Apply AI: Unleashing LLMs in your business

Discover Four Innovative Tactics to Propel Your Business Forward and Secure Your Advantage with Large Language Models

The AI revolution is here, and Large Language Models (LLMs) are leading the charge. Businesses that can harness and master these powerful AI systems will flourish, while those that can’t risk falling behind.

As someone who’s been implementing machine learning models since 2013, nothing I have seen so far has the transformative potential of LLMs. In this article, we’ll explore four game-changing tactics — Automated Sentiment Analysis, Personalized Marketing, Intelligent Chatbots and using LLMs for product development.

Unlike many other applications of AI in business, applying LLMs doesn’t necessarily require large amounts of pre-existing data. This makes the approaches outlined in this article applicable for a wide range of businesses. So let’s dive in and uncover how you can secure your advantage with LLMs!

Automated Sentiment Analysis

About 8 years ago I developed a bot that would read messages from a twitter feed and then upload those messages to a neural net for sentiment analysis. The system worked well, but the sentiment analysis wasn’t always the best and the error rate was significantly worse than a human. With the advent of chatbots like ChatGPT, detecting the sentiment of text has become much more accurate and easier to perform.

My original system was a prototype meant to be used by investment companies to automatically monitor sentiment on the companies in their portfolio, however automated sentiment analysis has several other use cases.

A few example use cases are listed below:

• Use customer sentiment to provide optimal adjustments to improve customer experience and satisfaction.
• Use the analysis as feedback for product development teams.
• Understand sentiment towards competitors to help provide strategic insights.

So how would one go about implementing such a system?

Firstly, identify what you want to track with the sentiment analysis. Depending on the overall objectives this could for example be sentiment of customer feedback.

Once objective is defined, we can start determining the optimal data feed. After identifying the data feed you can either connect to the data feed via an API, scrape the information source of your choosing or perform another data acquisition exercise.

When the data is obtained we would typically apply some preprocessing steps to prepare the data and then feed the text into an LLM. By creating the right prompt along with the text, we can give the LLM the needed context to analyze the text.

Once the data is analyzed we can execute an action based on the insight. For example, send a follow up mail to an unhappy customer or send a warning email to a customer service representative that could follow up on the event.

The key steps for automated sentiment analysis are outlined below:

1. Define Goals: Identify what you want to achieve with sentiment analysis.
2. Data Source Identification: Identify the optimal data feed.
3. Integrate data: Develop a system to automatically feed relevant text into the LLM.
4. Analyze: Develop a system to analyze the data.
5. Act: Execute on the actionable insights delivered by the system.

Personalized Marketing

Imagine a marketing strategy that speaks directly to each customer’s unique needs and preferences. That’s the power of personalized marketing. By utilizing LLMs, businesses can analyze customer data and generate unique marketing messages that resonate on a personal level.

Personalized marketing messages have can have multiple benefits:

• Improving customer engagement and conversion rates.
• Increasing customer loyalty and retention.
• Enhancing brand reputation by showing your customers you understand and value their unique needs.

While the benefits are obvious it is not necessarily trivial to implement a personalized marketing solution. I suggest to begin with integrating your customer data with the LLM. This could include demographic information, purchase history, browsing behavior, and any other relevant data.

The LLM can then analyze this data to understand each customer’s unique needs and preferences. Based on this understanding — and by crafting specific prompts — the LLM can generate personalized marketing messages for each customer.

After the system is initially implemented it needs to be monitored and tuned for the best performance. The key to successful personalized marketing lies in continuous testing and iteration.

The key steps to use a LLM for personalized marketing are outlined below:

1. Define Goals: Identify what you want to achieve with personalized marketing.
2. Integrate Data: Develop a system to feed customer data into the LLM and generate personalized marketing messages.
3. Test and Iterate: Test the effectiveness of the personalized marketing messages and iterate based on the results.

Intelligent Chatbots

Intelligent chatbots are currently revolutionizing customer service and the advent of LLMs have supercharged them! They are available 24/7, can handle a wide range of inquiries, provide instant responses and can free up human agents for more complex issues. While you previously needed vast amounts of data to develop them, the barriers to developing these AI systems have been greatly reduced.

The benefits of chatbots go beyond customer service. They can also drive sales, provide valuable customer insights, and scale to handle large volumes of inquiries. Not only that, in the long- to medium-term they are also a cost-effective solution that will help give a cost advantage. In addition, chatbots also makes it much easier and faster to translate strategy into concrete business to be acted upon by the the bots — rather than having to update a team of customer service reps.

When I have implemented machine learning products into the customer service pipeline previously, we almost always used a phased approach. This was done to mitigate risks and ensure continuous service.

Adopting a phased approach, the intelligent chatbots are first integrated into the customer service platform as assistants to the current support staff. This allows the team to familiarize themselves with the technology and provides an opportunity to fine-tune the system based on real-world interactions. In this first stage, the support staff still makes all the decisions and writes the replies, while the chatbot is only suggesting answers in the background.

In addition to current requests, other information about the customer should also be fed into the chatbot to provide extra context. This includes a lot of the same information used for personalized marketing, and makes the chatbot able to produce more personalized and unique responses. While the system is generating replies, the customers service agents should generate feedback on the replies and rank them, so we can improve the system in the second iteration.

Once the system has been optimized and feedback has been incorporated into the system we can transition into the second phase. Now, the customer service agent oversees the responses generated by the LLM and these are the default response. In this second stage, the chatbot is writing the content and the agents only intervene if the bots provide unsatisfactory replies.

Finally, once confidence in the system is established, transition to full automation, with the LLM handling customer inquiries independently and then only escalating to a real customer service employee in the most complex cases.

The key steps to implement a chatbot are outlined below:

1. Portal Implementation: Implement the chatbot into the customer service portal
2. Provide Data Feed: Feed relevant customer information and data into the chatbot
3. Initial Responses: Run the chatbot along the current customer service process
4. Supervised Mode: Let the chatbot provide the customer service response while being supervised by the customer representative
5. Fully Automated: Automate the system and handle only the most complex cases with a customer service agent.

Product Development

You might have heard about LLMs hallucinating or inventing ‘facts’ that are not true. In some cases this can of course be problematic, but when it comes to areas where we want to be creative — like product development — this is a feature and not a con of the LLMs. By combining ideas, and using different styles and smart prompts we can use the LLMs to constantly evolve and improve our product offering.

For example, let’s say you have a log of customer feedback. Feed this into the LLM to provide context, and then also feed it your company’s current strategy and product offering. Then get the LLM to generate novel products based on the feedback and your background information.

Or similarly, assume you have an idea of a product but would like to explore other possibilities. Speed up the ideation process by using the LLM to generate similar product ideas.

Perhaps you are producing code for a new project, the LLMs can help to write the code and speed up the development process significantly. Try the plugins available on ChatGPT https://openai.com/blog/chatgpt-plugins . While this plug is currently only available in Python it’s only a matter of time before more language support is added.

Or maybe you need to design a new product and are unsure what it should look like? You can combine your ideas for the product design, together with your company’s strategy and the competitive landscape and then have LLM blend that information and use it to create prompts for image generating tool like Midjourney. The following example illustrates how this can work in practice. I was recently working on a new logo for my company, so I fed the LLM information about my service offerings and my core values and asked the LLM to return 10 different ideas for logos.

Two of the suggestions are given below:

“Futuristic key in cool gray and electric blue”: A key designed with a futuristic or digital aesthetic in cool gray and electric blue, symbolizing unlocking potential through digital transformation and data science.

“Innovative light bulb in vibrant teal and cool gray”: A light bulb combined with a gear or circuit board in vibrant teal and cool gray, symbolizing innovative ideas in the realm of data science and digital transformation.

Now, tweaking these slightly to allow for a logo on white background gave the following results with Midjourney:

When it comes to product development and the use of LLMs, only the imagination sets the boundaries. The LLMs ability to merge multiple concepts and ingest vast amounts of information makes them perfectly suited as assistants in product development.

Conclusion

In this article we have explored the possibilities that Large Language Models (LLMs) offer businesses to drive growth, enhance efficiency and ensure customer satisfaction. The potential is quite significant and spans from marketing strategies to the development of intelligent chatbots.

However, it is important to remember that the exact details of an LLM implementation will depend on a few factors, specifically its interface and whether it’s a proprietary or open source solution. These factors will influence how you engineer prompts and determine whether fine tuning the model is feasible.