By now, you’ve doubtlessly heard the buzz surrounding Artificial Intelligence (AI) and its potential to revolutionize each and every industry. We’re talking about improved efficiency, heightened customer satisfaction, increased sales, decreased costs and more. Sounds fantastic, right? But, to realize those goals, you really need to have an AI application that uses your proprietary business data.

I help businesses build growth strategies that work, and AI is the hot topic right now. One of the most frequent questions I get asked is, “How can I implement AI in my business using my own data?” If you have the same question, fear not. I’m here to help you get started.

This article is intended for a wide business audience. To keep things accessible, I’ll explain technical concepts at a high level and focus on the fundamentals. However, to keep the article concise, I’ll assume you already have some understanding of the key business benefits of AI. If this is all new to you and you’re unsure if AI is relevant to your business (spoiler: it is), I recommend reading my earlier post for a quick introduction.

Types of AI

Before we get in too deep, it’s worth taking a step back to understand what Artificial Intelligence (AI) is. Simply put, AI is the ability of a computer system to learn and make decisions from data and experience, mimicking the cognitive processes of humans. Before AI, computer systems could only process a strict set of instructions that were defined by a human programmer (or more likely, a team of them). You can think of traditional computer software as tools in a workshop. Businesses collect huge sets of these tools, some simple and some complex. They have done a remarkable job of improving the way we work, but they can be rigid, complex, and expensive to install and maintain. AI, by comparison, is not just another tool. It’s a multi-talented workshop apprentice who is keen to learn and gets better by the day. It can use the tools you already have, and it can invent new ones all on its own.

There are numerous types of AI, and what suits your business may not be the cutting-edge headline-grabbing systems that aim to emulate Shakespeare or Van Gogh. For instance, machine learning (ML) is an application of AI that provides systems with the ability to learn and improve from experience without being explicitly programmed. Think autonomous vehicles and the like. Computer vision is another AI technology, used in applications such as security, medical screening, product recognition, and more. Natural language processing (NLP) is an application of AI that understands human language and is found in applications such as search engines, customer service chatbots, and the predictive text on your smartphone.

There are other types of AI too, but the type we will focus on in this article is NLP. Because of its ability to read and “understand” language, NLP is very well suited to interacting directly with humans. It’s also a natural choice for many business applications because businesses use language to store much of their data (manuals, emails, business plans, processes, websites, etc.) Likewise, the internet as a whole uses language, so an NLP system can digest a wealth of knowledge.

Large Language Models (LLMs)

The type of AI system that is, quite rightly, getting most of the press right now is a particular type of NLP system called a Large Language Model (LLM). The most famous of which is undoubtedly ChatGPT by OpenAI. There are many other types of NLP system, but to keep things simple, we’ll just focus on LLMs today.

An LLM is a type of NLP-based AI system that is capable of understanding large volumes of text. LLMs are trained on large datasets of natural language (often millions of words) to learn the nuances of a language, helping them understand and generate more meaningful and accurate responses. This makes them well-suited to responding to customers in natural language, as well as to other tasks such as sentiment analysis, text summarization, and document classification.

A simplistic way to consider how an LLM works is that it predicts the most likely words that will follow a given number of prior words. It can do this because it has seen similar phrases before during its training, so it uses probability to essentially guess what comes next. Of course, context is also important, so the LLM model looks back far enough to ascertain a basic understanding of what is being said. If this process is repeated iteratively, what results is a full response that gives the illusion of intelligence. Take the following example, produced using ChatGPT. I gave the model a prompt, and it generated a response that should seem familiar:

We got this response because ChatGPT has seen many examples of what follows the partial sentence in the prompt. But watch what happens when we modify the prompt and use a partial sentence that the model likely has not seen before:

In this case, the model appears to have just made some stuff up. In AI parlance, it hallucinated a response. To be fair, we gave the model no further information, and this is what LLM models do when we are vague. Unless we say things that violate some internal guardrail rule:

These are just very simple examples to help you understand the basics of how an LLM provides a response. It can provide some pretty amazing and useful responses if we are careful about what we ask and how we ask it. For example, if we ask ChatGPT to summarize the key news events in 2020, it will do so concisely and eloquently. We can therefore use the model directly as a research tool. However, if we ask it to summarize the news events in 2022, it will tell us that it can’t because its training ended in September 2021.

Perhaps you can spot a couple of key issues here? One is that the knowledge in ChatGPT is (currently) static, and 2021 was some time ago now. The other is that it only knows what was in its training dataset. That is still huge, by the way. The amount of data and time used to train ChatGPT 4 is a closely guarded secret, but it is safe to say that it is likely to have used hundreds of billions of words and taken several months. But, it would not have any knowledge of the internal workings of your business, such as customer records, stock movements, HR policies and so on. Those data were not publicly accessible, so the LLM could not have been trained on them.

So, returning to the core purpose of this blog, how do we implement an AI solution that can provide knowledge and ideas about our business, and be sure that the knowledge is current? In the following sections, I will suggest some methods. Please note that the pace of development in AI is astonishing. Every day, new AI tools and methodologies are being published, so this list won’t stay current for long!

Build Your Own LLM

One option is to build your own LLM, perhaps using one of the growing number of open-source LLMs. This would allow you to customize the model to perfectly fit your specific needs, as well as to have full control over the data it is trained with. You could ensure that the LLM is regularly updated with new customer data and other business-related information that your business could benefit from.

However, this would likely be a costly and time-consuming process. Building, training and managing an LLM requires a lot of resources and expertise, and AI data scientists and developers are not cheap. That said, resources for LLM DIY’ers are starting to appear which lower the bar for setting up and maintaining an LLM. For instance, Hugging Face hosts open source LLM models and datasets, and Replicate runs open-source LLMs in the cloud. Many of these have been fine-tuned to add additional domain-specific knowledge such as genetics or car manufacturing. You pay according to the model, computing time and processing power used.

For now, building and training your own LLM is an expensive luxury best reserved for larger enterprises or those with deep technical expertise.

Retrieval Augmented Generation (RAG)

A solution that is certainly within reach of any company is to use a publicly available LLM, such as ChatGPT, and to use a technique called Retrieval Augmented Generation (RAG) to augment the LLM with their company’s proprietary data. This sounds complex, but it is quite a simple concept.

Using RAG, you can leverage ChatGPT’s exceptional multi-lingual conversational capabilities to comprehend your prompt, consider both the information on which it was trained plus the additional data you have provided, then respond in natural language. Your proprietary data is available only to you and does not form part of the LLM’s training. The technique utilizes a server that you control, which runs specialist (likely open-source) software and a database that houses your company data (more accurately, a specially prepared subset of it). What happens under the hood is that when your prompt is sent to the LLM, a small but relevant subset of your proprietary data is sent along with it. For this reason, you should take measures to assure data privacy and security.

Implementing RAG still takes some planning and work, but you can probably achieve it using your current IT team and/or with some short-term assistance from an external consulting firm. One very important consideration is that the data in your database should be properly cleaned and prepared. That not only means removing any incorrect or out-of-date data but also any superfluous data (eg. citations, personal notes, etc.) that could both confuse the LLM and cause wasted processing costs.

Without getting too deep into technical detail here, your server might run a software library called langchain and use a vector database. This allows your IT team/partner to develop a relatively low-code custom solution that uses your data with a public LLM. No-code solutions are also appearing, such as Flowise and Langflow, both of which are user interfaces (UI’s) for langchain.

Function Calling

On June 13, 2023, OpenAI released an update to ChatGPT called function calling. This is a tremendously powerful addition and is one of the easiest ways for any business to make its own data available via an LLM.

Function calling allows the LLM to request additional help from an external system whenever it needs it. So, for instance, you could connect your inventory management system via one of these function calls. Then, when the LLM is asked to provide some insights into your inventory (eg. “compare widget shipments last month to the same month in previous years”), it will cause the necessary information to be retrieved from your inventory management system, which is then further analyzed by the LLM to provide a natural language response.

This is different from using RAG. With function calling, we do lose the ability for the LLM to directly analyze the data contextually and predict answers using it. However, this might not be a big deal, depending on the data in question. We can still converse with the LLM in natural language and the data will always be up-to-date (assuming the system you connect is current). This makes sense for things like inventory data and customer records, which change constantly. But language-rich information such as an employee handbook would be better integrated using RAG.

The really clever thing is that you can connect multiple systems via function calling, each with different purposes, and ChatGPT will automatically work out from the context which function to call. For example, say we had two inventory systems, one for shipments from London and another from New York. We just tell ChatGPT that for questions relating to inventory in London use the London function, and for New York use the New York function. And we do that using natural language when the system is configured — no coding needed.

As with RAG, function calling requires software to run on a server under your control, and again you can use langchain and its UI counterparts to do this. Combining RAG with function calling is a very powerful and relatively low-effort way to use AI in your business and one that I would recommend over building your own LLM. At least, as a first step into the bold world of AI.

AI Plugins and Integrated AI

The final method that I’ll discuss today is using AI technology that has been integrated into software that you already use, or is available via a plugin. This is probably the simplest, but not necessarily cheapest, way to use AI with your data.

For example, AI tools are now available in enterprise systems such as Salesforce.com and ServiceNow which will draw from their underlying databases. There are certainly advantages to using a provider’s own AI technology, including ease of use, assured compatibility, verified security, etc. The downside of using only this approach is that the business runs the risk of creating data silos. There is a lot to be gained by having an AI solution that has broad access to your operational and customer data (with security safeguards in place, of course). With a full view of how the company is running, an AI system can provide a level of business intelligence that has not been possible until now. And, by having a natural language interface, access to this information in the business becomes democratized and available to all (those whose roles permit it, of course).

Conclusion

The potential benefits of AI are immense and applicable to businesses of all sizes and in all industries. Every business should be evaluating not just how AI can help them, but also how their competitors are using it, or they risk being left behind.

While the technology might appear magical or at least complex at first glance, AI systems do not need to be difficult to work with. There are several ways to integrate your own operational and customer data into an AI system; building your own AI system, leveraging a public system with RAG and/or functional calling, and via plugins and integrated AI.

The process of integrating AI systems with your proprietary data is getting easier by the day, but you don’t need to wait. Solutions exist now, and businesses who leverage them will have an advantage over those who don’t. Stratimetrix can help you define and implement the AI strategy that will get you ahead of the pack. To get started, please contact us for a free no-obligation consultation.