Actual application of AI in production and future expectations: A conversation between Tilde and SBA Group experts

A conversation between two experts of artificial intelligence who read about AI daily and develop actual business solutions. Giedrius Karauskas, Lithuanian Head of the Technology Department of the language technology company Tilde, speaks with Audrius Patalauskas, Head of SBA Group Activity Data Management, about the actual use of AI and its challenges, how AI is accepted by employees, and how managers respond to the content produced by AI. 

GIEDRIUS KARAUSKAS, Tilde: We talk a lot about AI, but we never fully understand how companies have actually started looking for AI solutions to improve their performance. SBA is a large, strong group of companies operating in a number of sectors: manufacturing, real estate, construction, and investment management. Tell us, how did you start discussing AI possibilities internally? Who inspired it?

AUDRIUS PATALAUSKAS, SBA Group: When it comes to generic, non-generative artificial intelligence, formerly called machine learning, automation solutions – we’ve always been using them to a great extent. But there wasn’t any focus on any the section of artificial intelligence; automation solutions were naturally born in the search for efficiency in manufacturing processes and employee ergonomics. Recently, one SBA Group company called Inno Line, won an award in the intelligent industry competition for the use of a variety of technical solutions, which included some traces of traditional AI. These solutions were applied to internal logistics, image-based product quality assessment, and production planning. Such applications of AI are developing very holistically within our group of companies, but now my main task is to find ways to employ generative artificial intelligence, i.e., assistants and language models. The core value of SBA Group is leadership in innovation, so as soon as the market started buzzing about generative AI, the board allocated a budget and set us on our way to explore the best practices. So, we’re exploring, searching, and recruiting big partners to find the best ways to get the most out of generative AI. In my opinion, the companies that will be the first ones to master generative AI will get the greatest return.

GIEDRIUS: That’s right, the practice of having a dedicated person or even an entire team in an enterprise for the deployment and development of AI solutions is becoming increasingly popular in the world, but the company often doesn’t know which person to hire or appoint. What is your personal experience, how did you end up in an AI-related position, and what was your line of work before that?

AUDRIUS: I think most people have ended up working with AI just like me. My main occupation is data management and analytics, and AI likes orderly data because what you put in AI is what you get out in return. Therefore, I think it’s quite common for data analysts to focus on the implementation and development of AI.

GIEDRIUS: I’d like to return to machine learning methods, both neural networks or simpler ones. You said you had a lot of automation solutions in place, could you perhaps mention some really successful ones that have really made a significant impact on efficiency and cost savings? If this is not a secret. 

AUDRIUS: It’s certainly not a secret – perhaps the best example is the product quality control system that employs video cameras. Our plant has about 60 points in manufacturing processes that are supervised by AI cameras, which capture and detect faults, such as various inconsistencies and misaligned holes, and then classify and separate defective products from the good batch. It’s really easy to calculate the benefit of this solution as it is completely tangible: how many faulty products didn’t reach the customer. And when you remove them from the workflow, you don’t incur any losses because only high-quality items are delivered to the customer. We are also starting to use AI solutions to predict when certain machinery can wear out and suffer a breakdown. In this way, we can anticipate unscheduled downtime and turn it into planned maintenance, avoiding rushed solutions to an unexpected problem. This makes it possible to increase the efficiency of the production flow. Of course, it is more difficult to estimate the return of such a solution, as the breakdown predicted by AI might not have happened in the first place.

GIEDRIUS: According to McKinsey’s study, the companies that apply AI solutions see a 25% reduction in operating costs during the first year. Do you think that the full integration of artificial intelligence allows such achievements in your work? Are you a pessimist or an optimist regarding AI? 

AUDRIUS: I think McKinsey was simply cherry-picking within his scope of research. In the first year, I would anticipate 25% savings when only using generative AI in businesses with call centres. In such companies, AI can automate a significant part of the workflow. SBA Group is set to increase efficiency by producing more than the day before, and AI helps us make decisions and communicate them more clearly, although it doesn’t amount to traditional cost savings. In addition, generative AI has additional costs of its own. In June 2024, Goldmans Sachs provided data that were much more conservative than McKinsey’s. They report that only 1/2% of productivity growth and 1% GDP growth are projected over the next 10 years. One year ago, the same bank predicted growth rates of 9% and 6.1%. However, I am still an optimist regarding this technology: the speed at which it is developing is incredible.  

GIEDRIUS: Let’s talk about another tier of AI application: strategy, governance and decision-making. In the next 10 years, AI is expected to change the corporate governance structure of many companies with an increasing focus on the development of “learning organisations” where decisions are increasingly taken by AI. Do you think that in the future, we can expect managers to rely fully on AI models to shape business strategies? Will you still need a “human factor”?

AUDRIUS: In our group, the field of generative AI is evolving rather strangely. I had a conservative opinion myself. I had a hunch that the team and the clients would dislike and reject a solution if we told them that AI was involved in making it. But this is not the case: a month ago, we did this exercise, a miniature hackathon, where the SBA Group head of organisational development worked along with an AI assistant to prepare material for the strategy of some business directions. We gave the AI assistant the criteria and asked it to assess them and provide insights, as well as suggestions. Of course, on all occasions, we asked the AI to mention the sources from which it took the material in order to avoid some hallucinations. We generated a lot of valuable content, and since it was prepared in a valid, logical way, the other colleagues accepted it positively, and the discussion continued based on that content until the decisions were made. I think we saved a lot of time. This example shows that people welcome the participation of AI in decision-making. It is true that there are still AI errors, so-called hallucinations, and human involvement is therefore necessary. 

GIEDRIUS: When did that moment occur when people began trusting AI?

AUDRIUS: Good question. In my opinion, it occurred when it became convenient to use, when we crossed the quality threshold and generative AI stopped talking nonsense. There are many tests that assess the abilities of LLM (large language models), which demonstrate that LLM has long overtaken us in conventional intellectual skills. 

GIEDRIUS: Yes, I completely agree that this trust was gained when we cleared the bar of “good enough”. In our practice, we assess the quality of some AI-based systems by doing a lot of tests. For example, when testing machine translation, we have metrics that we measure in a set of test data, and we assess speech recognition based on the proportion of errors (word error rate), that is, how many errors AI makes in word recognition. And when we talk to customers, there’s always the question of whether AI is good enough, and whether it can provide a satisfactory answer. So, I can only concur that people started to rely on such systems when they saw that they were good enough. But I have a follow-up question: you mentioned that you and your colleague were checking all the sources of information given by AI IN this mini hackathon. But in your automation solutions, where you have 90% accuracy, for example, do you let it run independently without human supervision?

AUDRIUS: In the case of, for example, a solution that removes defective parts from a production flow, yes, the goal is to let it run without a human being if the test proves that 99% of the work is done correctly. Regarding business solutions, even EU legislation, in some cases, does not allow AI to make decisions as there has to be some human input. I think a human being will always be involved in decision-making; although AI ensures speed and efficiency, a real person will have to finalise it. 

GIEDRIUS: Yes, I absolutely agree that people must be informed if a certain system is based on AI or whether the decisions must be reviewed. A similar topic about GDPR (General Data Protection Regulation): how your companies deal with this, because we hear all sorts of things on the market: some prohibit everything, while others allow everything. What is your position on the implementation of GDPR when experimenting with AI?

AUDRIUS: The implementation of GDPR in our organisation is supervised by the department of Business risks, which maintains close relationships with my AI department. We consult with them without any reservations. We use completely anonymous data for all AI experiments. For example, one of our crazy experiments was to study employee goals: we wanted to know whether employees’ personal goals correlate with those of our company, how our employees understand the values of our group of companies and whether they stick to them in their work. I asked AI how the behaviour of a certain employee matches our company values. AI answered me: I’m just simply artificial intelligence and I can’t evaluate a human being, but if you ask how their behaviour fits into a value system, I can answer (laughing). This means that even AI itself has integrated fuses that ensure compliance with EU regulations. We use highly trusted partners, and we have agreements with Google and Microsoft that ensure that data won’t leave the EU, will not be used for other models of training, and will be managed in accordance with the requirements of GDPR. 

GIEDRIUS: I myself have been confronted by people’s resistance to AI; not even generative AI, but specifically speech recognition. People understand that AI decisions will take over their jobs, and they either need to change or die out like dinosaurs.  And statistics show that in the next five years, AI will take over around 30% of the jobs related to data analysis and management. Do you feel inner opposition to AI in your organisations? Do people feel anxious?

AUDRIUS: AI won’t replace you, but it’ll replace a colleague who doesn’t use AI (smiles). This idea perfectly reflects the importance of technology: using AI makes you more important and meaningful because you can work more efficiently and better. It will be necessary to adapt and learn how to use the AI, and those who will manage it will succeed.

GIEDRIUS: Finally, let’s consider the future. Judging from how generative AI has evolved, we can see that over the last four years, it has moved light-years forward. Some experts think that by 2027, we could have AGI; an artificial generic intelligence that is able to learn and perform tasks as well as a human being or even better. Unlike the current narrow artificial intelligence (DI) systems that specialise in one area, such as speech recognition or image recognition, it should be able to deal with any cognitive tasks similar to those that the human brain can perform. There are, of course, experts who maintain that it will take 10, 20, or even 50 years, or even that it will never happen. But let’s consider hypothetically: If we get AGI in a few years’ time, what changes would you envision in business? Where do we use it most?

AUDRIUS: I personally like the definition of general artificial intelligence by Australian scientist Dr Alan D. Thompson: it is a machine capable of acting like an average human being. The average person is now a thirty-year-old Indian woman.  He estimates that we have now reached 81% AGI. A year ago, this indicator was 55%. So we rose by 26 percentage points in 12 months. If the pace of development remains the same, we will achieve it within a few years. If this happens and this technology is cheaper than human labour, I think the predictions made in January 2024 by Sam Altman, OpenAI CEO, that we will have a one-person unicorn, will become a reality. This is a very significant change in society and can be both very positive and very negative.