Artificial intelligence will transform commerce. Which solutions to choose? Ready-made products, or maybe tailor-made? - interview with Barbara Sobkowiak
Barbara Sobkowiak - Data Science Manager at Britenet - in an interview with Michal Kokoszkiewicz of the Commercial News portal, talks about, among other things, issues related to artificial intelligence in commerce and its solutions and implementations.
What is the state of development of artificial intelligence in commerce? Is it a technology that is barely noticed or is it already quite clearly formed in commerce?
Commerce is increasingly turning to artificial intelligence in a bid to improve various types of processes or gain a competitive edge. This can already be seen not only in the company's internal operations themselves, but also at industry conferences and in conversations among end customers. We, as customers, also notice that the company is introducing innovations of some kind. This is a trend that is getting stronger and is being used more and more. I myself notice that customers, i.e. companies involved in trade or logistics, are more and more willing to use this type of technology. This is definitely already the moment when we say to ourselves that commerce is using artificial intelligence more and more widely and more boldly.
In which areas is artificial intelligence most often used? Is it in logistics, supply chain, procurement, or perhaps other areas?
Supply chain is the area where we reach for artificial intelligence most often, because the whole issue of inventory management, forecasting, demand, is fundamental in this area. Especially at a time when these supply chains have been a bit shaky in recent years, and logistics has had to work hard to make sure we have goods on the shelves. This is a major area that is constantly being talked about and where technology is being reached for on the demand side and on the route planning side. In warehouses, too, automation, streamlining, optimization, even if only in the arrangement of goods in the warehouse and on the shelves to make sure it's done as well as it can be.
In which areas do you think it can still be developed? Where can artificial intelligence in commerce still appear or will appear in the near future?
I think in the near future, all (retail) chains will move toward some kind of personalization. In e-commerce, things are a little easier because we can see who the customer is and we can track what they buy and what they ask for. The more classic networks, i.e. stationary, where we can go physically to the store, are also starting to do this. Of course, they implement it in a slightly different way, they use some kind of app, points or cards. We may not know who the customer is, however, we can determine that this is a certain ID that buys in a specific way. Here quietly this personalization will expand, so we as customers will probably get more and more personalized offers, prompts and promotions. An area that I don't think has been fully developed is issues related to managing and forecasting promotions. Here there are still some simpler algorithms in operation, which also work well, but I see a lot of room for artificial intelligence here still.
Do you see any opportunities in the possible use of artificial intelligence in the stationary store itself? In matters related to the distribution of the assortment in the store, guiding the consumer, can it also be used?
As for the unit store itself, here we have a whole range of applications - starting with selection of assortment, products for sale, purchasing power, up to statistics of the most frequently purchased items individually, as well as in pairs with others. There are also issues of seasonality trend - here as much as possible assortment selection. Going further is the arrangement of goods in the store, because it's not a well-known fact that where a product is located is important to the end customer, since we often buy with our eyes.
Another place where the algorithm can be applied is certainly in the management of inventory in the store, that is, how much of the item we need and how much to actually order. Just because we need a hundred pieces of a particular product doesn't mean we will order the same amount. We may have some kind of logistical minimums, packaging, financial issues that will indicate that we should order a different quantity of a given product. As you can see, we still have a lot of room for optimization algorithms here.
Autonomous stores, where there is no attendant, are becoming more and more popular, and at this point artificial intelligence is actually taking over the role of the salesman. It has to observe what the customer has purchased, bill them for the goods, as well as keep track of the products in stock. This trend has expanded a lot in Poland in the last year, but also around the world. Certainly another possibility for artificial intelligence is to look at where customers go. If we're talking about large stores, tracking where customers are most likely to go in order to be able to arrange the aisles differently and position the assortment differently.
How is it with the popularity and use of personalized solutions, tailored exclusively to the customer's needs, and more generic, off-the-shelf solutions, perhaps not necessarily perfectly tailored to a specific case? How does this look on the market?
It depends, a lot of companies like to go for so-called boxes, which are practically ready-made solutions that you just need to implement. They often work well in this type of application, where we have a repetitive problem and repetitive data. We're talking here, for example, about workflow - it's repetitive for a lot of companies, or some kind of image analysis, such as checking the type of products and the correctness of their price, as they may occur in different chains and stores.
It is a little different in solutions that touch very much on the business process created by a particular company. If we talk about promotions, each company manages them differently, which also translates into having different data and alignment in the systems. Here, however, I often observe that clients opt for more customized solutions, because the boxed ones are too simple or would require a lot of customization and adjustment to the data the company has anyway.
What are the other strengths and weaknesses of off-the-shelf and custom variants?
Let's start perhaps with off-the-shelf variants and their strengths. Certainly, the strong point is that a company, buying such a box, worries less about artificial intelligence specialists, because AI is already sewn into the package, integrated in some way. The whole issue of managing and hiring Data Scientists goes away. Such solutions are often faster to deploy and testable because they are half-finished or even completely ready. All we have to do is install them and do the integration with the systems, so we speed up when it comes to implementation time. In addition, however, they are often systems that already have a GUI (graphical user interface) ready. This is a plus and a minus in one, because on the one hand we get a GUI that does not require design. On the other hand, we can't interfere with it, it must remain as it is. The manufacturer can only agree to change the color to match our company's color scheme, but will not completely change the approach and the way we work with the system. Additional disadvantages of boxes are certainly the cost of licenses. The cost of implementation and purchase is one part, the other part is the license itself, for which we pay annually - here we bind ourselves immediately to a particular manufacturer for years. In addition, these are not very flexible solutions. A company buying a ready-made package usually has no, or very limited, possibilities to change algorithms, data formats or other GUI features.
When it comes to custom solutions, the strengths are certainly that they are highly customized to individual needs and specifics of the data. In this situation, we are working on two years of good data that the manufacturer has. We can use exactly the whole range of algorithms that exist on the market and are available in all software libraries. We are not limited to what a particular manufacturer has prepared. Data Scientist can use neural networks of different types, parameterize in different ways, test very different variables and data. We can fully adapt the system to what the company's business process looks like and what the customer's data looks like. In addition, these are completely flexible modifications in the code, not only from the algorithm side, but also from the side of integration with other systems in the company. It is known that it is never one integration - it involves multiple systems, notifications, data, couplings and other similar processes. What always works to the disadvantage in this type of solution is the requirement to have the right specialists. Either we have such a team in-house, or we need an outside supplier to prepare such a solution for us, which also entails that the project will be more time-consuming. In my opinion, this translates later into quality, because despite spending more time preparing the algorithm, it is actually tailored to the data the client has and the process they actually want to manage.
Or are there some areas where off-the-shelf solutions or custom solutions should always be used? Are we able to adapt these technologies enough to implement both solutions everywhere?
I think that off-the-shelf solutions cannot be implemented everywhere. Such ones are much worse for so-called tabular data, that is, data describing, for example, sales, promotions or orders. They work much better for document or image analysis.
As for custom solutions - I think they can probably be used anywhere, only it won't always be cost-effective. In fact, it all depends on what we have at our disposal as an ordering company and what we care about. There is probably no area where you can say with 100% certainty that a ready-made or custom solution will work.
How does the cost of the two solutions compare? How will this change in the future?
Today, the biggest cost, regardless of the choice of deployment version, is the cost of data preparation and machines, i.e. computing power. It will always be costly to prepare data until an organization takes great care of it and gets its data management in order. Even if we have a box ready, it needs data in a specific form and from a certain range. I know of companies whose data preparation time for one of their solutions has extended up to twice as long.
As for computing power, it represents a very large cost due to the training of models that have to be taught periodically, as they can get old and will need to be refreshed. These are always the two biggest design costs, on top of the cost of specialists. Whether we choose the boxed option, where we need implementers from a particular company, or an organization that creates the solution for us from scratch, the staff time is easy to calculate - it's quite a lot.
The cost also includes the license, but the issue of maintaining the entire system is often forgotten. Artificial intelligence systems are not systems that we simply install. They have to walk, they have to refresh, they have to be monitored, and very often they also touch quite sensitive processes. They can't fail to work for three days, they have to work in a continuous system. They need to be monitored, and they need to be very well protected against various oddities in IT systems. It is advisable to have people who will be able to help in case of a problem.
This is what the cost component of projects looks like. I think this will change as more and more companies in Poland try to organize their data. I personally hope to have a little less trouble in the future to organize data for projects, so it will be cheaper and faster. The second issue is computing capacity, which has been getting cheaper so far. The whole IT environment is trying to optimize code, so I hope to be able to reduce costs in the future. These are two points where it will be possible to cut expenses a little more for projects. Organizations and people are becoming more aware of how artificial intelligence works. This is causing fewer and fewer problems in the project and easier implementations.
You can also find Barbara Sobkowiak's conversation with Michal Kokoszkiewicz of Commercial News HERE