The development of large language models (LLM) and the competition between them attract interest in these technologies. You have probably already tried to communicate with ChatGPT or another model yourself, and were convinced of their extraordinary capabilities. But the applications of LLM in business are still limited. Despite the promises of all sorts of benefits and demands to use AI everywhere, there are very few real success stories in this area.
In this article, we will give several specific recipes on how to apply AI in business, get benefits and avoid risks.
How is AI currently used?
Let's first look at how AI is being commercialized in the global market. US corporate spending on generative AI, which includes LLMs, grew to $13.8 billion in 2024 (hereinafter - Menlo Ventures data) and increased 6 times compared to 2023. At the same time, 1/3 of US executives do not yet understand how exactly generative AI can be applied in their enterprises. Only 40% of AI costs come from operating budgets, while 60% of funds are spent from innovation development budgets - this shows that the most of AI implementation projects are perceived by managers as experiments. The TOP 5 types of tasks solved using generative AI are as follows:
- 51% of enterprises that participated in the Menlo Ventures survey use automation of program code creation. This is a specific task for IT companies, which is of little relevance for businesses in other industries.
- 31% of enterprises use chatbots and other customer or employee support tools. This is an important area, but it is obvious that even full automation of the support service will only slightly reduce the costs of large companies. In addition, the transition to fully automatic customer service can reduce their loyalty, since in B2C, emotional contact with the client is very important, as well as the use of social and communication mechanisms that are inaccessible to a robot.
- 28% of companies have tools for searching, extracting, and transforming information based on AI. Close to this category is the task of extracting insights from data (27%), distincted by the authors of the survey.
- 24% of companies use automatic creation of meeting and call minutes. Of course, when working remotely and in distributed teams, drawing up minutes becomes a tedious task, but it is difficult to call it a problem that significantly affects the productivity of personnel.
As we can see, the picture is not very cheerful. The technology is used to solve problems that are not of primary or even secondary importance for the business. Automation of the work of support specialists and secretaries is clearly not something worth investing billions in, and not an area in which an enterprise can receive significant benefits (by the way, AI does not work for free either). In terms of economic effect, these use cases are similar to attempts to save on cleaning or water delivery for coolers. Search and retrieval of information is a more promising scenario, which we will discuss further, but here too, we need to take into account the level of its implementation: for example, if we are talking only about searching through a corporate correspondence database, the value of the solution will not be very high.
It is not surprising that, according to the Lucidworks survey, 42% of companies noted that generative AI implementation projects have not yet brought significant benefits.
Can LLMs do more than help create meeting minutes, generate code, and prevent a client from reaching a support operator? Of course, yes, but this requires a little change in thinking and not looking at LLM as a "magic wand" that is in itself a universal tool for solving any problem. Gartner's 2024 report states that generative AI technologies have passed the peak of inflated expectations, and the path to practical implementation lies, among other things, through the creation of composite AI solutions that combine different types of tools, such as traditional ML, Enterprise Knowledge Graphs, and NLU (Natural Language Understanding). This is exactly the direction our company is working in.
As promised, in this article we will provide several specific examples of solving practical business problems using LLM, NLU, ML, and enterprise knowledge graphs. Behind each of these examples are real solutions created by our company. We will be happy to offer and implement ideas for using artificial intelligence and machine learning tools in business processes and your enterprise.
How to Benefit from AI?
Let's list the ways to monetize AI technologies for business.
1. Reduce costs by replacing people with AI agents. This is usually the first thought that comes to managers, but it is definitely not the most effective way to profit from AI. In some cases with this way, it is possible to cut costs, but this approach also has some risks. The same is applicable to the technologies that control people's work without helping them.
2. Increase human productivity. This method promises greater success. Technologies that help people do their job better, rather than replace them, can not only reduce costs, but also increase revenue and profit. This can be achieved by increasing the number of clients served, increasing the volume of products sold, and other options for scaling the business.
3. The ideal option is if the use of AI allows you to create new digital products and services, open new business processes, solve problems that a person cannot solve in principle. This is not possible in all areas of activity, but if this is your case, the economic effect will be the greatest.
Let's estimate the possible economic effect of each method.
1. Reducing costs can, at best, increase the profitability of the enterprise by several percent.
2. Increasing labor productivity under favorable conditions can increase revenue and profit several times.
3. Creating new products and processes can recoup investments made dozens of times or more.
Have you chosen a direction? Let's move on to specific scenarios for using AI to solve business problems.
Scenarios for using AI in business
In this section, we rely only on our own experience in implementing projects using AI. All the scenarios below definitely work and can be profitable!
1. Processing support requests. Completely replacing Service Desk with a chatbot is a dubious idea, but improving business support processes with AI is completely realistic. NLU and ML algorithms can be used to group incoming requests: automatically determine their topic and importance in order to route them to the right specialist. From the text of the request, you can extract essential details, for example, what product and class of problem it relates to. You can automatically select ready-made answers from the knowledge base, or generate an individual response using LLM and RAG (we will talk about this technology later). Finally, it is possible to automatically detect mass incidents and notify duty officers about them, and group all requests caused by one problem.
2. Automation of compliance, the control of fulfillment of certain conditions. Compliance procedures exist not only in the business processes the financial institutions and the corporate legal departments. Usually, there is a large volume of the state and organizational regulatory documents containing certain requirements. The task is to control the compliance of specific transactions or counterparties with these requirements. To do this, it is necessary to extract information from the documents that is related to the fulfillment of known requirements and evaluate it. This is an excellent example of a task optimally solved using the composite AI, an ensemble of models and tools that together allow to achieve a result. We emphasize that in such tasks, AI should not make the final decision for a person. It should only save employee's time by preparing a summary of the necessary information and arguments. This is exactly the method that allows you to increase labor productivity, leaving the work function itself (and responsibility) to a person.
3. Sometimes, people have a hard time understanding the regulatory requirements. In this case, NLU technologies combined with an enterprise knowledge graph, can be used to extract requirements from various documents. Requirements shall be annotated by the subject to which the requirement relates, the circumstances of its application, and so on. This will create a knowledge base with which can an employee can quickly prompt to obtain the requirements applicable in a particular case. In combination with the approach described above, the system can generate a summary of their compliance.
4. If a company works with a large volume of incoming documents, be it loan agreements or equipment supply specifications, their processing can be automated. NLU technologies can extract key facts from each document, classify it, annotate it, place it in a database, and make it searchable. LLM can summarize the contents of a document, including key facts in a summary.
5. Interaction with the data in the dialog mode. Solutions of this class use RAG (Retrieval-Augmented Generation) technologies. The essence of the technology is that LLM generates a response to user questions based on structured data. For example, the user asks - "What is the balance of my personal account?" NLU tools determine what the question is about and form the context for the answer, retrieving information from the business application database or from the DWH. In our example, the context will contain the value of the current balance of the client's personal account. LLM will generate a response using this value.
RAG technology allows LLM to "be aware" of current data without retraining it. RAG-based solutions allow you to create customer self-service tools in a mobile application or on the company's website, as well as interactive assistants for employees of the enterprise. They increase the availability of information and can be a component of a corporate knowledge management system, an intelligent search system that provides employees with access to all information accumulated at the enterprise.
6. For companies purchasing a wide range of products, NLU tools will help in determining the characteristics of the offered goods and comparing them. Let's say a supplier sent an Excel file with a list of goods and prices, where the product characteristics are indicated in one line, such as "Engine 3200 kW 500 rpm SDN3-1776". NLU tools will determine the product class - engine, its power - 3200 kW, rotation speed - 500 rpm, and model - SDNZ-1776. This will allow you to determine whether the proposed equipment meets your request and rank offers from different suppliers. Based on such tools, you can create services for selecting product analogs, automatically parsing specifications/purchase lists and converting them to the reference equipment catalog hosted by MDM, and integrating with trading platforms.
Similarly, you can parse other text strings containing partially structured information, such as postal addresses.
We have listed here only those use cases for NLU and LLM that we have implemented or piloted ourselves. There are many other ways to benefit from these technologies. If your business has tasks that can be solved using them, contact us: we will offer you a pilot project to assess the effectiveness of the solution. This could be automation of information collection from the Internet, generation of product descriptions, incident reports, and much more.
Recipes for Success in Implementing AI
In conclusion, we will share a few more considerations that need to be taken into account when implementing AI in the business processes of your enterprise.
Gartner in its 2024 AI Report emphasizes the importance of preparing data for use with AI tools: a person has at least a chance to notice questionable data, while AI accepts any information as true. If your enterprise has problems with disparate and poor quality data, it is worth starting with solving them - at least in terms of the information that AI will use.
It is important not to follow "tunnel thinking" and not automatically accept certain technologies and tools promoted by experts. It is necessary to calculate the planned cost-effectiveness of implementation, determine the available data, describe the desired end result, and then select the tools that will allow to achieve it within the existing resource constraints. This also applies to the use of composite AI technologies: using RAG on top of structured data can be much more profitable than retraining the model using this data.
When using AI, you need to pay attention to security issues. Transferring data to the model is also data transfer. If you intend to use a third-party model, such as ChatGPT, you need to determine whether the law and your obligations to counterparties allow you to transfer certain information to third parties. If this is not the case, you need to use privately deployed AI tools. But even in this case, you need to make sure that, for example, a conversational AI assistant will under no circumstances tell the user information that he should not have access to (this is much easier to achieve with RAG tools than with retrained models).
And, of course, we cannot lose sight of the ethical issues of using AI. Make sure that the solution you are implementing will benefit your company's employees and its customers, and will not reduce someone's income or worsen the quality of services.
