Transcript

Tobias: Welcome to our webinar: how large language models are changing the eCommerce customer experience.

[0:08] My name is Tobias. I’m doing the moderation today. I’d like to introduce Chewbacca, who's right here behind me. He’ll be watching over me during the webinar. And I’d also like to introduce our speakers for today, who are much more important than I am.

[0:30] We have our two product managers, Yuliia and Tufan — it makes me very happy that we’re on a first name basis in this company. We also have our business consultant, Felix. Like I said, I’m Tobias. I usually take care of offline event management at FactFinder. And since this webinar is definitely a large event, I'm taking care of this one as well.

[2:13] So that's my part. I'm not taking any more of your time. I’ll hand over to Felix, who’ll start with the first point, the rise of large language models, a realistic perspective.

The rise of large language models

[2:37] Felix: Thank you, Tobias. Hi, everyone. I'm Felix. I’d like to kick us off by going to the day where it all started for most of us, which is November 30th, 2022. That's the birthday of OpenAI's ChatGPT. And as you can see on the next slide, since then, a lot has happened.

It only took ChatGPT five days to surpass 1 million users. And after less than two months, in January, they had already accumulated over 100 million users. That's a feed that actually took TikTok over nine months and Instagram over two and a half years, which I think puts into perspective how important ChatGPT really is, compared to some other tools we use on a daily basis.

I think this UBS study quote sums it up quite nicely: “In 20 years following the internet space, we cannot recall a faster ramp in a consumer internet app.” So, I think that's quite fitting.

[3:41] Now it's May of 2025 and a lot has happened since then. OpenAI's ChatGPT is now almost 2.5 years old. If it was a human baby, it would already be able to walk, jump and talk. And I think it's safe to say that they have also jumped, or the AI has also jumped. It has made a big jump since then. So only in April, OpenAI reported having over 800 million users, which was up from only 400 two months earlier. They get about 5.2 billion visits per month. And the average user spends about 8m13s during one session, which adds up to about 82 years – a whole lifetime – that humanity spends in total every month on the app. So, I think that puts into perspective how important this app has become for most of us.

The AI hype

[4:51] Naturally, there's quite a big hype around ChatGPT and AI now. We already have over 1 billion devices on the planet running some kind of AI assistant. As of two years ago, there were about 15,000 artificial intelligence startups in the US alone. Back then, there were, I assume, hundreds of thousands of AI tools on the market, and those numbers have already increased. If you're an eCommerce manager, you've probably seen dozens of emails in your inbox from different companies informing you about the latest AI trends.

[5:46] Of course, there are many good reasons to start using AI in eCommerce. Maybe one of the most important reasons is to save time. You can automate, for example, content creation for SEO. You can automate newsletters or recommendations. You can reduce manual labor and the risk of human errors, increase productivity and introduce 24/7 operations.

For all those reasons and more, most eCommerce companies (80%) have started using or exploring AI. Only about 3% of companies don’t plan to use it in the future.

[6:59] According to a 2025 Stanford University report, AI adoption is still rising among companies. About 78% of companies are currently using or exploring using AI. Although, those rates include other countries, such as countries in Asia, which have higher adoption rates compared to Germany or other European countries.

[7:14] Investments grew to over $250 billion this year, and the AI is becoming even more intelligent and efficient over time. Performance has grown by up to 67.3% since last time they checked. It's also becoming more and more affordable for companies to use AI.

Challenges of AI Implementation

But still, AI isn’t all great. It can sometimes lie – which frustrates customers rather than helping them. And in some cases, it can result in financial damage. For example, Air Canada’s AI chatbot made up non-existent refund policies, which the company had to honor.

[8:34 ]Another example is the McDonald's Drive Thru disaster. McDonald's was one of the very first companies to start investing heavily in AI. They tried to automate the drive-through process using AI, which sometimes resulted in strange order combinations.

For example, the AI added hundreds of dollars’ worth of products to an order without asking the customer. Another example would be combining bacon with ice cream, which was quite comical for the customer, at least. I'm not sure that's actually a bad combination, but I guess it depends on your taste.

[9:31] Even Google can sometimes display wrong recommendations based on AI. So, for example, if you ask how many rocks you should eat each day, it will recommend eating at least one small rock a day, which shows the AI doesn’t truly understand the search query. There's no real nuanced grasp of the customer's intention or their preferences, and so it gives a completely irrelevant AI overview. There's no real benefit for the customer in seeing these kinds of recommendations.

[10:11] For all those reasons and more, Forbes has made the audacious quote, I guess, that in their opinion, about 95% of all AI products will fail. This would mean that many companies’ investments in AI would be wasted.

To be among the 5% of AI products that actually succeed, we should keep in mind three points:

1. The first point is to actually anchor all decisions around solving customer pain points. So, we have to know what those pain points are and then solve them using AI.

2. The next point is to deliver consistent innovation and unique benefits. That means keeping up to date with the latest technology and creating new benefits, which aren’t already on the market.

3. And then thirdly, harmonize rapid delivery with reliable service quality. If you don't have reliable service quality, you end up with the examples we just saw from McDonald's, Air Canada and so on. If you’re reliable with service quality, you can avoid those kinds of issues.

Introduction to vector search

[11:45] At FactFinder, we’re actually able to cover the three points I just mentioned. We've been working with our customers on product experience for over 20 years. We started working with AI and machine learning almost 20 years ago, so we have lots of experience in that area as well.

We know what our customers' pain points are and we want to solve them. We’ve been innovating for many years now, and our customers rate us very highly, which is, I think, also a good sign that we’re on the right track. Hence, we’ve started developing our own AI tools based on large language models. One of those tools is vector search, which we're going to talk about next.

So, why invent or why use vector for your eCommerce site search?

I think it's always a good idea to follow Google's lead. Usually, Google knows what they’re doing. They are the biggest search provider in the world, so it’s probably good to listen to them. And according to Google, their natural language queries increased by 60% between 2015 and 2022. Please keep in mind that's by 60%, not to 60%. Keyword search is still the most important kind of search, but still, using natural language queries has increased significantly since then. It’s growing even more popular.

Plus, 70% of shoppers think a lot of search engines need improvement and/or struggle to find good results using natural language queries. So, it's important to try to improve the results for that majority of customers who have issues.

[13:49] In addition, a voice-based search is only really possible with vector search and not with keyword search.

And so, for all those reasons, and more probably, Google launched BERT in 2019. BERT is their kind of vector search, which tries to better understand natural language queries.

On the left-hand side, the before example shows kind of fitting results for the query, but the right-hand side understands the query even better and displays even more fitting results, because it uses natural language. And so, because it can improve the Google experience, it can really improve the product discovery experience in your online store.

Now, I’ll hand over to Yuliia to tell you more about our vector search tool.

Yuliia: Thank you so much, Felix. Hello, everybody. My name is Yuliia. And let's talk about vector search. At Factfinder, we’re not chasing trends. We’re solving real problems faced by our eCommerce customers.

When we decided to evolve our search capabilities, we didn't start with technology. We started by listening to the data, specifically the search behaviors and frustrations of the shoppers across our customer platforms. And we’re systematically analyzing millions of eCommerce search queries, focusing on critical metrics, like:

• Zero-hit searches – queries that returned no products

• Click-through rates – whether users found something relevant or not

• Post-search engagement – did shoppers continue browsing or abandon?

• And of course, customer feedback from interviews

This analysis revealed a simple but powerful truth. Traditional keyword search is not enough anymore.

Understanding user search behavior

[16:03] Our analysis showed consistent patterns in how users search, allowing us to categorize queries into distinct types, each reflecting different user behaviors and challenges.

• Natural speech queries – let's talk about natural speech queries. Apparently, AI chatbots are influencing how users are searching right now. Shoppers are using full sentences or descriptive phrases like “I want a dress for Christmas.”

• Intent-based queries – users search based on needs, not exact product names, for example, “something to keep my coffee hot during commute” or “wedding guest.”

• Synonyms – different shoppers use different words for the same things, for example, “winter coat”, “parka”, or “jacket.”

• Mistakes and typos – quick typing, especially on mobile, can lead to errors. For example, we can write “iPhnoe” instead of “iPhone” case.

• Plurals and grammatical variations – small grammatical differences impact results. For example, “men's boots” versus “boots for men.”

• Mixed language queries – and last but not least, mixed language queries. In multilingual markets, shoppers often blend languages. For example, “cheap Jacke,” where “cheap” would be English and “Jacke” is in German.

These challenges are common across all industries, from DIY, fashion, to B2B and food, underscoring the need of more intelligent semantic search solutions to go beyond traditional keyword matching.

How do we approach vector search?

[17:49] So, how do we approach vector search? This image shows a high-level overview of the process. Vector search retrieves information based on meaning, not just exact keyword matches.

It starts when a user submits a query — for example, “winter sleeping bag”.

An embedding model then converts this query into a vector: a list of numbers that captures its meaning. These vectors, also called embeddings, are dense numerical representations of data, like text or images, that reflect their semantic content.

The same model also transforms product data into vectors in the same space, allowing for meaningful comparisons.

The system then compares the query vector to the product vectors and identifies the ones with the closest meaning. Finally, the most similar items are returned as a search result to our user.

Privacy and control

[19:02] Let's talk about privacy and control. When designing our vector eCommerce sight search solution, we made a deliberate decision.

We do not use OpenAI APIs or external cloud-based LLMs for embedding generations. Instead, we focus on self-hosted offline embedding models, because:

• Data privacy – self-hosting our embedding models offline ensures data privacy. No customer search data is sent to third-party servers.

• Customization – we can fine-tune these models on specific customer catalogs and industry-specific languages.

• Performance and stability – hosting embeddings locally ensures predictable latency and system control.

• Compliance – and, of course, compliance. It's easier to meet strict industry regulations around data handling, like GDPR and others.

Demonstrating Vector Search Capabilities

[19:55] I’m going to share a few examples with you. So, what I'm sharing right now is basically the backend of FactFinder. And here on the left side, we have keyword search activated. And on the right side, we have vector search activated. Let's run a few queries together.

Here, we have “günstige kletterschuh für anfänger,” which in English, translates to “affordable climbing shoes for beginners.” So, as you can see, Keyword Search here doesn't perform well, because it doesn't understand the meaning. The query is super long, but vector understands the meaning here and can identify the intent.

Here, we have these climbing shoes and all products are relevant. I can go to the third page, and you can see that vector still delivers us relevant products.

Let's try another query.

[21:52] “Winterschalsack.” So, we have sleeping bags on the keyword side. Let's try the very same query here on the vector side. As you can see, keyword search provides us some relevant results, but for some reason, it's quite limited. But vector search shows lots of relevant results here. This shows us that vector search understands the query better and can provide much more relevant results than keyword search. Let's try something with a mistake.

[22:54] This is “Jakke,” which is written with a common mistake users make on online shops. (It should have a C.) And if I search for “Jakke” with the mistake on the keyword side, it returns mostly irrelevant results. Only one is relevant, but the others don’t fit our criteria, unfortunately. Let's try vector search here. And as you can see, vector search, even with the mistake, understands the intent. All results are relevant. I can go to the next page, and the results are still all jackets, not any missed products. That's all the examples I’ll show you today.

Launching vector search

[24:23] We have very exciting news today. After testing our vector search with some customers during the last couple of months, we’re ready to launch it. If you want to test our vector search with your product data, you can reach out to us. Just contact your Customer Success Manager (CSM) or email info@factfinder.com.

This launch only marks a starting point. We’ll soon be:

• Expanding our vector search beyond zero-hit queries

• Adding more control options to merge vector and keyword search results, which would shape our hybrid intelligence search

• And last but not least, adding image search in the future.

So, stay tuned for what's coming next. And over to you, Felix. Let's talk about our next initiative, AI Guided Selling (AIGS).

Transition to AI-Guided Selling

[25:10] Felix: Thank you, Yuliia. So, the next AI tool we’d like to introduce today is AI Guided Selling. Generally speaking, guided selling isn't new. We’ve been using guided selling for many years now. Many of our customers are very happy with the tool. It’s kind of a campaign, which helps customers determine and find the right product for their use case.

For example, guided selling campaigns are often used if customers only really have a rough idea of what they're looking for. And maybe if you have a very wide product range, you can use guided selling campaigns to help customers narrow down the results. The advantage is that they’re enjoyable to use, so customers are happy to return to your store.

[26:27] Here, you can see how guided selling looks in real time. So, this is the Bergfreunde channel. As you can see, I went to the jacket category. I chose my gender. I chose my use case. And after three or four clicks, I've narrowed down the number of results from over 1,000 to fewer than 100.

So, if we go through the process again, you can follow what I mean. I'm on the starting page. I go to clothing, then I go to jackets. I get around 1,500 results. Then I choose cycling. This leaves only about a tenth of the results. Then, everyday life as the next choice. And next up, the materials, in this case, windproof material. And I end up with a smaller selection of products tailored to my specific needs.

[27:23] So, of course, this is just a good example for jackets, but the same thing goes for any other category, whether it's DIY or food or anything else. You can customize the campaigns to your use case.

[27:39] The campaign we just saw was from the Bergfreunde online shop. They've been using guided selling successfully for a long time and have achieved great results:

• Up to 50% more guided selling usage rate per category

• Up to 40% of all customers who start using a guided selling campaign also finished using it

• Exit rates decreased by up to 28%

• Conversions increased by up to 50% per category

• Up to 70% more sales per session

• Fewer returns — because customers were more likely to find exactly what they were looking for with no need to return wrong items

• Customers enjoyed the experience and remembered it fondly

• And lastly, Bergfreunde also generated more tracking data and a better understanding of their customers' needs and wishes, which they could then use to improve their offering.

Overview of AI-Guided Selling

[28:57 ] Like I said, now we can automate the creation of guided selling campaigns using AI, and Yuliia is going to tell you more about how it works.

[29:14] Yuliia: Thank you, Felix.

So, what is AI Guided Selling?

Our AI Guided Selling enhances traditional advisor campaigns by introducing pre-generated advisor trees, created by FactFinder using large language models and Worldmatch® searches.

Instead of starting from scratch, campaign managers can simply enter a search term and the system generates an initial advisor tree.

Use cases. So, product guidance helps visitors find the right products through natural language interactions, especially helpful for complex products requiring explanation.

Decision support. You can assist undecided visitors in navigating large product catalogs.

Of course, sales boost. Guided selling campaigns engage customers who prefer not to use traditional filters or sorting, offering a conversational and intuitive buying experience.

[30:11] So, how it works. As you can see in the picture, the journey starts with the search term.

1. Users should enter a product-related term, for example, “camping equipment”, to kickstart the campaign.

2. FactFinder analyzes the product feed facets, filters and categories to build a structured advisor tree.

3. Every branch of the tree is processed through our integrated large language model to generate natural conversational questions.

4. Dialogues and answer options give the LLM full context, ensuring human-like advisory flows.

5. And of course, no manual product mapping is needed. Search-based filtering dynamically updates product recommendations based on user responses.

[31:04 ] It's important to note that AI Guided Selling is non-deterministic. While AIGS significantly accelerates campaign creation, human verification is still essential to ensure quality and accuracy. All AI-generated campaigns are set to inactive by default, giving campaign managers the full control to review, adjust and approve before going live. This module is designed to empower campaign managers, but not to replace them.

Why is human review critical? Let's look at some examples. Food products. A campaign for vegan tomato soup might miss hidden animal-based ingredients, like milk or eggs listed in detailed product descriptions.

Another example in specialized electronics. While AI reliably recommends standard electronics like phones, TVs or laptops, it may need some technical specifications for specialized devices, such as those that operate under specific external temperatures. AIGS delivers speed and scalability, but requires human expertise to guarantee relevance, safety and trust, especially for complex or sensitive product categories.

[32:25] So, what's the current status for AI Guided Selling? AIGS was built based on direct customer feedback gathered during the product discovery phase. This feature has been tested by a group of customers across industries like fashion, DIY, toys, electronics and dental devices. We are actively collecting feedback to refine and enhance the advisor tree generation for different eCommerce sectors. Development is already underway to boost the quality, accuracy and relevance of advisor trees based on real world use cases. Any customer interested in testing AIGS can get access by reaching out to their Customer Success Manager.

Thank you. Over to you, Tufan.

Importance of data enrichment

[33:17] Tufan: Thanks, Yuliia. Hello, everybody. I'm Tufan, and I'm a product manager at FactFinder. Today, I'll talk about our latest topic, data enrichment. Then, I'll do a wrap-up of our vision on how LLMs will change the eCommerce landscape before we move forward to our Q&A session. So, let's get into it.

When we talk about AI features like vector search or AIGS or any intelligent search and navigation experience, we often focus on the end result or the fancy way it's presented. But in reality, those features are only as good as the product data behind them. And that's where data enrichment comes in.

So, what is data enrichment? At its core, data enrichment means improving the quality, consistency and completeness of product data. We're talking about things like generating useful tags or attributes, standardizing values across fields, fixing spelling mistakes or adding missing information.

[34:18] In the past, and still today, most of the time, it's been done manually or rule based. But increasingly now, AI is stepping in to automate and scale this process. In eCommerce, this matters because better product data means better discoverability. Not just for keyword-based search, but also for vector search. And what better discoverability means for customers is they can find the right product faster and then for retailers, that they will deliver a smoother experience and push their numbers up.

[35:05] Now, if you're asking why we're talking about this and why this has become a hot topic in eCommerce now, it’s due to several interconnected reasons. First off, large language models have reached a point where they can interpret and transform product data at scale. What's been done in the past with rigid rules or manual human effort can now be done with smart, context-aware enrichment, if done right, of course. But this is the main enabler.

Secondly, better data not only means better discoverability, but also means better AI features. AI needs good data. So it's a catalyst to any AI feature that is in use right now, like vector search or AI Guided Selling or personalization or conversational search and so on. But also, any AI feature that will be invented in the future.

And last but not least, it's not only a supply matter. Everybody thinks they need better data. And now that LLM technology is maturing and AI features are becoming more and more profitable, the market is becoming aware of this need and starting to demand data enrichment solutions.

And I'm not making this up. We asked around. We made a survey with our customers, and 80% said their product feed has room for improvement. That's a huge number. And it was confirmed again and again in direct customer interviews.

Let me share a few examples. One customer went so far as to build their own in-house enrichment tool. Their reason is that their product data is generated by a different team than eCommerce management and isn't ideal for search and navigation. They wanted to generate helpful attributes that better reflected how their end customers show.

Another customer struggles with inconsistent field values from multiple suppliers differing in things like units or naming, which makes faceted navigation and filtering a very painful experience.

And of course, there were a bunch of customers who wanted to fill in missing fields. So, when you add all these on top of each other, you can see that it's only natural that this is becoming a thing now.

[37:39] Let's talk about where we are at FactFinder. We are actively working on a data enrichment MVP, a minimum viable product, that will be a first step in tracking this problem. The scope of the MVP covers the low-hanging fruit use case, tag generation or attribute generation, however you want to call it. And here's how it works. You select which products you want to enrich. Our AI-based enrichment service takes the most valuable information available in your data for those products and analyzes it. And it returns a list of suggested tags, attributes, to enhance the metadata of those products. Then you review these enrichment suggestions per product and either accept or decline them.

The goal is to improve keyword and vector search relevance, but we are designing it to be flexible so that you can use them anywhere else in your UI, for example, or choose how much they should influence search behavior.

[38:41] But tag generation or attribute generation or keyword generation, however you want to call it, is just the first step. As we collect feedback and iterate on the MVP, we are preparing to tackle more advanced use cases, like standardization of different values across fields, or eventually, even filling in missing field values, one of the hardest but most impactful use cases.

We are also exploring more interactive enrichment options, something like a free prompt field, where you can ask the enrichment service to do whatever you want, kind of like a ChatGPT for your product feed.

Challenges with data enrichment

[39:19] So, let's talk about the challenges. Everything sounds great, great plans, great opportunities, but of course it's not without challenges, so let's take a moment here to talk about them. One of the biggest problems we face is you can't blindly trust AI-generated data, especially when it comes to product data that your whole shop experience relies on. Remember what Felix talked about earlier. If an LLM suggests a wrong or misleading enrichment, it can confuse customers, damage trust and even hurt your numbers, your performance.

[40:08] So, how is this problem addressed right now when you look at the market? Some players offer manual rule-based enrichment, which works, but it's not smart. It doesn't generate anything new. And like I said, it’s very manual. It's even hard to call this data enrichment. It's like a light EIM solution where you can play around with your data. Others offer custom data services that require long, expensive collaboration cycles with their data teams. Neither of these feels efficient for us. They're either not scalable or not autonomous or both.

[40:36] At FactFinder, we believe there's a better way forward. Our vision is to develop a data enrichment solution that is autonomous, reliable, scalable, affordable and fast. Because a really great enrichment solution isn't just about generating data, it's about doing it in a reliable and efficient way. Without doing this, data enrichment stays as a marginal, costly and slow solution.

I'm not saying here that we are doing these perfectly. We are still at the very early stages of our solution. But we believe that it is possible to fulfill our vision by incorporating automatic assessment systems into the enrichment pipeline, like confidence scoring, anomaly detection methods, constant improvements to our models via feedback loops, through which we will be able to generate more and more reliable results or allowing A/B testing of enrichments to mitigate risks or ensure trust.

If we can tackle this problem — and we believe we have the means and we will tackle them — data enrichment will become a highly profitable solution that will catalyze eCommerce itself and all the innovations in it.

Final thoughts

[41:58] So let's wrap it up. Why is data enrichment important beyond just clean data? Because data quality is the invisible backbone of everything else we were talking about in this webinar. And AI or not, every new innovation we will be building. If you want better vector search results, you need clean, rich, contextual data. If you want AI assistance when creating guided shopping experiences, it needs structured and comparable product information. Even with the existing features. To get better results, your data has to be rich and relevant.

But again, it's not just about fixing bad data. Our goal at FactFinder is to offer a solution that is accurate, scalable, affordable and of course, delightful to use. We believe the challenges can be overcome with the right combination of AI, UX thinking and of course, smart automation. And the technology has matured to the level that allows us to start building a good solution.

The future of the eCommerce customer experience

[43:00] All right, so now let's leave data enrichment behind and go back to our broader vision on how LLMs are changing the eCommerce landscape before we conclude our webinar. In software product management, the best experience is what is natural for the user. And you will understand it better when you see the image in the next slide.

[43:29] Yes, as human beings, the most natural way of communicating for us is through natural language like this.

[43:36] Think of the shopping experience, for example. For thousands of years, we shopped like the image you see here. This is what is natural for human beings in this context, right? But over the past decades, we made this trade-off — and while we weren’t exactly thrilled about it, we got used to speaking in keywords when interacting with machines. In return, we gained some benefits, like the convenience of browsing and ordering from the comfort of our homes.

I call this a trade-off because it was a compromise. We didn't do it because keyword-based queries were the best solution or the best experience. We did it because it was what the technology allowed, and the benefits were so compelling.

But as LLMs are getting better and more reliable, we’re seeing search behavior becoming more natural. For example, more and more people are starting to search on ChatGPT instead of Google. And Google’s trying to prevent that by developing its own LLM solutions and integrating them even into its conventional search.

That's just the nature of things. It's destined to be like this, because you can't beat what comes naturally. We think that the same will happen in eCommerce too. and the future of eCommerce product discovery will be more conversational. Maybe not 100%, maybe it will be 100%, but it will definitely have a major role.

[45:03] Everything we talked about in this webinar so far, every feature that we are building right now is a step, a building block that will prepare us for this.

And of course, this paradigm shift won't happen overnight. This is a slow process. But it will eventually become the new norm. Yes, there is still a lot of hype around the technology and maybe more useless features than useful ones, which makes the whole thing feel like a scam sometimes. But once the dust settles, the real value will stand out from the noise, and we believe it will disrupt eCommerce space like everything else. And this is why we’re investing in this right now. We are not doing this because of the hype. We’re focusing on the other side of this cloud of dust.

Felix also has a few words to say about this.

[46:01] Felix: Thanks, Tufan. So, I think also from a business user perspective, talking to my customers, it seems like most companies, even if they aren’t using AI right now, they’re very aware of its potential and they’re exploring how to start using it. So maybe they aren’t doing it right now, but they are looking for partners like FactFinder to help them integrate AI solutions in their daily business.

Q&As

[46:39] Tobias: We are almost perfectly on time for the actual webinar. We are over actually, but I still want to take time for at least one or two quick questions, so we don't keep you guys too long. We’ve received lots of questions and we will address all of them, even those we can't address right now after the webinar, but we had one question about the models we are using for our solutions like vector search. I mentioned that it's hosted locally, it’s our own. It's not like an OpenAI solution. Can we say something about what we are using here?

[47:43] Yuliia: Thank you for the question. Indeed, as I mentioned, we’re not using OpenAI models and we do not use any external cloud-based LLMs for embedding generations. I’d like to keep what we’re using currently a small secret because this is our technology that we’re building.

But in any case, we’re ensuring all compliance, data protection and everything related to our customers’ concerns is fully covered. So, please be in touch with us and you’ll see the models’ performance will keep getting better, because we’re constantly training them on customer data, as I mentioned. Thank you.

[48:35] Tobias: Thank you. We also received lots of questions about specific use cases for vector search. For everyone who wants to try it out with their own data, we invite you to reach out for a personalized demo. I think that's the best approach, so you can actually try out what you would use it for.

I have one more question I'd like to ask you guys: what to do if AI Guided Selling generates something wrong. I think you mentioned it in passing, but if you could just talk about it a little bit more.

Yuliia: Thank you. So, of course, since it's AI and it's not deterministic, basically, it can create, still natural language speech questions, right? And it can make mistakes.

Of course, we do not push you to use this tree completely untouched. You can easily manually edit, delete or change any question it creates. You can also change the order of the questions and answers with the drag-and-drop function.

So basically, it’s important to understand that the AIGS-generated tree is a draft that you can customize to your needs. It's just to reduce your manual effort, because based on conversations with our customers, sometimes it takes days to create a tree, right? And with AIGS, you have a draft in just a few seconds, which you can tweak wherever you like.

[50:30] So, feel free to play with it. If you don't like one tree, you can just delete it and create a new tree by running the very same query. We’ll add even more customization in the future, so you can define how many questions to include, how long the Q&A sequence goes on for, etc. This is all based on our customers' feedback. So, the more feedback we receive, the better we can make the feature. Thank you.

[51:17] Tufan: I also want to say a few words about this. Please also, for this question, refer to my data enrichment presentation. You can rest assured that this is our number one nemesis right now. We’re perfectly aware of this problem, that this is the biggest challenge in all AI features. We need to solve this manual human review problem to make the AI innovations perfectly scalable, fast and free of human effort. But you will see this now in every solution. I would say our difference in FactFinder is we’re aware that this is the number one barrier to mass adoption, and we’re focused on tackling it.

[52:19] Tobias: So, watching the clock, I’d say we wrap it up for today. Again, all your questions will be addressed after this. I just wanted to thank everyone, of course, our dear listeners, for taking the time out of your day to listen to us for this whole time. Also, of course, thank you to our great speakers.