales Assistance
Buying things online is great, but as online marketplaces grow, finding exactly what you are looking for is getting harder and harder. Search functionality returns thousands if not hundreds of similar products for one single search. This leads to Search Abandonment by the customer.
A survey by Google showed that 77% of users avoid sites where they have experienced difficulties with search. They also showed that $300B is lost every year just because of bad search experiences.
This is where AI comes in. GPT-4 or ChatGPT-based chatbots can assist users in narrowing down the search and finding exactly what they are looking for.
Here’s an example of a chatbot we built to assist buyers in their buying journey.
You can see that the chatbot is capable of asking follow-up questions and collecting relevant information required to accurately search for the product.
Once the chatbot has collected enough information, it will use the search functionality to find the product and return it:
We can have a Python pipeline to replace the search(...) with actual product images and links. This allows us to integrate search and product browsing features directly into the chat experience. We can also ask ChatGPT to extract information in a JSON format which then can be used to implement as filters on the search itself.
How to build a Sales Assistance Chatbot?
A good sales/search chatbot has to be able to be good at certain tasks. It has to be able to ask for more information and should be able to form search queries to look up products. It should also have access to product inventory and should be able to show information from the product database right in the chat for ease of use. These are the key capabilities of a good chatbot:
- Ask follow-up questions: Asking follow-up questions is necessary for knowledge gathering and forming a proper search query. It helps the model understand what the user wants and execute a better search query. In the example shown above, you can see it asked multiple questions before using the search(...) function.
- Form search queries: After gathering the information, the chatbot should be able to form proper search queries to look up the products in the inventory database.
- Access to Product Inventory: After forming the search query, Chatbot should be able to access the products returned and show them in a proper manner. It should also be able to take the search further if too many products are returned.
At Mercity AI we have built our pipelines and techniques to tackle all these problems. Here is our pipeline:
Let’s go through all these components step by step.
Chatbot
Chatbot is a large language model here. It can be GPT3 finetuned for chatting purposes, or ChatGPT GPT-4. Chat models like GPT3.5 and GPT-4 perform better as they have been pretty much trained to follow a chat-like structure.
Chatbot here is responsible for interacting with the user and gathering information about what kind of product the user is looking for. And after that, writing a search query to search for the relevant products and return them to the user.
Here we use a modified version of ReACT prompting to give the model search capabilities. Here’s an example of original ReACT prompting:
Here you can see the model thinks and uses the search function and search results to give a better answer to the user.
Embedding Generator
This component is responsible for taking all the pre-existing product data and converting it into embedding vectors.
Embeddings are numerical representations of textual data, these are learned and used by AI models. These vectors are responsible for encoding and learning the semantic meanings of the text. Because these vectors “carry” meaning, they have a property: Texts with similar meanings or semantics will be located closer to each other in the vector space than the texts with different meanings. This property allows us to use these embeddings for semantic search.
We can use OpenAI’s embeddings or Sentence Transformers for embeddings.
Embedding Search Pipeline
This component takes search queries from the chatbot and finds the relevant products for it.
It works by converting the search query into embedding vectors and then using cosine similarity to find the most similar product embeddings. These product embeddings can then be tracked back to the original products, which then can be returned to the user via the chatbot.
This graph by OpenAI shows how sentences with similar meanings and contexts are placed together and texts with different meanings are placed farther away.
Vector Database
Vector Database here is used to store the embeddings generated by the Embedding Generator. It is responsible for the quick retrieval of these embeddings when required.
Some popular choices of vector databases are Pinecone, Milvus, and Weaviate.
Advantages of using a Sales Chatbot
As presented here, a sales chatbot can greatly help with product search and can make it easy for buyers to interact with the platform. They can answer questions, make recommendations, and complete transactions.
Chatbots are becoming increasingly popular as a way to improve the customer shopping experience. They offer a number of benefits for both customers and businesses.
Visual Product Search
Visual Product Search has emerged as a game-changing technology in the realm of e-commerce and retail. With the rise of smartphones and advancements in computer vision, consumers now have the ability to search and discover products simply by capturing or uploading images. This innovative approach allows users to bypass traditional keyword searches and instead rely on visual cues to find exactly what they are looking for.
Google lens is the best example of this feature being used in the wild. One can simply take an image of a dress and find it online amongst millions of products. All under a few seconds.
Previously, if you like a dress or piece of furniture, you would have to ask or spend hours online trying to find the same product you saw. But google lens can take care of this end to end.
Let’s see how can you build a visual search AI for products on your own. You can see our more comprehensive coding guide to build your visual product search engine here.
How to Build a Visual Product Search Pipeline
A pipeline to search for similar products is not very difficult to build. So much product matching depends on the data the AI is trained on. The goal of the AI here is not to do a specific task, but to learn the similarities and differences between product images. This AI model can then be used to find products similar to the one passed by the user.
To be more specific, we will use the AI model to get embeddings from it and then use them for similarity matching, just as we used in the chatbot pipeline to find relevant products from search queries.
The pipeline is not very different from the data retrieval pipeline in the chatbot section:
Only a very few things have changed here, mainly the model and how we feed the data has changed. This is because we are using the CLIP model here, it is a special model that learns to combine both text and images in the same embedding space. Let’s take a deeper look at how the CLIP model can be used for this.
CLIP Model
CLIP is a model by OpenAI. It has been trained specifically to learn to put together similar images AND their text descriptions closer together in the embedding space. Meaning that CLIP embedding space will understand the textual meaning of the images, as the texts which can describe the image will be located closer to the image itself.
This simple capability of the model allows us to create embedding spaces which can find similar products with very high accuracy. This works not only for image search but also for text search as the CLIP model learns with both image and text.
But this is not where it ends, CLIP model can be further finetuned on custom datasets to learn more about a task specifically. This is perfect for our task.
In this paper, researchers trained a CLIP model to work specifically on clothes datasets. They trained the model on clothes along with their text descriptions:
The CLIP model slowly learned to put together the image and its descriptions together in the latent space.
You can check out FashionCLIP here.
Alternatively, MetaAI has also released a model similar to CLIP, but it can combine much more data than just text and images. This is ImageBind. One can also use this model to achieve similar capabilities as CLIP.
Attribute Extraction from Product Images
There are not only search but also tagging capabilities that are highly desirable from tools like these:
Attribute extraction allows us to understand the product better, and as a result, understand the customer better. This allows us to create better recommendations, provide better results, and even better search capabilities.
This data can be further used to search or even improve customer recommendations on an e-commerce platform. Some companies offering such capabilities are: Pixyle.ai, and Visenze.
Why use Visual Product Search?
Visual Product Search offers a ton of benefits that make it a valuable tool for both businesses and consumers.
- Enhanced User Experience: Visual Product Search simplifies the search process by allowing users to find products simply by uploading or capturing an image. This intuitive approach eliminates the need for keyword searches and provides a more seamless and user-friendly experience. This leads to increased customer satisfaction.
- Improved Product Discovery: Visual Search expands the possibilities of product discovery by enabling users to explore visually similar items that they may not have found through traditional search methods. This significantly decreases search abandonment rates.
- Competitive Advantage: By implementing Visual Product Search, businesses gain a competitive edge in the crowded e-commerce landscape. Providing an innovative and seamless search experience sets them apart from competitors and helps attract and retain customers.
- Better Upsells: Once we have a base idea of what the user is looking for, we can sell other products or accessories to go with the product. This enables businesses to enhance upselling opportunities by offering relevant and appealing add-on products, thereby increasing the average order value and maximizing revenue potential.
There have been studies that show that around 62% of millennials and Gen-Z want Visual search capabilities. Google has also been working on using more and more visual elements in its search platform. And using visual search simply allows us to shop better and provide a better experience to the customers.
Summarization of Sales Calls
Sales calls can become increasingly lengthy and time-consuming, posing challenges for sales teams to effectively utilize the insights gained from these interactions. By employing automated summarization techniques, businesses can condense lengthy sales calls into concise and meaningful summaries, extracting key points, important details, and actionable insights.
Source: Firefly.ai
Call summarization software can simplify the process for sales teams by saving time and resources. Instead of manually reviewing lengthy sales call recordings, the software condenses the most important information, allowing sales professionals to quickly access key insights. This efficiency lets teams focus on strategic tasks like follow-ups and lead nurturing. Summaries can be easily shared among team members, promoting collaboration and consistent messaging.
Here we can use GPT3 for the summarization of these sales calls. By providing it with the text or recordings of these calls, GPT-3 can analyze and understand the conversations, picking out the important information. It then generates short summaries that capture the main points, giving sales teams a clear understanding of the conversation and helping them make better decisions.
Here we asked GPT3 to summarize an hour-long call. We told it to give a quick summary with essential points of the conversation. You can see that GPT not only was able to compress the information in the call, but also able to extract the important points with exact figures from the call.
How to summarize very long transcripts?
To summarize long transcripts or text using GPT3, you can use a chunking strategy. You can split your long transcript into smaller chunks, and then feed those chunks into GPT3 one by one, asking it to summarize it. Once you have the summaries of the chunks, you can prompt GPT3 to combine all those summaries into one.
We use a chunking pipeline to create chunks of the passed text. One important property of this pipeline is that there always should be some overlap between the beginning and end of the consequent chunks. This ensures that there is no data loss between the chunks.
Once we have the chunks, we can then feed them to GPT3 or GPT4 asking it to summarize and put them all together.
Sales Call Analytics
Summarization is not the only thing you might want out of your sales calls. There is so much data condensed that it can be hard to put it in a summary. We can ask GPT-3 or GPT-4 to perform all kinds of tasks. Here are some examples:
Extract Questions, Topics, and Tasks from Calls
GPT and other models can be used to extract important information from call transcripts. All we need is a prompt to pass the call transcript. And if the transcript doesn’t fit the context length limits, we can do chunking and combining here just as we did in the summarization process.
Review Sales Performance
GPT3 and GPT-4 can be used to give valuable reviews and gain insights from sales calls. You can put in your transcript and ask GPT3 to give a “detailed analysis” of the salesman and it will.
This capability can be used to evaluate large sales teams and improve their performance as a result. Salespeople can also practice selling with ChatGPT to improve themselves.
Sentiment Analysis
GPT models can also be used to do sentiment analysis based on what the people on the calls said. Here’s how you do it:
The prompt here has been modified to output one single label for every speaker. We can ask it to answer in a more detailed manner too:
Sales Message Writing Assistant
In the competitive world of sales, the ability to communicate persuasively and convey the right message is crucial for success. A sales message writing assistant can be really powerful in optimizing sales communication. Such an assistant can provide real-time suggestions, improvements, and guidance to craft compelling messages. It helps professionals improve their messaging by optimizing the choice of words, tone, and structure, enabling them to create persuasive messages that resonate with potential customers.
How can an AI Sales Assitant help?
Let’s go over some ways AI can help you improve your sales.
Improve your Sales messages
An AI assistant can help you improve your sales messages and interactions with your customers by improving your language. Here is how you do that:
Here we can see that without a very complex prompt, GPT4 is able to rephrase the message with better language.
You can check out our detailed guide on how to build a custom model to rewrite sales messages for simplicity and more persuasion.
Recommending Cross and Upsells
Having an AI assistant while you go on your sales calls or interact with customers over text can help you recommend what products you can sell them. More importantly, it can suggest you right upsells and cross-sells to maximize the sales potential.
Here we gave GPT4 a sales conversation and asked it what products from the product list can be upsold to the customer. This can be integrated directly into the sales platform to provide suggestions in real time.
Extract Valuable Data
GPT4 can be used to extract valuable information from sales conversations, here’s how:
Here we have used a very weak prompt to extract dynamic information from a sales conversation. We can use similar prompts to extract specific information from a conversation. This can be customer data, user behavior, user preferences, etc.
Key data extraction from Sales Documentation
AI methods can also be used to extract information from massive documents. This is different from summarization. Summarization is dynamic by nature, here we can extract specific data and work with that. This can be used to analyze specific bits or to extract specific information from long documents, calls, email chains, etc.
The extracted information can then be used to maintain CRMs and improve sales performance. This also helps in maintaining information across different mediums.
Extracting data from Chat Conversations and Emails
Conversation analytics can be a huge help in understanding customer interactions and extracting valuable insights. AI techniques can be employed to analyze chat conversations and emails, extracting important data such as customer preferences, sentiment analysis, frequently asked questions, and emerging trends. This data collected over a long period of time and for a big user base can prove really valuable.
Here’s how you extract key data from chat conversations:
Here we use GPT3 to extract valuable info and specific data from an internet chat between a customer and a salesperson. We have extracted names, objects of interest, and specific reasons for the customer to not buy. This is valuable information for the sales agents on the field who interact with the customers face to face. This information can be used to follow up and do targeted outreach toward this customer.
Sentiment Analysis on Customer Feedback
By leveraging advanced natural language processing algorithms, businesses can uncover customer preferences, sentiment analysis frequently asked questions, and emerging trends from these interactions. This data can be used to understand what customers want and what their major issues are and solve them accordingly.
Here we asked GPT4 to analyze the past reviews and it returned a quick summary of how customers are reacting to the product. It also returns information about the issues, this can be used to make the product better.
Personalized Sales Outreach with GPT4
GPT-4 can enhance personalized sales outreach by using its advanced language processing capabilities. Businesses can create customized messages, emails, and pitches that resonate with each prospect. By understanding customer needs and preferences, GPT-4 empowers sales teams to deliver targeted and persuasive communications. Using GPT-4 for personalized outreach can boost customer engagement, build trust, and drive sales growth.
Moreover, GPT-4 can help sales teams save time and increase efficiency. Automated processes can be used to generate personalized messages quickly and accurately. This can help sales teams reach more prospects in less time, resulting in higher conversion rates. GPT-4 can be integrated into existing sales pipelines to help sales teams maximize their outreach efforts and drive better results.
Personalizing Emails using GPT4
As mentioned before, people often use AI to personalize the first lines of emails for better open rates. But what if you could use AI to personalize the entire email? GPT-4 can help you do that. GPT-4 can generate emails that are tailored to each individual customer, based on their preferences and needs. This can help sales teams build trust and increase engagement with their prospects.
Here we are going to generate cold emails for a customer based on their reviews left on the website for a specific product. We pass the review left by the customer to GPT and ask it to write an email specific to the issue faced by them. This can help bring lost customers back to the platform.
Prompts can be further modified to generate more specific emails and responses.
We can use the same technique to generate a text message using GPT4.
Similarly, GPT4 can also be used to generate newsletters and other marketing material which can be targeted specifically to individual customers or customer segments.
Enhance Sales Performance with NLP and Cutting-Edge Language Models
Discover the incredible potential of NLP-powered large language models like GPT-3, GPT-4, Claude, and LlaMA in revolutionizing your sales strategies. With their advanced language processing capabilities, these models can help you optimize customer interactions, drive conversions, and boost sales growth. Reach out to us today and unlock the full power of NLP in your sales journey.