KPMG Expert: AI Solutions for Automating Routine Processes Deliver the Quickest Returns

What has been our current experience with artificial intelligence so far?

Computer-based statistical analysis began 30–40 years ago, when computers became sufficiently powerful. A couple of decades ago, they started calling it machine learning, although the underlying mathematics remained the same. Now we call it artificial intelligence.

Yes, we have used artificial intelligence in some specific areas, but the use cases have been very limited so far. When it comes to large language models (LLMs), there are specific niches where AI is currently performing exceptionally well and we can talk about new use cases. At the same time, the old statistical predictive models are under-utilised today, despite having the greatest potential for value creation.

For many years, I was involved in the implementation of various models in the United States. More than 20 years ago, the large retail chain Macy’s was already able to adjust prices in its stores in real time in order to optimise its stock levels. When they needed to launch a campaign and had goods available in the warehouses, they analysed forecasted customer numbers and stock volumes in stores. They also considered other indicators, such as whether the weather was rainy or sunny.

Based on these various data points and stock levels, the system suggested campaign prices for different products. Such a solution is highly profitable for companies, and this capability has been in place for over 20 years.

To bring another example, machine learning, now called AI, has long been used to decide whether a credit card transaction should be blocked or sent to a human for review. Such decisions were made by a predictive model based on dozens of data points, helping to prevent credit card fraud.

These models, developed over 20 years ago, are highly effective and have a significant financial impact on businesses. While this is nothing new for banks, many of these things are very new for the retail sector in Estonia.

For example, optimising discounts with a machine-learning model delivers a return on investment in just three months. Show me another technological solution with an ROI of three months – I cannot think of one! In other words, we are basically talking about the same old things again. True, large language models have different use cases.

What do clients ask you when they seek advice on AI?

They want to reduce labour costs and the number of staff in certain functions, such as finance, sales or warehousing. Today, quite basic robots can automate certain tedious tasks, all at a low cost.

Computer technology enables automation, including decision-making in some cases, such as in the context of credit card fraud. If there is a clear signal of potential fraud, the card will be instantly closed without the need for human intervention. This means the decision is made by artificial intelligence, if we may call it that.

An important factor is the size of a company. In Estonia, we have few large companies and many small and medium-sized enterprises. Regrettably, the ability of these enterprises to implement such new technologies is limited. There are exceptions, of course. Optimising logistics and warehouses delivers significant benefits.

What we sell is results. We do not sell analytics, as it is too abstract; we sell solutions to specific problems. For example, if a client’s stock levels are consistently higher than those of competitors, the task is to optimise inventory management to prevent excessive capital from being tied up in goods.

Where should a business manager or owner start when considering the adoption of artificial intelligence?

They should begin by assessing their current performance within their sector. In other words, they should explore whether they are falling behind international competitors, for instance.

The average profit margin of Estonian companies in terms of EBITA is two to three percentage points lower than that of companies in the Nordic countries. The reasons are, on the one hand, the size of the companies and, on the other, access to higher paying customers. However, another factor is the overall quality of management, which can be understood in various ways. One aspect is automation, while another is that we may be dealing with the wrong customers who are paying too little.

So, the answer to the question is that the source of new knowledge is comparison. A business manager must be able to compare their performance not only with Estonian companies but also with international firms, which are likely to have higher margins. This leads to the question of why they are able to achieve higher margins.

For example, at KPMG we have worked with Estonian retail companies. Their EBITA margins are two to three percentage points lower than those of companies in Finland. True, Finland has an oligopolistic, very consolidated market.

However, in terms of analytics, Kesko in Finland employs a dozen PhD-level data analysts who help optimise pricing, inventory, store locations, number of staff per time period and other aspects. The goal is to get the most out of the relatively expensive resources, which in retail include both space and staff.

How many PhD-level data analysts are employed by Estonian chains such as Selver, Coop or Rimi? This is where the difference begins to show. Managers must identify their needs by comparing themselves with foreign companies.

We were discussing that AI has been around for several decades, although under different names. How many grey areas or unregulated aspects are associated with it?

There are some. We even use the expression ‘AI in control’ in our sales activities and aim to ensure that AI remains under our control. I rather think that there is little grey area in Estonia, as the overall implementation of data analytics is limited, and the readiness to adopt it also remains relatively low.

As consumers, we have seen the grey area where AI is used to make certain decisions that we do not like and that are often based on very limited or inaccurate information. For example, Facebook may remove some of our posts or images. Its model is trained to make decisions based on certain data, but these can contain errors which ultimately affect the consumer.

Let us take the case of the Estonian FinTech startup Planet42, now in liquidation, which has received significant media attention. They, too, used machine learning models to make credit decisions. This is a borderline case, and the challenge, or the grey area, begins with the training data. If the training data is flawed, there will also be many of flawed responses. One example is Facebook, another is the credit decisions. We do not know which models companies use to make their decisions.

Several years ago, while working with a telecom company, we discovered that when a salesperson wanted to offer additional products to a certain customer, they could see that customer’s potential profitability displayed on their screen. Specifically, they could see the customer’s profitability for the company and determine how much they could sell or offer in discounts without reducing the profitability. This, for instance, is one of the aspects the EU AI Act aims to regulate – whether the decision on the size of the discount offered to the customer for an internet or telephone connection is fair.

Clearly, the European Union is leading the way, ahead of both Asia and America, in terms of AI regulation. We can argue whether this is good or bad, but others will follow the European regulation.

The same was true for data protection regulation, where companies outside Europe started to adopt it. Companies are global, after all. Once you are in Europe, the same rules are likely to be applied in Asia and America. The same will happen to AI regulation.

Has regulation been an issue in the field of AI?

The issue has been that most models based on neural networks, including LLMs, or large language models, are often described as black boxes, which we cannot understand. In fact, science has advanced to the point where these models and the decisions they make are now auditable.

Some time ago, there was a big debate about whether these models could be used at all if we did not understand how they generated decisions. Now, science has advanced to the point where these models can be ‘broken down’. I recently came across two scientific articles that suggest that there has been a significant breakthrough in the field of auditing.

Last autumn, KPMG stated that ChatGPT was unsuitable for business use due to the high risks involved. Given the recent developments, do you believe ChatGPT is now more suitable for business use?

Yes, it is, significantly. We have now also implemented licensed GPT solutions, which offer different privacy measures and terms of service compared to the free version of ChatGPT. These solutions are available to everyone, though with some limitations.

The core issue lies in the responses a GPT provides. Using it to generate a draft text is not an issue. However, when it comes to providing factual information, the GPT consistently falls short compared to tools like Google Search. Based on its training data, it may provide incorrect answers, which is why we cannot simply rely on LLMs to generate texts for us.

By training our people, we have reached a point where we feel confident in making it available to them, which has been a great help. For example, we can reduce translation costs and deliver higher-quality texts to the client much more quickly. However, editing remains crucial – you cannot remove people from the process.

If I have a basic knowledge of a foreign language, I can use a language model for translation somewhat blindly, as I am unable to assess how many mistakes it makes. However, an advanced language user will spot numerous mistakes and these must be corrected.

Another issue is search functionality, which remains highly unreliable today. These models are not well-suited for conducting searches. For example, I tested several models by asking them to find information about Tarmo Toiger. The results are incredibly interesting. I have never done any of the things listed in the search results, but they certainly sound impressive.

How can KPMG assist an organisation in establishing secure and controllable generative AI capabilities?

The quickest returns are achieved through solutions that focus on automating routine tasks that are performed manually, such as data entry, data transfer, or generating basic texts in sales or marketing. Another area to consider is the transfer of information between systems – for example, between warehouse and sales teams, sales and marketing teams, or finance, accounting and sales. Much of this data is still transferred manually between the systems, but today, it could be done by a robot, with appropriate controls in place. This is where organisational efficiency can be significantly enhanced.

And then there is the public sector. The Estonian government is gradually deploying bots, which is a highly positive development. However, deploying bots requires data analysts, and we simply do not have enough of them. We are too small a country and lack capacity in this area.

What does the collaboration between your team and the client look like during the implementation of AI? What is your role, and what is expected from the client?

Our first task is to assess the potential, to identify the processes with the greatest benefits, where we can save hours of work and some activities. We will then ask the client to select the processes from a list of 10 to 20 that are likely to deliver the greatest benefits to them and try to automate these processes.

Ultimately, the goal is to apply technology where it can deliver the greatest value. For example, when it comes to simple office robots, or office AI, the implementation process typically takes two to three weeks until it is up and running. So, it is very fast.

Deploying more complex, statistical models, however, requires significantly more preparation. The first step is to identify the data that most accurately represents the situation. In other words, we need training data and must determine which data points can be used to build the model. The next step is testing, which typically takes months rather than weeks.

The benefits of simple data analytics and robots can be reaped very quickly today. It is a matter of will; the need is always there.

How can the AI implementation process be managed given the rapid pace of AI development and evolution?

This is more relevant to language models than to other algorithms used for process optimisation. Having received much attention and investment in recent years, language models have indeed advanced very quickly in just a month or two. However, this progress is now beginning to slow down.

As regards other statistical models, they have not evolved as quickly. The underlying mathematics available today dates back to the 1960s; the issue is that we do not use it.

The development of large language models is unlikely to be as rapid in the future and it is expected to stabilise at some point. Language models can be used, for example, as office assistants that use language to provide advice on solving problems in customer service or sales, or to generate text. The larger the datasets, the better these models become.

Today, the quality of translation, for example, is worlds apart from what it was a year ago. This includes the translation of relatively exotic languages. If today we use a computer for translation, it is very likely that in the near future, simple earbuds will be available that can listen to a text in a foreign language and translate it for us. There are already prototypes.

As for other models, their development has been relatively stable, without the same rapid advancements. Instead, computing power and the capacity to leverage large data volumes have increased, thanks to the availability of cloud computing.

Have your clients encountered any legal risks, and if so, what are the most significant ones?

The greatest risk is making the wrong decisions. Leaving aside self-driving cars and other 'hard problems’, the biggest concern with simpler issues, such as whether to approve a loan or the publication of a document, is likely the auditing of these decisions.

The more widely AI is adopted, the more such challenges are likely to arise. For example, if texts sent to clients are not properly reviewed or edited, and clients make decisions based on such texts, legal disputes could certainly follow.

This can be avoided by having an editor function in the loop, or by having stricter rules about what conditions the generated text must meet when there is no human in the loop. In any case, there will be a lot of interesting case law, I predict.

Such problems have already arisen around the world in industries such as aviation and insurance, where inadequately tested solutions have been implemented, resulting in obligations for companies. And then, you will face a host of issues.

How important are privacy and copyright concerns for those implementing artificial intelligence?

They are crucial, as this is a very specific area governed by complex legislation. When a new model is introduced, we need to consider which team will be responsible for analysing it. We need a minimum of three roles: a data analyst to evaluate the model, a personal data protection specialist, and a lawyer to assess it from a purely legal perspective. They must analyse the model to see if it works as intended and to understand how it makes decisions. The specifics are, of course, different for each model.

How big are the changes that AI will bring to businesses in the near future?

The changes will not be very significant. We tend to overestimate the capabilities of LLMs, but they have certain limitations. Language models as assets are pure hype and there is no rational basis when it comes to valuing these companies. Their valuations are too high for their real potential and are bound to come down. While it is great that LLMs are being developed and a new way to create value has been discovered, the value creation is certainly overrated.

I have also spoken with our leading data analytics experts on the science side. They confirm that the mathematics of universal AI, or artificial general intelligence (AGI), has not yet been created. We are still a long way from AI being able to replace humans. At the same time, there are specific tasks that AI can perform effectively. Although it makes a lot of mistakes and is not perfect, it does have an impact.

However, the statistical machine learning models developed earlier, which optimise various processes, can have an even greater impact. When implemented more effectively, they can deliver more significant results than language models, bringing greater financial benefits to companies.

Is it accurate to say that the impact of AI on businesses is not yet comparable to the impact of the advent of the internet?

You are right, it is not really comparable yet. The current hype surrounding language models is overstated. The very definition of ‘hype’ implies the exaggeration of something’s impact. The impact of language models is definitely overrated at the moment. Then again, who can predict the future? We do not know what is going on in the labs. However, my view at the moment is that, given the current state of scientific knowledge, it is not yet possible to predict the next rapid breakthrough.

How can we improve accuracy and increase the value of AI?

Increasing the accuracy of language models largely depends on the training data. Thoroughly verified data can enhance accuracy and help avoid factual errors. However, due to its inherent structure, a language model tends to generate text in a relatively arbitrary manner, based on the structure of the training data.

A language model knows nothing about the meaning of words. It orders them based on statistical probabilities, on how people have written similar texts before.

The model’s behaviour can be manipulated by manipulating the training data that we feed into it. Indeed, outstanding results have been achieved with highly specialised models, such as those used for various exams. When such a model is trained on a specific type of data, it typically delivers excellent results. However, that does not mean it actually understands anything. It simply provides the correct answers within the context of a specific test.

However, you could not send it to court to act as your lawyer. It would fail entirely because it lacks the ability to understand the meaning of words.

So, AI still has some way to go before it achieves true intelligence?

Actually, a very long way! The general public does not realise this. It may seem as though AI is thinking when it generates text that appears human-written, but there is no thinking involved, at least not in the way we understand it. It simply generates words in a more or less logical order, and nothing more.

So, the quality of the training data is one of the biggest risks in implementing artificial intelligence?

Yes, certainly. What could happen today is that only large companies will advance in developing language models, which is concerning, as they have access to vast volumes of data.

If you try to set up a start-up and apply the same technology but you do not have access to training data, you will not achieve much. Given the models and technology we use, training data is the key to everything.

The same applies to self-driving cars, for example, though this is not their only challenge. It is not just about the training data but also about sensors that trigger false responses and alarms.

Estonia does not have large companies on a global scale, but AI is not a technology that benefits only the big players. As a smaller Estonian company, you can implement AI within your niche and reap the benefits immediately, right? This could also inspire other small businesses to consider integrating AI into their operations.

Yes, absolutely. The question, however, is what type of artificial intelligence we are talking about. For example, if we want to use artificial intelligence or machine learning to reduce inventory costs – by optimising the amount of stock we hold at any given time – that is based on our own data. We can analyse the past three years of stock movements and transactions, and enrich the data with factors such as weather patterns, competitor activities, marketing campaigns, etc. This dataset is within our own control and ownership. That is why we achieve the maximum benefit.

When it comes to language models, the situation is more complex. This is where the topic we discussed earlier – personal data and copyrighted texts – becomes relevant. Not everyone has access to such data. Those who do are in a very strong position and can fine-tune their model accordingly.

Now, let us consider the example of a call centre where calls are recorded. These recordings can be transcribed into written text, creating a personal dataset that can be used to generate intelligent responses using artificial intelligence. This does not require third-party data, only proper preparation.